[{"data":1,"prerenderedAt":1631},["ShallowReactive",2],{"work-index":3,"profile":1591,"work-domains":1602},[4,165,236,310,493,606,737,869,938,1021,1086,1217,1349,1447,1516],{"id":5,"title":6,"body":7,"context":136,"description":127,"domain":137,"draft":138,"extension":139,"hero":140,"impact":141,"meta":142,"navigation":143,"path":144,"role":145,"scale":146,"seo":147,"slug":148,"stack":149,"status":156,"stem":157,"tags":158,"year":163,"__hash__":164},"work\u002Fwork\u002Fcognitive-erp.md","Cognitive ERP: Augment-First SAP Modernization",{"type":8,"value":9,"toc":126},"minimark",[10,15,19,23,26,30,59,63,119,123],[11,12,14],"h2",{"id":13},"the-mandate","The mandate",[16,17,18],"p",{},"I lead the Cognitive ERP work at Zettamine: an augment-first approach to modernizing the\nSAP systems enterprises actually run on. The strategy is deliberate. Rip-and-replace ERP\nprograms are where budgets and careers go to die. The opportunity is to make the core the\nbusiness already trusts noticeably smarter, without moving it.",[11,20,22],{"id":21},"the-problem","The problem",[16,24,25],{},"The ERP core is the least glamorous and most load-bearing system in the enterprise. It works,\nit is deeply entrenched, and it is also where most of the friction lives: rigid interfaces,\nreporting that lags reality, and armies of people doing routine process by hand. The instinct\nto replace it is expensive and usually wrong. The better move is to augment it, but augmentation\nthat respects the core is harder to design than a greenfield rebuild that ignores it.",[11,27,29],{"id":28},"the-system","The system",[31,32,33,41,47,53],"ul",{},[34,35,36,40],"li",{},[37,38,39],"strong",{},"Conversational UX over the core."," Natural-language interfaces onto SAP, so people ask the\nsystem for what they need instead of navigating it.",[34,42,43,46],{},[37,44,45],{},"Real-time analytics."," Reporting that tracks the business as it moves, rather than\nyesterday's batch.",[34,48,49,52],{},[37,50,51],{},"Autonomous process offloading."," Routine, rule-bound process handed to automation, freeing\npeople for the judgment work that is left.",[34,54,55,58],{},[37,56,57],{},"Application and mainframe modernization."," AI-driven modernization of legacy application and\nmainframe estates, using the NSL-era modernization work, so the core evolves rather than being\nfrozen or dumped.",[11,60,62],{"id":61},"key-decisions-and-what-they-cost","Key decisions and what they cost",[64,65,66,82],"table",{},[67,68,69],"thead",{},[70,71,72,76,79],"tr",{},[73,74,75],"th",{},"Decision",[73,77,78],{},"Why",[73,80,81],{},"What it traded",[83,84,85,97,108],"tbody",{},[70,86,87,91,94],{},[88,89,90],"td",{},"Augment the core rather than replace it",[88,92,93],{},"Lower risk, faster value, no bet-the-company rebuild",[88,95,96],{},"Harder architecture, constrained by the existing system",[70,98,99,102,105],{},[88,100,101],{},"Conversational and analytics layers on top",[88,103,104],{},"Meet people where the friction is",[88,106,107],{},"Deep integration with an entrenched core",[70,109,110,113,116],{},[88,111,112],{},"Offload process, keep judgment human",[88,114,115],{},"Automate the routine, not the decision",[88,117,118],{},"Careful scoping of what is safe to automate",[11,120,122],{"id":121},"what-it-proves","What it proves",[16,124,125],{},"The hardest enterprise-AI work is not greenfield, it is landing AI on the systems the business\ncannot afford to break. Cognitive ERP is that discipline: modernization measured by what still\nruns on Monday, not by how much got replaced.",{"title":127,"searchDepth":128,"depth":128,"links":129},"",3,[130,132,133,134,135],{"id":13,"depth":131,"text":14},2,{"id":21,"depth":131,"text":22},{"id":28,"depth":131,"text":29},{"id":61,"depth":131,"text":62},{"id":121,"depth":131,"text":122},"Zettamine","enterprise-core",false,"md","[object Object]","AI layered onto the SAP core, not ripped out",{},true,"\u002Fwork\u002Fcognitive-erp","Product Architect","Enterprise core systems",{"title":6,"description":127},"cognitive-erp",[150,151,152,153,154,155],"SAP","LLM","Conversational UX","Real-Time Analytics","Process Automation","Modernization","building","work\u002Fcognitive-erp",[159,160,161,162],"build","transform","sap","enterprise","2025","iBFmM5mMTKGYVbbVYQzvkW1oMhY34u7qXLgCSVwr9a0",{"id":166,"title":167,"body":168,"context":213,"description":127,"domain":214,"draft":138,"extension":139,"hero":215,"impact":216,"meta":217,"navigation":143,"path":218,"role":219,"scale":220,"seo":221,"slug":222,"stack":223,"status":156,"stem":229,"tags":230,"year":234,"__hash__":235},"work\u002Fwork\u002Ffactory-vision-predictive-maintenance.md","Factory Computer Vision and Predictive Maintenance",{"type":8,"value":169,"toc":207},[170,172,175,177,180,182,202,204],[11,171,14],{"id":13},[16,173,174],{},"I lead the industrial AI work at Zettamine aimed at mid-sized manufacturing: factory\ncomputer vision, predictive maintenance and cognitive digital twins. The largest\nmanufacturers can fund bespoke industrial AI. The mid-market cannot, and that is exactly\nwhere the systems are worth the most and least available.",[11,176,22],{"id":21},[16,178,179],{},"A mid-sized factory runs on equipment that fails expensively and quality processes that\ncatch defects late. The data to predict the failure and see the defect is being generated\ncontinuously, on sensors and cameras that are already installed, and almost none of it is\nused. The barrier is not the technology, it is that industrial AI has been priced and\nscoped for the few, not the many.",[11,181,29],{"id":28},[31,183,184,190,196],{},[34,185,186,189],{},[37,187,188],{},"Factory-floor vision."," Computer-vision models for defect detection and process\nmonitoring, running on the cameras already on the line.",[34,191,192,195],{},[37,193,194],{},"Predictive maintenance."," Time-series models over machine telemetry that forecast\nfailure before it stops the line, turning unplanned downtime into scheduled work.",[34,197,198,201],{},[37,199,200],{},"Cognitive digital twins."," A live model of the plant that fuses vision and telemetry,\nso operators can see state, anticipate failure and reason about change in one place.",[11,203,122],{"id":121},[16,205,206],{},"Bringing industrial AI down to the mid-market is an architecture and economics problem,\nnot a research one. The work is making systems that earn their keep on modest data and\nexisting hardware, which is a harder constraint than an unlimited lab.",{"title":127,"searchDepth":128,"depth":128,"links":208},[209,210,211,212],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":28,"depth":131,"text":29},{"id":121,"depth":131,"text":122},"Zettamine Industrial AI","industrial-vision","Factory-floor computer vision, predictive maintenance and cognitive digital twins built for mid-sized manufacturers, the segment large industrial AI keeps skipping.","Cognitive digital twins for the factory floor",{},"\u002Fwork\u002Ffactory-vision-predictive-maintenance","Architect","Mid-sized manufacturing",{"title":167,"description":127},"factory-vision-predictive-maintenance",[224,225,226,227,228],"Computer Vision","Predictive Maintenance","Time-Series ML","Digital Twin","IoT","work\u002Ffactory-vision-predictive-maintenance",[159,231,232,233],"vision","industrial","manufacturing","2024-2025","sAR0WzvEwLzANoVBgqPYPjYNK4eTIyJWHfH5qPXtp4c",{"id":237,"title":238,"body":239,"context":136,"description":127,"domain":290,"draft":138,"extension":139,"hero":291,"impact":292,"meta":293,"navigation":143,"path":294,"role":145,"scale":295,"seo":296,"slug":297,"stack":298,"status":156,"stem":304,"tags":305,"year":234,"__hash__":309},"work\u002Fwork\u002Fguildtrek.md","GuildTrek: Talent and Upskilling Platform",{"type":8,"value":240,"toc":284},[241,243,246,248,251,253,279,281],[11,242,14],{"id":13},[16,244,245],{},"I own the product architecture for GuildTrek, the talent and upskilling half of the\nZettamine product studio, and a natural companion to the hiring work in DigiHire. Where\nDigiHire decides who to bring in, GuildTrek decides how the people you already have grow.",[11,247,22],{"id":21},[16,249,250],{},"Corporate upskilling is mostly a content problem dressed up as a learning problem. Firms\nbuy a course library, assign it, and measure completion instead of capability. The result\nis expensive and inert: everyone finishes the module, nobody closes the gap that mattered.