Five years building applied ML and about 21 months on LLM systems across manufacturing, supply chain, healthcare, and education. I ship RAG and agent products — hybrid retrieval, fail-closed ACLs, citation grounding, eval harnesses . Production computer-vision, NLP and time-series models I built are used by Mahindra, Flexera, and Dorman.I take a problem from data to a FastAPI service.
Manaswi — AI co-learner and teacher assistant. Mock interviews, Feynman-style concept checks, assignment/test design, classroom translation, and progress tracking. Plagiarism review is a class-local lexical detector, not a web-scale “% copied” score: word 5-grams, containment (not Jaccard), and an adaptive window W(n) = clamp(floor(n/3), 30, 100) with a higher match bar on short CBSE answers. Scoring is closed-form (no HTTP in the scorer); Sarvam-105B only narrates spans already found. Evidence for the teacher, not a verdict. How it works
ClinAssistIndia — PHC case workspace (POC, rural India). A Hinglish / Hindi / English note becomes an ICMR-grounded clinical card: urgency, ICD-10, PHC-feasible steps, referral. RAG over official ICMR Standard Treatment Workflows (Chroma MiniLM → Sarvam-105B). Human-triggered ops for beds, ambulance, pharmacy, expert, SOS. Silent FHIR R4 bundles on disk. Decision support only — treating judgment stays with the clinician. Project secured 3rd place, LatentForce 48-hour Build Sprint 2026 (28–30 Aug).
Enterprise Knowledge Assistant — Agent-free Azure RAG over HR / Finance / IT / Legal / Sales policy. Hybrid BM25 + vector search, fail-closed ACL filters, evidence gate, citation allowlist. ~40-case golden set on Azure AI Foundry: groundedness ~4.6/5, relevance ~4.5/5, citation_ok 1.0. Live demo. Ingest is structure-first, then length: split on policy headings and Excel sheets, then C(n) = clamp(floor(n/α), 128, 512). On the live Northwind pack (112 chunks / 11 files) the proportional band is empty — what shipped is one heading, one chunk, plus atomic sheets. Length-adaptive windows are implemented for longer sections; they did not produce this index. Write-up
Shipped for Mahindra, Flexera, Dorman, and Corning. Intern → Associate Data Scientist → Data Scientist. Out-of-cycle promotion for the Mahindra NLP work.
| Client | What shipped | Result |
|---|---|---|
| Manaswi · plagiarism | Class-local 5-gram detector with length-adaptive windows and τ. Stylometry is a side channel and cannot move the risk band. LLM explains; it does not score. | Labelled synthetic CBSE physics (200 + 500): verbatim 1.00, near-verbatim 0.99; paraphrase miss by design; independent overlap ~0.3–0.6. Part II |
| ClinAssistIndia · LatentForce Build Sprint | ICMR-grounded PHC case workspace: Hinglish / Hindi / English note → clinical card (urgency, ICD-10, PHC-feasible steps, referral). RAG over official ICMR STWs. | 3rd place, 48-hour Build Sprint 2026 (28–30 Aug) |
| Enterprise Knowledge Assistant · ingest | Structure-first policy chunker: do not merge short headings to fill 512 tokens; keep sheets whole; prefix vectors, leave BM25 unprefixed. C(n) is implemented. |
112 chunks from 11 files (98 section / 4 sheet / 10 window). Length-adaptive band unused on this short pack. Part I |
| Mahindra Research Valley, Chennai | NLP retrieval that extracted root causes from English mechanical-failure writeups | ~60% average reduction in ORC closing time |
| Mahindra plants | Computer-vision sealant-fault detector on poor-quality line images | VAPT reported ~90% increase in fault detection without stopping the line. Recognized by the VP, Mahindra Emerging Tech Division, Mumbai |
| Flexera / Dorman | Time-series license-demand and units-sold forecasts (holidays, attrition, sparse history, overlapping seasonality, concept drift) | Production forecasts on non-stationary demand |
| Mahindra shop floor + tenders | Real-time YOLO proximity alerts; Azure Custom Vision ANPR retrained for graffiti, occlusion, and degraded Indian plates; Azure OCR (Read API) + clause-level tender diffs on App Service | Live CV + document-integrity review |
| justuju.in · EvalAI | Automated evaluation of student submissions (flowcharts → code, code → Mermaid, Math/Physics tests). The same junior team of 7–8 used the tool instead of grading by hand | ~50–60 submissions/day manually → ~1,000/day with the tool; ~98% accuracy on handwritten flowchart → Python |
| Role | Where | When | One-liner |
|---|---|---|---|
AI Engineer and Creator  |
Manaswi · ClinAssistIndia | Apr 2026 – Present | Independent product work: Sarvam-105B co-learner and ICMR-grounded PHC workspace |
| AI Engineer | justuju.in | Jan 2025 – Mar 2026 | AI consultant and SME; EvalAI for ed-tech assessments |
| Career break | — | Oct 2023 – Oct 2024 | Family and travel; ICPC India AI plagiarism-checker POC |
| Data Scientist | Bristlecone | Jan 2020 – Oct 2023 | Intern → Associate DS → DS; NLP, CV, and forecasting for Mahindra, Flexera, Dorman, Corning |
Languages & ML — Python, SQL, PyTorch, scikit-learn, pandas, NumPy, OpenCV, YOLO, NLTK
LLM systems — RAG, hybrid BM25 + vector search, agentic workflows, evaluation, guardrails, token/cost telemetry, DSPy, LangChain, CrewAI
Platform — FastAPI, PostgreSQL, MongoDB, ChromaDB, Azure (OpenAI, AI Search, Foundry, App Service), AWS
Writing down how the pieces actually work:
- How Manaswi detects plagiarism — Part II — adaptive windows, containment, measured limits
- Structure-first adaptive chunking — Part I — what shipped on the 112-chunk Northwind index vs what is still Phase II
- How RAG Works: Part I — why a plain LLM falls short, and how retrieval before generation fixes it
- How Agentic RAG Works: Part I — from single-pass retrieval to self-correcting agents
- How Claude CLI Works: Part I — sessions, context, memory limits, and where a coding agent can be attacked
B.Tech, Electrical and Electronics Engineering — Manipal University Jaipur, 2020
Open to remote work on RAG, agents, eval, and production ML.