~/intro.md
UTF-8Markdown
Explorer
1# Keshav Madhav
2.
3> @keshav-madhav · AI Engineer @ [VerbaFlo](https://www.verbaflo.ai/)
4.
5## TL;DR
6.
7**I build the systems behind production AI: orchestration, retrieval, observability, and developer tooling.**
8.
9- **4M chunks** Retrieval scale — from 20K+ docs across 70+ client portfolios
10- **95%** Root-cause accuracy — on internally benchmarked debugging tickets
11- **170K LOC** Platform footprint — LLM engineering platform, 236 API endpoints
12- **10K+** OSS reach — installs of Live Jinja Renderer
13.
14## What I actually build
15.
16### Agent orchestration `Runtime`
17.
18A custom plan → delegate → reflect runtime that fans work out to parallel workers across MongoDB, Postgres, and Milvus and streams cited answers over SSE. Built from scratch, not a framework wrapper.
19.
20### Retrieval and RAG `Pipelines`
21.
22Production vector retrieval: ~4M chunks from 20K+ documents, scraped content, and PMS data, serving 70+ client portfolios. Hybrid search, query expansion, reranking, and text-to-SQL on top.
23.
24### Internal tooling `Tooling`
25.
26A unified MCP debugger the whole team runs daily. Hand it a ticket and it walks code, traces, and databases to a root cause — 95% of internally benchmarked tickets, no human needed.
27.
28### Full-stack surfaces `Full-stack`
29.
30Electron apps, VS Code extensions, tracing UIs, dashboards. I build the runtime and the interfaces.
31.
32## Quick facts
33.
34- **Role:** AI Engineer @ [VerbaFlo](https://www.verbaflo.ai/)
35- **Location:** New Delhi / Gurgaon, IN
36- **Education:** B.Tech Computer Science, AI / ML specialization
37- **GitHub:** [98 repos](https://github.com/Kes...)
38.
39## The longer read
40.
41I'm Keshav Madhav, an AI engineer who builds the systems around the model: orchestration platforms, retrieval infrastructure, and the tooling that makes LLM pipelines debuggable.
42.
43## Philosophy
44.
45> If you can't see what a pipeline did, you can't fix it. Most of my time goes into tracing, replay, and eval harnesses — the unglamorous infrastructure that makes LLM systems debuggable.
46.
47---
48.
49## Navigate
50.
51- [](now.md) — Current work
52- [](before.md) — Previous experience
53- [](projects.md) — Side projects
54- [](stack.json) — Tech stack
55- [](contact.md) — Get in touch
56.
ReadyLn 1, Col 1
main