\nThe missing piece is a model of what a person can actually do, and what the role in front\nof them actually needs.",[11,252,29],{"id":28},[31,254,255,261,267,273],{},[34,256,257,260],{},[37,258,259],{},"Skill graph."," Roles, skills and evidence modeled as a graph rather than a flat\ncatalog, so a gap is a real distance between where someone is and where the role needs them.",[34,262,263,266],{},[37,264,265],{},"Personalized pathing."," LLM-driven recommendation that routes each person along the\nshortest credible path across that graph, tuned to their current level and target role.",[34,268,269,272],{},[37,270,271],{},"Evidence over completion."," Progress is measured against demonstrated capability, not\ntime spent, so the platform reports readiness instead of attendance.",[34,274,275,278],{},[37,276,277],{},"Shared spine with the hiring stack."," The same skill model underneath sourcing,\nscreening and growth, so the organization reasons about talent with one vocabulary.",[11,280,122],{"id":121},[16,282,283],{},"GuildTrek and DigiHire are two products over one idea: that talent decisions get better\nwhen skills are a model the system can reason about, not prose in a document. Architecting\nboth is how the studio keeps its products coherent rather than adjacent.",{"title":127,"searchDepth":128,"depth":128,"links":285},[286,287,288,289],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":28,"depth":131,"text":29},{"id":121,"depth":131,"text":122},"ai-products","A talent and upskilling platform that models skills as a living graph and routes each person along the path that actually closes their gap.","Skills modeled as a graph, not a course catalog",{},"\u002Fwork\u002Fguildtrek","Product platform",{"title":238,"description":127},"guildtrek",[151,299,300,301,302,303],"Skill Graph","Recommendation","RAG","Node.js","Nuxt","work\u002Fguildtrek",[159,306,307,308],"product","llm","talent","OlGjq2koIZerMrDM8RizdJhmltIve1olXDXgnFsnFPM",{"id":311,"title":312,"body":313,"context":470,"description":127,"domain":290,"draft":138,"extension":139,"hero":471,"impact":472,"meta":473,"navigation":143,"path":474,"role":475,"scale":476,"seo":477,"slug":478,"stack":479,"status":485,"stem":486,"tags":487,"year":491,"__hash__":492},"work\u002Fwork\u002Fai-interviewer.md","Real-Time Audio-Video Human-Interaction Agents",{"type":8,"value":314,"toc":463},[315,317,320,322,325,346,350,382,384,442,446,449,452],[11,316,14],{"id":13},[16,318,319],{},"I architected and led the build of a real-time conversational agent that had to\nhold a natural voice and video conversation, and produce a structured, comparable\nevaluation at the end of it. The hard part was never the language model. It was\nbuilding a system that stays conversational under a latency budget most LLM stacks\nquietly break.",[11,321,22],{"id":21},[16,323,324],{},"Recruiting teams burn the most expensive hour of every funnel on first-round\nscreenings. Most candidates do not survive that hour. The work is repetitive, the\nrubric is consistent, the conversation is bounded, in other words a textbook job\nfor an agent. But textbook hides three real problems:",[326,327,328,334,340],"ol",{},[34,329,330,333],{},[37,331,332],{},"Latency."," Voice conversations break the moment turn-taking lag crosses\nroughly 700 ms. Most LLM stacks are not built for that budget.",[34,335,336,339],{},[37,337,338],{},"Structure."," Hiring teams do not need transcripts. They need a rubric-aligned,\ncomparable, structured evaluation per candidate.",[34,341,342,345],{},[37,343,344],{},"Mode."," A real interview is voice plus video plus screen and sometimes code.\nSingle-mode bots feel like phone trees.",[11,347,349],{"id":348},"the-architecture","The architecture",[31,351,352,358,364,370,376],{},[34,353,354,357],{},[37,355,356],{},"Orchestration:"," Pipecat for pipeline-style agent composition (ASR, then\nreasoning, then TTS, with interruption handling).",[34,359,360,363],{},[37,361,362],{},"Transport:"," Daily.co for WebRTC video and audio with a WebSocket fallback,\nselected over rolling our own to compress build time on a non-differentiating layer.",[34,365,366,369],{},[37,367,368],{},"Reasoning core:"," the LLM is constrained to a custom evaluation JSON schema\nper candidate and per round, so the output is queryable, comparable and\nrubric-aligned at write time, not parsed after the fact.",[34,371,372,375],{},[37,373,374],{},"Latency budget:"," a total round-trip target under one second. Optimized\nWebSocket transitions and parallelized ASR and reasoning to keep the floor low.",[34,377,378,381],{},[37,379,380],{},"Round routing:"," technical, HR and managerial flows are configuration, not\ncode, so recruiters compose new flows without engineering.",[11,383,62],{"id":61},[64,385,386,396],{},[67,387,388],{},[70,389,390,392,394],{},[73,391,75],{},[73,393,78],{},[73,395,81],{},[83,397,398,409,420,431],{},[70,399,400,403,406],{},[88,401,402],{},"Pipecat over a custom orchestrator",[88,404,405],{},"Mature interruption and barge-in semantics out of the box",[88,407,408],{},"Some flexibility on novel turn-taking patterns",[70,410,411,414,417],{},[88,412,413],{},"Daily.co over self-hosted WebRTC",[88,415,416],{},"Time to production",[88,418,419],{},"Per-minute cost above a usage threshold",[70,421,422,425,428],{},[88,423,424],{},"Structured JSON output over free text",[88,426,427],{},"Comparability across candidates",[88,429,430],{},"Some loss of qualitative texture, contained with a notes field",[70,432,433,436,439],{},[88,434,435],{},"Sub-second target",[88,437,438],{},"Conversational naturalness",[88,440,441],{},"Forced eager inference and parallelization, and real compute cost",[11,443,445],{"id":444},"outcome","Outcome",[16,447,448],{},"A screening agent that talks like a person and reports like a rubric. Recruiters\ncompose the interview flow, the system runs it, and the evaluation lands\nstructured and comparable across every candidate.",[450,451],"hr",{},[16,453,454,457,458],{},[37,455,456],{},"Related writing:"," ",[459,460,462],"a",{"href":461},"\u002Fwriting\u002Fagentic-architecture-patterns","Five patterns I keep reaching for when designing agentic systems",{"title":127,"searchDepth":128,"depth":128,"links":464},[465,466,467,468,469],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":348,"depth":131,"text":349},{"id":61,"depth":131,"text":62},{"id":444,"depth":131,"text":445},"Brane