I'm an AI Engineer Lead at Syngenta, based in Pune. I design and ship LLM-powered products — RAG pipelines, evaluation loops, and the MLOps plumbing that keeps them running in production. My focus is AI/ML — powered by deep full-stack engineering roots. That combination lets me bridge the gap between a working model and a product people can actually rely on. The model is the easy part. The hard part is everything around it — retrieval, evals, failure modes, cost, and the team that maintains it six months from now. That's what I write about.

I'd rather show working systems than list skills. Everything here runs and is inspectable — the eval scores below are re-measured and republished by CI on every push, not typed in by hand:

Working system	What it demonstrates	Source
FrameSleuth agent	The open-source MCP server behind FrameSleuth — local video → structured Context Bundle, no frames leaving the machine. I wrote the initial commit and most of what followed.	repo · site
ask-santosh	Retrieval-augmented Q&A over my writing, gated in CI two ways: a key-free retrieval gate (recall@k · MRR over a labelled corpus) that runs on every push, and a DeepEval answer suite (faithfulness · answer-relevancy · contextual-relevancy) when a judge key is configured.	repo · gate
Multi-agent on Databricks	Production multi-agent orchestration, beyond Genie code	repo · write-up
This profile	The section below is regenerated weekly from my real GitHub activity by a scheduled workflow. Its footnote states how each run was produced, so the claim can't drift from what actually ran.	workflow · script

• graph-over-gold-agent — Graph over Gold: governed multi-hop routes vs naive text-to-SQL, end to end (+3 more)
• kappa-iot-temperature — chore: add Apache-2.0 LICENSE
• delta-live-table-databricks-sample-use-case — chore: add Apache-2.0 LICENSE
• delta-lake-databricks-sample-use-case — chore: add Apache-2.0 LICENSE
• santoshshinde2012 — docs(readme): correct FrameSleuth credit, refresh stale content (+5 more)

Built from my last 21 days of public GitHub activity · workflow · updated Aug 24, 2026

"Turn any video into code your agent can ship."

FrameSleuth is a local-first AI system that converts screen recordings into structured, evidence-cited context bundles for coding agents. Record a bug or a feature demo — it reads every frame, transcribes the narration, and captures console and network activity, then hands your agent repro steps, error evidence, and code candidates it can act on.

I build and maintain framesleuth-agent — the open-source Python MCP server underneath it. Local video in, structured Context Bundle out, over MCP. It's public, so you can read the pipeline rather than take my word for it: keyframe extraction, transcript alignment, bundle contract, Docker image built from a pinned lockfile in CI.

• End-to-end systems — not the model in isolation, the full pipeline. Most of the value (and most of the bugs) live between the boxes on the architecture diagram.
• Tradeoffs that matter to the business — Lakebase vs Lakehouse, batch vs streaming, RAG vs fine-tuning. These show up in latency, cost, and risk — not just engineering preference.
• The unglamorous production work — eval harnesses, observability for non-deterministic systems, drift, and guardrails. The stuff that separates a demo from something you can trust on a Tuesday morning.

Shipping something is one thing. Keeping it working for five years, while people build on it, is a different discipline — and it's the half of "Lead" that a commit graph doesn't show.

• Five years of upkeep, not a weekend project. node-boilerplate has been maintained since March 2021 and is still on a current Node CI matrix, wired to SonarCloud, Snyk, CodeQL and njsscan. 82 forks depend on it staying correct.
• Dependencies stay current. 274 merged upgrade and security PRs — almost all automated, which is the point: the pipeline does the work and I keep it unblocked.
• Breadth across a team, not one repo. Beyond my own projects I contribute across thestackhub1 — the FrameSleuth agent, a voice agent, a logistics platform, and the web surfaces around them.

Being straight about the limits: this is sustained maintenance and dependency discipline, not a large contributor community. One external contributor, three issues ever. The reviewing half of leading happens in private repos at work, where a public graph can't see it.

A decade of building — from data pipelines and full-stack apps to production AI, with a habit of sharing what I learn along the way.

