AI Lead Engineer in Pune.
I work on problems where a system can look fine and still be wrong: retrieval that misses the right context, evaluations typed in by hand, and data quality issues that hide between Silver and Gold. I would rather show a repo you can run than list skills.
LinkedIn · Medium · AI That Ships (Substack) · X · Website
- Retrieval and evaluation that can be rerun, not just described. ask-santosh measures recall and MRR in CI and publishes the latest scores.
- Lakehouse quality problems as executable examples, with a write-up next to the code. local-data-lakehouse and the Silver and Gold repair repos below are the working copies.
- Applied ML with the limits written down. crop-disease-prediction ships a model card that states where the reported accuracy stops being a useful number.
- Small tools I can run locally, with a demo when there is one: holdmap, Catalogfore, and deck-smith-local.
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.
Each row is a problem, a repo, and a piece of evidence you can open.
| Project | Problem it takes on | Proof |
|---|---|---|
| ask-santosh | A skills list does not prove that retrieval works over a real corpus. | Retrieval tests run in CI. Latest published scores: metrics · live demo |
| ai-ml-ground-truth | It is hard to learn AI and ML as a working developer without labs you can actually run. | Live book with interactive explanations and Python labs. Archived at DOI 10.5281/zenodo.23815936. |
| crop-disease-prediction | A leaf-image classifier is easy to over-claim if the number comes from a friendly dataset split. | The model card reports 97.83% validation accuracy on a held-out PlantVillage split, and explains why that number will not hold on messy field photos. The repo includes Streamlit, FastAPI, a WhatsApp path, and a React Native TFLite path, with screenshots. |
| local-data-lakehouse | Lakehouse architecture is often read about and rarely started on a laptop. | Runnable local stack (Iceberg, REST catalog, DuckDB, Spark, Trino) plus the article Stop Reading About Lakehouses. Build One Locally. |
| deck-smith-local | Business decks built from governed data need a trace back to reviewed SQL, not a one-off slide export. | Local agent that builds an editable PowerPoint from a plain-English request, with bootstrap evidence and CI including e2e and supply-chain checks. |
| churn-vs-risk-poc | A churn score is often treated as a risk score, and the two problems are not the same. | Repo plus Why Your Churn Model Is Not a Risk Model. |
| retention-radar | A renewal queue needs a score, an explanation, and a way to check whether the action helped. | Calibrated model with SHAP and a holdout, plus Building a Churn Risk Queue and a UI screenshot. |
| silver-to-gold-totals | When one Silver row changes, it is unclear which Gold totals must change, and whether a repair matches a full rebuild. | Executable Delta Lake example with a result snapshot and HTML report. Companion to One Silver Row Changed. Which Gold Totals Must Change?. |
| holdmap | It is easy to stop the wrong process when many agents and tools are holding ports. | Live demo of a CLI, TUI, desktop app, and MCP server, with screenshots in the repo. |
| Catalogfore | Databricks cost plans are hard to inspect if the rates and assumptions are hidden. | Independent, browser-local planner, with a screenshot and rate evidence in the docs. This is a personal open-source project, not an official calculator. |
| node-boilerplate | A TypeScript service starter has to stay current, not just look complete on day one. | Created March 2021 and still maintained (last push May 2026), with CI for CodeQL, Snyk, and njsscan. |
- holdmap - fix: post-#132 test pass: accurate agent roots, CPU, unknown memory and UI polish (#133) (+5 more)
- ai-providers - feat(ai): add a canonical provider protocol package and Keychain fallback tests (+4 more)
- ai-ml-ground-truth - Label the lab report's charts, and sharpen the README's short version (+5 more)
- festive-offer-engine - Link the published Databricks system-design article (+3 more)
Built from my last 21 days of public GitHub activity · workflow · updated Nov 16, 2026
I write at Medium and in the AI That Ships publication, and I publish the same weekly short pieces on Substack. The list below is filled by a scheduled workflow.
- A Knowledge Graph Is Not a Knowledge Base: How I Navigate One for AI Agents
- The Beginner’s Path to Data Engineering on Databricks
- One Silver Row Changed. Which Gold Totals Must Change?
- Why Does Every Gold Table Clean the Same Rows?
- Can You Rebuild Silver Without Reloading the Source?
- Silver Already Keeps History. Why Is Gold Building It Again?
- Bronze Loaded the New Schema. Silver Lost a Column
- Defining the AI Product Beyond the Happy Path (Part 2)
LinkedIn · X @sanshinde2012 · Substack · Medium · santoshshinde.com