Enterprises","A voice and video AI that conducts technical, HR and managerial screenings end to end, with sub-second response and structured evaluation output.","Full voice and video screening under a one-second turn budget",{},"\u002Fwork\u002Fai-interviewer","Architect and Technical Lead","Production, sub-second latency budget",{"title":312,"description":127},"ai-interviewer",[480,481,482,483,151,484],"Pipecat","Daily.co","WebRTC","WebSocket","JSON Schema","shipped","work\u002Fai-interviewer",[159,488,489,490],"agentic","multimodal","latency","2024","GdNg-XnZz8BFTHsRWAZiEPz51_QpBpE0bAOXG4KH0ps",{"id":494,"title":495,"body":496,"context":470,"description":127,"domain":290,"draft":138,"extension":139,"hero":588,"impact":589,"meta":590,"navigation":143,"path":591,"role":592,"scale":593,"seo":594,"slug":595,"stack":596,"status":485,"stem":602,"tags":603,"year":491,"__hash__":605},"work\u002Fwork\u002Fenterprise-digital-twin.md","Enterprise Digital Twin: CEO to Trainee",{"type":8,"value":497,"toc":581},[498,500,503,505,508,512,538,542,562,564,567,569],[11,499,14],{"id":13},[16,501,502],{},"I conceived and architected a digital twin of the organization itself. Enterprises\nmodel their factories and supply chains in software, then run the business that\nowns them on intuition. The mandate was to close that gap: give leadership a system\nthey could simulate decisions against, and give every role a mentor that understands\nwhat that role is supposed to be doing.",[11,504,22],{"id":21},[16,506,507],{},"Enterprises have digital twins of factories, supply chains, even buildings. Almost\nnone have a digital twin of themselves, of how decisions flow, who owns what, how a\nchange at the top propagates to the work that gets done. Which means two things break\nin parallel: executives cannot reliably model the downstream impact of decisions, and\nindividual contributors get generic training instead of context-aware mentorship. This\ntwin closes both gaps with one architecture.",[11,509,511],{"id":510},"architecture","Architecture",[31,513,514,520,526,532],{},[34,515,516,519],{},[37,517,518],{},"Hierarchy model:"," every role from CEO to trainee is represented as a node with a\ntyped responsibility surface, decision rights and observable outputs.",[34,521,522,525],{},[37,523,524],{},"Simulation engine:"," neural-network-backed what-if runs across budget, resource\nallocation and organizational change. Inputs are decisions, outputs are propagated\nstate changes across the hierarchy.",[34,527,528,531],{},[37,529,530],{},"Personalized agents:"," specialized LLM agents bound to individual roles that provide\nreal-time mentorship, automated skill-gap analysis and context-aware learning paths.\nThe agent knows what the role is supposed to be doing, not just what the person searched for.",[34,533,534,537],{},[37,535,536],{},"Feedback loop:"," real outcomes feed back into the simulation, so the twin sharpens\nover time rather than drifting.",[11,539,541],{"id":540},"why-this-is-hard","Why this is hard",[31,543,544,550,556],{},[34,545,546,549],{},[37,547,548],{},"Modeling decision rights"," is harder than modeling org charts. The formal structure\nis rarely the real structure.",[34,551,552,555],{},[37,553,554],{},"Simulation calibration."," A simulation that is always optimistic or always pessimistic\nis worse than no simulation. Earning trust required disciplined back-testing against\nhistorical decisions.",[34,557,558,561],{},[37,559,560],{},"Personalization without surveillance."," An agent that knows enough about your work to\nmentor you is also an agent that knows a lot about your work. Governance and consent were\narchitecture decisions, not afterthoughts.",[11,563,445],{"id":444},[16,565,566],{},"One architecture that lets executives rehearse a decision before they make it, and gives\nevery individual contributor a mentor tuned to their actual role. The twin gets sharper\nas real outcomes flow back into it.",[450,568],{},[16,570,571,457,573,576,577],{},[37,572,456],{},[459,574,575],{"href":461},"Five patterns for agentic systems"," and ",[459,578,580],{"href":579},"\u002Fwriting\u002Fenterprise-ai-adoption-playbook","The enterprise AI adoption playbook",{"title":127,"searchDepth":128,"depth":128,"links":582},[583,584,585,586,587],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":510,"depth":131,"text":511},{"id":540,"depth":131,"text":541},{"id":444,"depth":131,"text":445},"A hierarchy-wide digital twin of the organization, from CEO down to trainee, that lets executives simulate decisions and gives every role a personalized AI mentor.","Decision simulation plus role-aware mentorship on one model",{},"\u002Fwork\u002Fenterprise-digital-twin","Architect and Product Visionary","Organization-wide, CEO to trainee",{"title":495,"description":127},"enterprise-digital-twin",[597,598,599,600,601],"Neural Networks","LLM Agents","Simulation","Vector DB","Custom Models","work\u002Fenterprise-digital-twin",[159,160,488,604],"simulation","hH_e3T8dIvryx3ttGBNH6BP-Iq4MFP0J_-neDKDo6JQ",{"id":607,"title":608,"body":609,"context":136,"description":127,"domain":290,"draft":138,"extension":139,"hero":722,"impact":723,"meta":724,"navigation":143,"path":725,"role":726,"scale":727,"seo":728,"slug":729,"stack":730,"status":485,"stem":733,"tags":734,"year":735,"__hash__":736},"work\u002Fwork\u002Fdigihire-ai.md","DigiHire.ai: End-to-End AI Recruitment Platform",{"type":8,"value":610,"toc":711},[611,613,616,618,621,623,626,631,634,638,641,645,648,650,697,699,702,704],[11,612,14],{"id":13},[16,614,615],{},"I own the product architecture and technology direction for DigiHire.ai, one of\nthe flagship products at Zettamine. The brief was not a feature bolted onto an\napplicant tracking system. It was to redesign the recruitment workflow around\nwhat AI can now do, and to ship something recruiters would trust with a decision\nthat is expensive to get wrong.",[11,617,22],{"id":21},[16,619,620],{},"Recruitment in most enterprises is three disconnected workflows pretending to be\none: sourcing on professional networks, screening through applicant tracking\nsystems, and evaluating in spreadsheets. The seams between them eat the bulk of\nrecruiter time and most of the signal. The premise of DigiHire.ai: collapse the\nthree into one AI-native loop where the system sources, the system screens, and\nthe system ranks, with humans intervening on judgment rather than janitorial work.",[11,622,511],{"id":510},[16,624,625],{},"Three engines under one product.",[627,628,630],"h3",{"id":629},"_1-deep-sourcing-engine","1. Deep-sourcing engine",[16,632,633],{},"A custom Chrome extension using segmented scrolling and DOM hydration to extract\nstructured profile data from professional networks that expose no usable public\nAPI, without tripping anti-bot heuristics. Output is piped into a typed candidate\nmodel, one adapter per source.",[627,635,637],{"id":636},"_2-llm-based-semantic-screening","2. LLM-based semantic screening",[16,639,640],{},"Resumes and, when available, interview transcripts are passed through an LLM with\nrole-specific prompts that return structured evaluations rather than free text.