%%{init: {'theme':'base','themeVariables':{'fontFamily':'Fira Code, monospace','taskBkgColor':'#1F6FEB','taskBorderColor':'#2F81F7','activeTaskBkgColor':'#2F81F7','activeTaskBorderColor':'#58A6FF','doneTaskBkgColor':'#8B949E','doneTaskBorderColor':'#6E7681','critBkgColor':'#D97706','critBorderColor':'#B45309','critTextColor':'#ffffff','milestoneBkgColor':'#238636','milestoneBorderColor':'#2EA043','taskTextColor':'#ffffff','taskTextDarkColor':'#ffffff','taskTextLightColor':'#ffffff','taskTextOutsideColor':'#8B949E','gridColor':'#30363D','todayLineColor':'#F85149','titleColor':'#8B949E','sectionBkgColor':'#2F81F714','altSectionBkgColor':'#2F81F70A'}}}%%
gantt
    title A decade of building — 2014 to today
    dateFormat YYYY-MM-DD
    axisFormat %Y
    tickInterval 1year

    section Craft
    Data engineering (pipelines, Spark, Databricks)      :active, de,   2014-01-01, 2026-08-24
    Full-stack engineering (TypeScript, Node, React, AWS) :done,   fs,   2014-01-01, 2021-06-01
    Architecture and platform engineering                :active, arch, 2018-01-01, 2026-08-24

    section AI / ML
    ML and data science                                  :active, ml,   2021-06-01, 2026-08-24
    LLM, RAG and agentic systems                         :active, llm,  2023-06-01, 2026-08-24
    MLOps, evals and guardrails                          :active, ops,  2023-06-01, 2026-08-24

    section Building in public
    Answering on Stack Overflow                          :crit, so,   2015-01-01, 2026-08-24
    Open source on GitHub                                :crit, gh,   2016-01-01, 2026-08-24
    Writing on Medium                                    :crit, med,  2019-01-01, 2026-08-24

    section Milestones
    AI Engineer Lead at Syngenta                         :active,    syn, 2024-05-01, 2026-08-24
    FrameSleuth launch                                   :milestone, fsl, 2025-09-01, 0d

%%{init: {'theme':'base','themeVariables':{'pie1':'#2F81F7','pie2':'#238636','pie3':'#8B949E','pie4':'#D97706','pieOpacity':'1','pieStrokeColor':'#ffffff','pieStrokeWidth':'2px','pieOuterStrokeColor':'#8B949E','pieTitleTextSize':'18px','pieTitleTextColor':'#7A7A7A','pieSectionTextSize':'13px','pieSectionTextColor':'#1F2328','pieLegendTextColor':'#7A7A7A'}}}%%
pie showData
    title Focus areas, right now
    "AI agents, LLM and MCP systems" : 40
    "Data and AI on Databricks" : 25
    "Full-stack product engineering" : 20
    "Technical writing" : 15

Not every problem lives in the same place. I map what I build to the Cynefin domains — because the right approach for a known CRUD API is the wrong approach for a non-deterministic agent.

AI, Agents & Products

Data Engineering & Platform

Project	What it is
medallion-architecture-databrics	Medallion Architecture — principles and a practical Databricks exploration. Read →
dataset-atlas	A map-first way to discover and download datasets — Region → Domain → Get. Live demo →
node-boilerplate	Production-ready Node.js + TypeScript skeleton for microservices — ESLint, Prettier, Husky, CI wired in.
genie-reality-check	Chart data and methodology behind my Databricks Genie evaluation series.
delta-live-table-databricks-sample-use-case	Declarative pipelines with Delta Live Tables — expectations, lineage and incremental loads.
delta-lake-databricks-sample-use-case	Delta Lake from first principles — ACID on object storage, time travel, schema evolution.
kappa-iot-temperature	Kappa architecture for real-time IoT temperature monitoring — one streaming path, no batch twin.
next-genie	A Next.js front end over Databricks Genie.

• Medallion Architecture Has a Write-Back Problem
• A Knowledge Graph Over Gold: What Multi‑Hop Questions Need, and How to Build It Inside Databricks
• Stack Convergence vs. Cloud Sprawl: The Cost of Unified AI/BI
• Why Your Churn Model Is Not a Risk Model (And Why It Matters)
• Databricks Genie: Agent Mode vs Chat Mode — Should You Switch Yet?
• Feature Store vs. Gold Data Products: Where Should Your ML Features Live?
• What Genie Ontology Actually Automates, and What It Leaves to You
• Data Governance on the Ground — and the Challenges AI Agents Bring

I publish most of this through AI That Ships — my Medium publication on getting AI systems into production, also on Substack. More on Medium →

Most of my work is in private repositories — 1,893 of my last 2,027 contributions, about 93%. These are recounted by the weekly workflow, not typed in by hand.

What I'm up for.
Production LLM systems that have to survive contact with real users — retrieval that holds up, evals that fail loudly, and the cost and failure-mode work underneath. Happy to talk shop, review an eval harness, or compare notes on agent tooling.