\nEvery evaluation is vector-indexed for similarity search across the candidate pool.",[627,642,644],{"id":643},"_3-ranking-and-recommendation","3. Ranking and recommendation",[16,646,647],{},"Candidates are ranked against the role profile using a hybrid of semantic\nsimilarity, structured-criterion matching and explicit recruiter weights. Every\nrank is explainable down to the contributing signals.",[11,649,62],{"id":61},[64,651,652,662],{},[67,653,654],{},[70,655,656,658,660],{},[73,657,75],{},[73,659,78],{},[73,661,81],{},[83,663,664,675,686],{},[70,665,666,669,672],{},[88,667,668],{},"Browser-extension sourcing over API ingestion",[88,670,671],{},"Coverage of networks with no public API",[88,673,674],{},"Brittleness when the DOM changes, contained with an adapter per source",[70,676,677,680,683],{},[88,678,679],{},"Structured LLM evaluation over free text plus parsing",[88,681,682],{},"Trust and explainability",[88,684,685],{},"Higher prompt-engineering cost up front",[70,687,688,691,694],{},[88,689,690],{},"Hybrid ranking (vector, criterion, weight) over pure vector",[88,692,693],{},"Recruiters need to steer the outcome",[88,695,696],{},"More moving parts, longer evaluation cycles",[11,698,445],{"id":444},[16,700,701],{},"Time-to-hire fell by more than 60 percent for pilot accounts, and every ranking\nstayed auditable, which is what turned it from a demo into a system a recruiting\nteam would actually run.",[450,703],{},[16,705,706,457,708],{},[37,707,456],{},[459,709,710],{"href":579},"The enterprise AI adoption playbook nobody writes",{"title":127,"searchDepth":128,"depth":128,"links":712},[713,714,715,720,721],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":510,"depth":131,"text":511,"children":716},[717,718,719],{"id":629,"depth":128,"text":630},{"id":636,"depth":128,"text":637},{"id":643,"depth":128,"text":644},{"id":61,"depth":131,"text":62},{"id":444,"depth":131,"text":445},"An AI-native recruitment platform that sources, screens and ranks candidates in one loop, and cut time-to-hire by more than 60 percent for pilot accounts.","Time-to-hire cut by more than 60%",{},"\u002Fwork\u002Fdigihire-ai","Product Architect and Technology Owner","Production platform, multi-account",{"title":608,"description":127},"digihire-ai",[151,600,731,732,302,303],"Semantic Search","Chrome Extension","work\u002Fdigihire-ai",[159,160,307,306],"2023-2024","C3h-5ajUq2PxC9eeX2YyvBXSAJ5_wGNo4-x4T04TzwQ",{"id":738,"title":739,"body":740,"context":851,"description":127,"domain":290,"draft":138,"extension":139,"hero":852,"impact":853,"meta":854,"navigation":143,"path":855,"role":219,"scale":856,"seo":857,"slug":858,"stack":859,"status":485,"stem":864,"tags":865,"year":735,"__hash__":868},"work\u002Fwork\u002Fenterprise-rag-conversational-ai.md","Enterprise RAG and Conversational AI",{"type":8,"value":741,"toc":844},[742,744,747,749,752,754,780,782,829,831,834,836],[11,743,14],{"id":13},[16,745,746],{},"I architected a set of retrieval-augmented and conversational AI systems across\nenterprise engagements, several of them in finance and healthcare. The common thread\nwas not the model. It was building systems that could be trusted in domains where a\nconfident wrong answer carries real consequences.",[11,748,22],{"id":21},[16,750,751],{},"A raw LLM in a regulated enterprise is a fluent liability. It answers from a frozen,\ngeneric memory, it cannot cite where an answer came from, and it will invent detail\nrather than admit a gap. In finance and healthcare that is not a rough edge, it is a\nreason not to ship. The work is turning a general model into a system that answers from\nthe enterprise's own current knowledge and can show its sources.",[11,753,29],{"id":28},[31,755,756,762,768,774],{},[34,757,758,761],{},[37,759,760],{},"Retrieval-augmented generation."," Answers grounded in vector-indexed enterprise\nknowledge, retrieved at query time, so responses track the current corpus rather than\nthe model's training snapshot.",[34,763,764,767],{},[37,765,766],{},"Citations by construction."," Every answer carries the sources it was built from,\nwhich is what makes the output auditable rather than merely plausible.",[34,769,770,773],{},[37,771,772],{},"Voice and chat surfaces."," Both voicebot and chatbot front ends over the same\nretrieval core, so the channel is a choice and the grounding is constant.",[34,775,776,779],{},[37,777,778],{},"Guardrails for regulated use."," Refusal and escalation paths for out-of-scope or\nlow-confidence queries, because in these domains a careful no beats a fluent guess.",[11,781,62],{"id":61},[64,783,784,794],{},[67,785,786],{},[70,787,788,790,792],{},[73,789,75],{},[73,791,78],{},[73,793,81],{},[83,795,796,807,818],{},[70,797,798,801,804],{},[88,799,800],{},"RAG over fine-tuning as the default",[88,802,803],{},"Freshness and citable sources",[88,805,806],{},"A retrieval and indexing pipeline to own and keep current",[70,808,809,812,815],{},[88,810,811],{},"Citations required, not optional",[88,813,814],{},"Auditability in regulated domains",[88,816,817],{},"Extra plumbing on every answer path",[70,819,820,823,826],{},[88,821,822],{},"Explicit refusal and escalation",[88,824,825],{},"A careful no is safer than a confident guess",[88,827,828],{},"Some coverage traded for trust",[11,830,122],{"id":121},[16,832,833],{},"Shipping conversational AI into finance and healthcare is a governance problem as much\nas a model problem. These systems are the pattern I keep reusing: ground the answer, cite\nthe source, and design the failure mode on purpose.",[450,835],{},[16,837,838,457,840],{},[37,839,456],{},[459,841,843],{"href":842},"\u002Fwriting\u002Fdata-problem-in-costume","The data problem in a costume",{"title":127,"searchDepth":128,"depth":128,"links":845},[846,847,848,849,850],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":28,"depth":131,"text":29},{"id":61,"depth":131,"text":62},{"id":121,"depth":131,"text":122},"Multiple enterprise engagements","Retrieval-augmented assistants and voice and chat agents built for finance and healthcare, where an ungrounded answer is not a bug, it is a liability.","Grounded answers with citations, in finance and healthcare",{},"\u002Fwork\u002Fenterprise-rag-conversational-ai","Cross-industry, regulated domains",{"title":739,"description":127},"enterprise-rag-conversational-ai",[301,600,151,860,861,862,863],"Voicebot","Chatbot","Retrieval","Guardrails","work\u002Fenterprise-rag-conversational-ai",[159,307,866,867],"rag","conversational","N3GXI3a4hiMHe9Y_z-eqnjvK2WALdX5_5F-nYsonfWg",{"id":870,"title":871,"body":872,"context":917,"description":127,"domain":214,"draft":138,"extension":139,"hero":918,"impact":919,"meta":920,"navigation":143,"path":921,"role":922,"scale":923,"seo":924,"slug":925,"stack":926,"status":485,"stem":931,"tags":932,"year":936,"__hash__":937},"work\u002Fwork\u002Fdiagnostic-vision-models.md","Diagnostic Vision Models for Medicine and Agriculture",{"type":8,"value":873,"toc":911},[874,876,879,881,884,886,906,908],[11,875,14],{"id":13},[16,877,878],{},"I built and fine-tuned deep-learning models for diagnosis across two domains where a\nmissed positive is the expensive error: medical imaging (brain-tumor and skin-cancer\ndetection) and agriculture (plant-disease detection). Different fields, same discipline:\na model whose mistakes have a real-world cost.",[11,880,22],{"id":21},[16,882,883],{},"Diagnosis from images is a classification problem where the two kinds of error are not\nequal. In tumor and skin-cancer detection a false negative can be a life. In crop-disease\ndetection a missed outbreak can be a season. A model tuned to maximize headline accuracy\nwill happily trade away the exact errors that matter, because the rare, dangerous case is\nunderrepresented in the data. Getting this right is about the loss you optimize and the\ndata you curate, not the size of the network.",[11,885,29],{"id":28},[31,887,888,894,900],{},[34,889,890,893],{},[37,891,892],{},"Fine-tuned detectors."," Models adapted to each diagnostic task rather than trained from\nscratch, so limited labeled data goes further.",[34,895,896,899],{},[37,897,898],{},"Cost-aware evaluation."," Tuned and measured against the error that hurts, sensitivity to\nthe dangerous class weighted above raw accuracy.",[34,901,902,905],{},[37,903,904],{},"Domain-honest data curation."," Careful handling of class imbalance and edge cases, because\nin diagnosis the rare case is the whole point.",[11,907,122],{"id":121},[16,909,910],{},"The same engineering discipline moves across medicine and agriculture: understand which\nmistake is unacceptable, then build and evaluate the model around that, not around a\nleaderboard. High-stakes classification is where that discipline shows.",{"title":127,"searchDepth":128,"depth":128,"links":912},[913,914,915,916],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":28,"depth":131,"text":29},{"id":121,"depth":131,"text":122},"Applied ML engagements","Fine-tuned image models for high-stakes diagnosis, from brain-tumor and skin-cancer detection in medical imaging to plant-disease detection across crops.","Fine-tuned detectors for tumors, skin cancer and crop disease",{},"\u002Fwork\u002Fdiagnostic-vision-models","ML Architect","Fine-tuned models, high-stakes classification",{"title":871,"description":127},"diagnostic-vision-models",[927,928,929,930],"Deep Learning","Medical Imaging","Model Fine-Tuning","Image Classification","work\u002Fdiagnostic-vision-models",[159,231,933,934,935],"ml","healthcare","agriculture","2022-2023","RbExKqbOdJkrMrS8UFnAOuqb6JSgUdu7uYQNLcM2ZAY",{"id":939,"title":940,"body":941,"context":997,"description":127,"domain":998,"draft":138,"extension":139,"hero":999,"impact":1000,"meta":1001,"navigation":143,"path":1002,"role":1003,"scale":1004,"seo":1005,"slug":1006,"stack":1007,"status":485,"stem":1013,"tags":1014,"year":1019,"__hash__":1020},"work\u002Fwork\u002Fdrone-swarm-surveillance.md","Autonomous Drone Swarm for Surveillance",{"type":8,"value":942,"toc":991},[943,945,948,950,953,955,975,977,980,982],[11,944,14],{"id":13},[16,946,947],{},"I architected an autonomous drone swarm for surveillance in defense and internal-security\nsettings. A single drone extends one operator's reach. A swarm changes the problem: many\naircraft covering an area together, coordinating without a human flying each one.",[11,949,22],{"id":21},[16,951,952],{},"Wide-area surveillance with individually piloted drones does not scale. Every aircraft\nneeds an operator, coverage is limited by headcount, and the drones do not cooperate, they\njust happen to be nearby. The step change is autonomy plus coordination: a group that\nallocates coverage among itself, adapts as conditions change, and behaves as one system\nrather than a fleet of singles.",[11,954,29],{"id":28},[31,956,957,963,969],{},[34,958,959,962],{},[37,960,961],{},"Autonomous coordination."," Drones that divide and cover an area cooperatively, without\na pilot bound to each aircraft.",[34,964,965,968],{},[37,966,967],{},"Swarm behavior."," Coverage that reorganizes as the situation changes, so the group adapts\nrather than flying a fixed script.",[34,970,971,974],{},[37,972,973],{},"Mission-grade constraints."," Designed for the reliability and operational demands of\ndefense and internal-security use, where the surveillance has to hold up under real stakes.",[11,976,122],{"id":121},[16,978,979],{},"Swarm autonomy is coordination engineering under hard constraints: many autonomous agents,\none coherent behavior, in a setting that does not forgive a sloppy failure mode. It is the\nrobotics counterpart to the multi-agent thinking that runs through the software work.",[450,981],{},[16,983,984,457,987],{},[37,985,986],{},"Related work:",[459,988,990],{"href":989},"\u002Fwork\u002Fautonomous-aerial-systems","Autonomous Aerial Systems",{"title":127,"searchDepth":128,"depth":128,"links":992},[993,994,995,996],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":28,"depth":131,"text":29},{"id":121,"depth":131,"text":122},"Defense and internal-security program","autonomous","A coordinated autonomous drone swarm for defense and internal-security surveillance, where many aircraft cover ground together without a pilot per drone.","Coordinated autonomous coverage across a wide area",{},"\u002Fwork\u002Fdrone-swarm-surveillance","System Architect","Coordinated multi-drone system",{"title":940,"description":127},"drone-swarm-surveillance",[1008,1009,1010,1011,1012],"Swarm Robotics","Autonomy","Coordination","Surveillance","Edge Systems","work\u002Fdrone-swarm-surveillance",[159,1015,1016,1017,1018],"drones","swarm","autonomy","defense","2021-2022","xH_M21gIz_xCRw37r3UOQnRZzi_EzlZyG1M8GDF4Aw0",{"id":1022,"title":1023,"body":1024,"context":1069,"description":127,"domain":214,"draft":138,"extension":139,"hero":1070,"impact":1071,"meta":1072,"navigation":143,"path":1073,"role":1074,"scale":1075,"seo":1076,"slug":1077,"stack":1078,"status":485,"stem":1081,"tags":1082,"year":1084,"__hash__":1085},"work\u002Fwork\u002Findustrial-safety-vision.md","Industrial Safety Vision Analytics",{"type":8,"value":1025,"toc":1063},[1026,1028,1031,1033,1036,1038,1058,1060],[11,1027,14],{"id":13},[16,1029,1030],{},"I architected a vision-analytics safety system for an industrial site in the Australian\nmining sector. The goal was blunt: use the cameras already on site to catch the conditions\nthat precede an injury, and raise them while there is still time to act.",[11,1032,22],{"id":21},[16,1034,1035],{},"Heavy industrial environments are dangerous in ways that are visible but not watched.\nA person in the wrong zone, missing protective equipment, a machine interaction that\nshould not be happening: all of it is on camera, and none of it is seen until after\nsomething goes wrong, because no human can watch every feed continuously. Safety in these\nsettings is a monitoring problem at a scale humans cannot hold.",[11,1037,29],{"id":28},[31,1039,1040,1046,1052],{},[34,1041,1042,1045],{},[37,1043,1044],{},"Real-time video analytics."," Deep-learning models reading live camera feeds to detect\npeople, zones, equipment and unsafe interactions frame by frame.",[34,1047,1048,1051],{},[37,1049,1050],{},"Protective-equipment and zone checks."," Detection of missing safety gear and presence\nin restricted or hazardous zones, turned into alerts rather than after-the-fact reports.",[34,1053,1054,1057],{},[37,1055,1056],{},"Edge-aware inference."," Inference placed to meet the latency and connectivity reality\nof an industrial site, so a warning arrives in time to matter.",[11,1059,122],{"id":121},[16,1061,1062],{},"Industrial vision is where computer vision stops being a benchmark and starts carrying a\nphysical cost for a wrong call. Building a safety system that operators trust means tuning\nfor the failure that hurts a person, not the one that dents an accuracy metric.",{"title":127,"searchDepth":128,"depth":128,"links":1064},[1065,1066,1067,1068],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":28,"depth":131,"text":29},{"id":121,"depth":131,"text":122},"Australian mining engagement","A computer-vision safety system for the Australian mining industry that reads live site feeds and flags hazards and protective-equipment violations before they become incidents.","Hazards flagged from live camera feeds in real time",{},"\u002Fwork\u002Findustrial-safety-vision","Solution Architect","Operational safety, industrial site",{"title":1023,"description":127},"industrial-safety-vision",[224,927,1079,1080],"Video Analytics","Edge Inference","work\u002Findustrial-safety-vision",[159,231,232,1083],"safety","2021","uwSCjSkTkJazjmf3qx2texnv412K-A69ybLk4ZqpUlM",{"id":1087,"title":1088,"body":1089,"context":470,"description":127,"domain":1194,"draft":138,"extension":139,"hero":1195,"impact":1196,"meta":1197,"navigation":143,"path":1198,"role":1199,"scale":1200,"seo":1201,"slug":1202,"stack":1203,"status":485,"stem":1210,"tags":1211,"year":1215,"__hash__":1216},"work\u002Fwork\u002Fnsl-silicon-hardware.md","NSL Custom Silicon and Hardware Interfacing Layer",{"type":8,"value":1090,"toc":1187},[1091,1093,1096,1098,1101,1103,1123,1125,1172,1174,1177,1179],[11,1092,14],{"id":13},[16,1094,1095],{},"The Natural Solution Language program did not stop at software. I led its descent to\nhardware: designing a custom microprocessor tuned for the NSL model of execution, and\na hardware-interfacing layer that let NSL-defined solutions drive real machines. This is\nthe rarest span an architect gets to own, from a language grammar at the top to custom\nsilicon at the bottom, in one coherent program.",[11,1097,22],{"id":21},[16,1099,1100],{},"A language that compiles to a general-purpose CPU inherits every assumption that CPU was\ndesigned around. If the whole point of NSL is a different way of expressing and running\nsolutions, then at some altitude the general-purpose substrate stops being neutral and\nstarts being a tax. The question was whether NSL execution deserved silicon designed for\nit, and whether that silicon could actually reach the field: drones, robots, IoT nodes and\nAR\u002FVR devices, not just a lab bench.",[11,1102,29],{"id":28},[31,1104,1105,1111,1117],{},[34,1106,1107,1110],{},[37,1108,1109],{},"Custom microprocessor (chipset) design."," A processor architecture designed around the\nNSL execution model rather than retrofitted to it, taken through design into fabricated silicon.",[34,1112,1113,1116],{},[37,1114,1115],{},"Hardware-interfacing layer."," The bridge between compiled NSL and physical devices, so a\nsolution expressed in the language could actuate and sense real hardware.",[34,1118,1119,1122],{},[37,1120,1121],{},"Production across device classes."," The layer was carried from prototype to production\nacross drones, robots, industrial IoT and AR\u002FVR, proving the stack held under the messiness\nof real devices, not just simulation.",[11,1124,62],{"id":61},[64,1126,1127,1137],{},[67,1128,1129],{},[70,1130,1131,1133,1135],{},[73,1132,75],{},[73,1134,78],{},[73,1136,81],{},[83,1138,1139,1150,1161],{},[70,1140,1141,1144,1147],{},[88,1142,1143],{},"Design custom silicon rather than target a stock CPU",[88,1145,1146],{},"Match the processor to the NSL execution model",[88,1148,1149],{},"The cost, risk and cycle time of a chipset program",[70,1151,1152,1155,1158],{},[88,1153,1154],{},"One hardware layer across drones, robots, IoT and AR\u002FVR",[88,1156,1157],{},"Prove generality, not a single-device demo",[88,1159,1160],{},"A far larger integration and testing surface",[70,1162,1163,1166,1169],{},[88,1164,1165],{},"Prototype to production, not proof of concept",[88,1167,1168],{},"Field reality is the only honest test",[88,1170,1171],{},"Sustained hardening under real-world failure modes",[11,1173,122],{"id":121},[16,1175,1176],{},"Language design, compiler engineering and custom silicon are usually three different careers.\nOwning the architecture across all three, and landing it on production hardware, is the clearest\nevidence of depth in this portfolio.",[450,1178],{},[16,1180,1181,457,1183],{},[37,1182,986],{},[459,1184,1186],{"href":1185},"\u002Fwork\u002Fnsl-language-platform","Natural Solution Language: Language, Compiler and No-Code Platform",{"title":127,"searchDepth":128,"depth":128,"links":1188},[1189,1190,1191,1192,1193],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":28,"depth":131,"text":29},{"id":61,"depth":131,"text":62},{"id":121,"depth":131,"text":122},"frontier","A custom NSL microprocessor and a hardware-interfacing layer that carried the language all the way down to drones, robots, IoT devices and AR\u002FVR, from prototype to production.","Custom microprocessor plus a hardware layer taken to production",{},"\u002Fwork\u002Fnsl-silicon-hardware","Program Architect","Language to silicon to production hardware",{"title":1088,"description":127},"nsl-silicon-hardware",[1204,1205,1206,1207,1208,228,1209],"Chipset Design","Custom Microprocessor","Embedded Systems","Drones","Robotics","AR\u002FVR","work\u002Fnsl-silicon-hardware",[159,1194,1212,1213,1214],"silicon","hardware","robotics","2020-2024","fphYBXljfugNp-ATpN0SgMG5MpQJEq80_RddloTBa8k",{"id":1218,"title":1219,"body":1220,"context":470,"description":127,"domain":1194,"draft":138,"extension":139,"hero":1329,"impact":1330,"meta":1331,"navigation":143,"path":1185,"role":1332,"scale":1333,"seo":1334,"slug":1335,"stack":1336,"status":485,"stem":1342,"tags":1343,"year":1347,"__hash__":1348},"work\u002Fwork\u002Fnsl-language-platform.md","Natural Solution Language: A Language, Compiler and No-Code Platform",{"type":8,"value":1221,"toc":1322},[1222,1224,1227,1229,1232,1234,1260,1262,1309,1311,1314,1316],[11,1223,14],{"id":13},[16,1225,1226],{},"I conceived and led the Natural Solution Language (NSL) program: an attempt to move\nthe act of building software up a level, from writing code to expressing intent.\nMost no-code tools are form builders with a database behind them. NSL was the harder\nversion of the idea: a real language, with real semantics, a real compiler, and a\nruntime that could take a solution described in near-natural terms and execute it.",[11,1228,22],{"id":21},[16,1230,1231],{},"Enterprises do not have a shortage of developers so much as a shortage of translation.\nA business owns the problem, an engineering team owns the solution, and most of the cost\nand most of the drift lives in the gap between them. Existing low-code tools narrow that\ngap for trivial apps and widen it for anything real, because the moment logic gets\ncomplex you fall off the visual editor and back into code the business cannot read.\nNSL set out to make the description itself the program.",[11,1233,29],{"id":28},[31,1235,1236,1242,1248,1254],{},[34,1237,1238,1241],{},[37,1239,1240],{},"Language design."," A grammar and semantics for expressing solutions declaratively,\nclose to how a domain owner would state the rule, with enough formal structure to compile.",[34,1243,1244,1247],{},[37,1245,1246],{},"Compiler toolchain."," Lexer, parser, semantic analysis and code generation, taking NSL\nsource to executable artifacts. This is where most no-code claims stop and NSL kept going.",[34,1249,1250,1253],{},[37,1251,1252],{},"No-code runtime and platform."," An execution environment that runs compiled NSL,\nwiring data, logic and interface without hand-written glue, taken through to production use.",[34,1255,1256,1259],{},[37,1257,1258],{},"Transformer-model research."," Advanced Transformer research feeding the platform, so\nintent expressed in natural language could be lifted toward valid NSL, closing the loop\nbetween how people describe a solution and how the machine runs it.",[11,1261,62],{"id":61},[64,1263,1264,1274],{},[67,1265,1266],{},[70,1267,1268,1270,1272],{},[73,1269,75],{},[73,1271,78],{},[73,1273,81],{},[83,1275,1276,1287,1298],{},[70,1277,1278,1281,1284],{},[88,1279,1280],{},"Design a real language over extending an existing one",[88,1282,1283],{},"Full control of semantics and the path to hardware",[88,1285,1286],{},"Years of compiler and tooling work before anything shipped",[70,1288,1289,1292,1295],{},[88,1290,1291],{},"Compile rather than interpret at the top layer",[88,1293,1294],{},"Performance and a real optimization surface",[88,1296,1297],{},"A heavier toolchain to build and maintain",[70,1299,1300,1303,1306],{},[88,1301,1302],{},"Couple language design to model research early",[88,1304,1305],{},"Natural-language intent could feed the compiler",[88,1307,1308],{},"Two hard research fronts advancing at once",[11,1310,122],{"id":121},[16,1312,1313],{},"Very few people get to take an idea from grammar and semantics through a working compiler\nand into a production runtime. NSL is the anchor of the depth in this portfolio: it is the\nsoftware half of a program whose other half went all the way to custom silicon.",[450,1315],{},[16,1317,1318,457,1320],{},[37,1319,986],{},[459,1321,1088],{"href":1198},{"title":127,"searchDepth":128,"depth":128,"links":1323},[1324,1325,1326,1327,1328],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":28,"depth":131,"text":29},{"id":61,"depth":131,"text":62},{"id":121,"depth":131,"text":122},"A ground-up programming language for expressing business solutions in near-natural terms, its compiler toolchain and a no-code platform that turns intent into running software.","A new language, its compilers and a production no-code platform",{},"Program Architect and Originator","Multi-year deep-tech program",{"title":1219,"description":127},"nsl-language-platform",[1337,1338,1339,1340,1341],"Language Design","Compiler Design","DSL","No-Code Runtime","Transformer Models","work\u002Fnsl-language-platform",[159,1194,1344,1345,1346],"language","compiler","platform","2019-2024","oEcdiMA_W70T-5YFz2CXaJYW-QB8sYdOIYGUtECjRbI",{"id":1350,"title":1351,"body":1352,"context":1430,"description":127,"domain":998,"draft":138,"extension":139,"hero":1431,"impact":1432,"meta":1433,"navigation":143,"path":989,"role":1434,"scale":1435,"seo":1436,"slug":1437,"stack":1438,"status":485,"stem":1443,"tags":1444,"year":1445,"__hash__":1446},"work\u002Fwork\u002Fautonomous-aerial-systems.md","Autonomous Aerial Systems: Heavy-Lift, High-Altitude and Survey",{"type":8,"value":1353,"toc":1424},[1354,1356,1359,1363,1389,1391,1411,1413,1416,1418],[11,1355,14],{"id":13},[16,1357,1358],{},"I have designed, built and flown drones since 2016, not as a hobby but as a series of\nreal missions with real payloads and real constraints. The through-line is engineering a\nplatform that survives the field: heavy loads, thin air, long range and results a\ngovernment or a scientific program will actually use.",[11,1360,1362],{"id":1361},"the-missions","The missions",[31,1364,1365,1371,1377,1383],{},[34,1366,1367,1370],{},[37,1368,1369],{},"Heavy-lift Falcon platform."," A drone built to carry roughly 15 kg of payload, which\ncompleted a 46 km flight carrying a 10 kg load. Range and lift together, which is the hard\ncombination.",[34,1372,1373,1376],{},[37,1374,1375],{},"High-altitude trials, Siachen."," Flown in Himalayan conditions to a maximum altitude of\nabout 14,500 ft, where thin air punishes every assumption a lowland design makes.",[34,1378,1379,1382],{},[37,1380,1381],{},"Nation-scale land survey."," Piloted a land-revenue survey program for the Andhra Pradesh\ngovernment, mapping 25 villages across roughly 600 sq km in three days, work that would take\nground teams far longer.",[34,1384,1385,1388],{},[37,1386,1387],{},"Ecological survey."," Built my first drone in 2016 for ecological studies in collaboration\nwith the Australian government's maritime department.",[11,1390,541],{"id":540},[31,1392,1393,1399,1405],{},[34,1394,1395,1398],{},[37,1396,1397],{},"Payload versus range."," Lift costs energy and energy costs range. Carrying 10 kg for\n46 km is a design compromise won at the airframe, propulsion and power level, not bought off\na shelf.",[34,1400,1401,1404],{},[37,1402,1403],{},"Altitude."," At 14,500 ft the air that provides lift and cooling is thin. A platform that\nflies at sea level is a different machine in the Himalayas.",[34,1406,1407,1410],{},[37,1408,1409],{},"Field reliability."," A survey that has to cover 600 sq km in three days cannot afford the\nfailure rate a demo tolerates. The engineering standard is the mission, not the flight.",[11,1412,122],{"id":121},[16,1414,1415],{},"Owning an autonomous system from airframe to flight to delivered survey data is full-stack\nin the most literal sense. The drones are the physical proof of the same instinct that runs\nthrough the software work: build the thing, then make it survive contact with the real world.",[450,1417],{},[16,1419,1420,457,1422],{},[37,1421,986],{},[459,1423,940],{"href":1002},{"title":127,"searchDepth":128,"depth":128,"links":1425},[1426,1427,1428,1429],{"id":13,"depth":131,"text":14},{"id":1361,"depth":131,"text":1362},{"id":540,"depth":131,"text":541},{"id":121,"depth":131,"text":122},"Self-directed and government programs","Custom drones designed, built and flown across real missions, from a heavy-lift Falcon platform and high-altitude Himalayan trials to nation-scale land-survey and ecological work.","10 kg payload over 46 km, flown at 14,500 ft",{},"Designer, Builder and Pilot","Prototype to field missions",{"title":1351,"description":127},"autonomous-aerial-systems",[1439,1440,1009,1441,1442],"Drone Design","Flight Systems","Payload Engineering","Geospatial","work\u002Fautonomous-aerial-systems",[159,1015,1214,1017],"2016-2022","D1F1BdFNDvMbWFkijfO4eEGErv-7gdTYOaBfIk14MFo",{"id":1448,"title":1449,"body":1450,"context":1497,"description":127,"domain":137,"draft":138,"extension":139,"hero":1498,"impact":1499,"meta":1500,"navigation":143,"path":1501,"role":1502,"scale":1503,"seo":1504,"slug":1505,"stack":1506,"status":485,"stem":1511,"tags":1512,"year":1514,"__hash__":1515},"work\u002Fwork\u002Fsap-globe-nestle.md","SAP Globe Global Template Rollout, Nestle",{"type":8,"value":1451,"toc":1491},[1452,1454,1457,1459,1462,1466,1486,1488],[11,1453,14],{"id":13},[16,1455,1456],{},"Early in my enterprise career I worked delivery on Nestle's SAP Globe program in\nSwitzerland, one of the largest global SAP template efforts in consumer goods. This is\nwhere I learned what enterprise scale actually costs, and why the systems the business runs\non are the hardest ones to change.",[11,1458,22],{"id":21},[16,1460,1461],{},"A global manufacturer running different processes in every country is a business fighting\nitself: no common view, no shared data, no way to move a practice from one market to another.\nThe Globe answer was a single SAP template rolled out worldwide. The difficulty is never the\ntemplate in the abstract. It is reconciling one intended way of working with the reality of\ndozens of countries that each do it differently, and doing it without stopping the business.",[11,1463,1465],{"id":1464},"the-work","The work",[31,1467,1468,1474,1480],{},[34,1469,1470,1473],{},[37,1471,1472],{},"Global template rollout."," Deploying a common SAP core across countries, holding the line\non the standard while absorbing genuine local difference.",[34,1475,1476,1479],{},[37,1477,1478],{},"Application management and testing."," Owning the delivery discipline that keeps a rollout\nof this size stable, where a regression is not an inconvenience, it is an operational risk.",[34,1481,1482,1485],{},[37,1483,1484],{},"Scale as the teacher."," Working at the size where process, data and people all have to move\ntogether, and where nothing succeeds on cleverness alone.",[11,1487,122],{"id":121},[16,1489,1490],{},"Before the AI work, this is the grounding: enterprise programs where the technology had to\nwork at global scale and the cost of getting it wrong was measured in a running business. It\nis the same respect for the load-bearing core that shapes how I approach modernization now.",{"title":127,"searchDepth":128,"depth":128,"links":1492},[1493,1494,1495,1496],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":1464,"depth":131,"text":1465},{"id":121,"depth":131,"text":122},"Satyam, Nestle Switzerland","Delivery on Nestle's SAP Globe program, one of the largest global SAP template rollouts in the FMCG world, rolling a common core out across countries.","One template rolled out across a global FMCG operation",{},"\u002Fwork\u002Fsap-globe-nestle","Consultant and Delivery Lead","Global template, multi-country",{"title":1449,"description":127},"sap-globe-nestle",[150,1507,1508,1509,1510],"Global Template","Rollout","Application Management","Testing","work\u002Fsap-globe-nestle",[160,161,1513,162],"delivery","2007","0HMQR53z-mFdkCqudsJovhCzWckKysqCGNlb-IJ115c",{"id":1517,"title":1518,"body":1519,"context":1570,"description":127,"domain":1571,"draft":138,"extension":139,"hero":1572,"impact":1573,"meta":1574,"navigation":143,"path":1575,"role":1576,"scale":1577,"seo":1578,"slug":1579,"stack":1580,"status":485,"stem":1586,"tags":1587,"year":1589,"__hash__":1590},"work\u002Fwork\u002Fport-digitalization.md","Digitalization of India's Major Port Trusts",{"type":8,"value":1520,"toc":1564},[1521,1523,1526,1528,1531,1533,1559,1561],[11,1522,14],{"id":13},[16,1524,1525],{},"As a nodal officer with the Ministry of Shipping, Government of India, I led digital\nadoption across major Indian port trusts including Visakhapatnam, Chennai and Jawaharlal\nNehru (JNPT). This was my first leadership posting, and it was at the scale of national\ninfrastructure, with progress reported quarterly to the Prime Minister's Office.",[11,1527,22],{"id":21},[16,1529,1530],{},"A nation's ports are its trade choke points. When port operations run on paper, the cost is\nnot local inefficiency, it is friction on the whole economy: slow clearance, poor visibility,\nand no common language between the many parties a single shipment touches. Digitalizing a port\nis not installing software, it is coordinating operators, customs, shipping lines and freight\nforwarders around new systems and shared standards, under public accountability.",[11,1532,1465],{"id":1464},[31,1534,1535,1541,1547,1553],{},[34,1536,1537,1540],{},[37,1538,1539],{},"Port operations systems."," Oversaw implementation of terminal operating systems, gate\nmanagement and yard management across the port trusts, working closely with the vendors who\nbuilt them.",[34,1542,1543,1546],{},[37,1544,1545],{},"Standards, not just software."," Served on the committee designing EDIFACT-based standards\nfor shipping and cargo operations in India, so systems across ports could actually talk to\neach other.",[34,1548,1549,1552],{},[37,1550,1551],{},"Many stakeholders, one system."," Coordinated port officials, customs, shipping lines and\nfreight forwarders, the constituency that has to agree for a digital port to work.",[34,1554,1555,1558],{},[37,1556,1557],{},"Public accountability."," Submitted quarterly progress on port digitalization directly to\nthe Prime Minister's Office.",[11,1560,122],{"id":121},[16,1562,1563],{},"Two decades before the AI work, this set the altitude: delivering technology at the scale of a\nnation's infrastructure, where the stakeholders are institutions and the accountability runs to\nthe top of government. It is where I learned that the hardest part of a large system is rarely\nthe code.",{"title":127,"searchDepth":128,"depth":128,"links":1565},[1566,1567,1568,1569],{"id":13,"depth":131,"text":14},{"id":21,"depth":131,"text":22},{"id":1464,"depth":131,"text":1465},{"id":121,"depth":131,"text":122},"Ministry of Shipping, Govt. of India","public-sector","Driving digital adoption across India's major port trusts, from terminal, gate and yard systems to national EDI standards, with quarterly progress reported to the Prime Minister's Office.","Terminal, gate and yard systems across major ports, reported to the PMO",{},"\u002Fwork\u002Fport-digitalization","Nodal Officer and Program Lead","National maritime infrastructure",{"title":1518,"description":127},"port-digitalization",[1581,1582,1583,1584,1585],"Port Operations","Terminal Operating System","EDI","EDIFACT","Systems Integration","work\u002Fport-digitalization",[160,1571,1588,1513],"infrastructure","2003","ODpHJ5E0zCndKM7W63nGfpRf0RvAz2jLDU41IgHdl-I",{"name":1592,"role":1593,"tagline":1594,"location":1595,"contact":1596},"Sankar Vema","AI Builder & Architect of Agentic Systems","I help enterprise leaders turn AI ambition into capability that actually ships.","India · open to global advisory engagements",{"email":1597,"linkedin":1598,"github":1599,"blog":1600,"twitter":1601},"sankar.vema@gmail.com","https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Fsankarvema\u002F","http:\u002F\u002Fsankarvema.github.io\u002F","http:\u002F\u002Fsankarvema.blogspot.com\u002F","https:\u002F\u002Ftwitter.com\u002Fsansvema",{"id":1603,"extension":1604,"items":1605,"meta":1628,"stem":1629,"__hash__":1630},"domains\u002Fdomains.yml","yml",[1606,1610,1613,1616,1620,1624],{"key":1194,"title":1607,"thesis":1608,"order":1609},"Frontier & Deep-Tech Engineering","Programs that go all the way down the stack, from language and compiler design through custom silicon and production hardware.",1,{"key":290,"title":1611,"thesis":1612,"order":131},"Agentic & Applied AI Products","Agentic and multimodal AI products architected end to end and shipped into real enterprise use, not left as demos.",{"key":214,"title":1614,"thesis":1615,"order":128},"Industrial AI & Computer Vision","Vision and machine-learning systems that run on the factory floor, the mine and the clinic, where a wrong answer has a physical cost.",{"key":998,"title":1617,"thesis":1618,"order":1619},"Autonomous Systems & Robotics","Aerial and autonomous platforms designed, built and flown, from heavy-lift and high-altitude payloads to coordinated swarms.",4,{"key":137,"title":1621,"thesis":1622,"order":1623},"Enterprise Modernization & Cognitive ERP","Modernizing the systems the business actually runs on, layering AI onto the enterprise core rather than ripping it out.",5,{"key":1571,"title":1625,"thesis":1626,"order":1627},"Nation-Scale & Public-Sector Platforms","Platforms delivered at the scale of a nation's infrastructure, under public accountability and real operational stakes.",6,{},"domains","g5JLHV2kdHGEcXRBJy02EYeE86WNAfHh460wFF2BSPs",1790601541506]