Case {{ cs.n }} / 04 · {{ cs.org }} · {{ cs.when }}
{{ cs.title }}
{{ cs.line }}
{{ m.v }}{{ m.l }}
{{ cs.problem }}
{{ s.n }}{{ s.t }}{{ s.d }}
→{{ r }}
Vatsa Joshi
AI/ML engineer — production LLMs, vision, agents.
Bengaluru · {{ time }} ISTNow · 2× Founding AI Engineer · Founding Engineer{{ ghLine }}Scroll
01 — Selected work · 01 / 04
Four systems, all in production.
02 — Approach
I build the parts of an LLM system nobody demos: the evals, the grounding, the cost model.
Ground everythingEvery number and citation traces to a tool output or a source, or it's dropped.
Gate on evalsRegression gates, LLM-as-judge and red-teaming run before anything ships.
Engineer the cost25× cheaper extraction, 15 paise per agent query. Cost is a design input.
03 — Experience
Every checkpoint, saved.
Six roles, Jun 2025 → now. Keep scrolling along the helix.2,000 users50–60 law firms8M judgments25× cheaper0 ungrounded answers
04 — Open source · live from GitHub
{{ ghHead }}
{{ ghActive }}active days
{{ ghLongest }}longest streak
{{ osRepos }}public repos
{{ osLangs }}languages
{{ ghRange }}github.com/Vatsa10 ↗
Languages · by repo
{{ l.name }} {{ l.pct }}
Recently shippedAll projects →
05 — Adversarial eval
Crank the learning rate.
Press and hold anywhere. Push the optimiser until training diverges and the model leaks its training data, which is my résumé.
Diverged · training data leakedv2026.10
Vatsa JoshiAI Engineer · Multi-Agent Systems · Production LLM Platforms
{{ r.org }} · {{ r.role }}{{ r.when }}→ {{ r.imp }}
Agents OpenAI Agents SDK, LangGraph, BAML, CrewAI, FastMCP · LLM systems Hybrid RAG, GraphRAG, evals, guardrails · Training PyTorch, PEFT/TRL, LoRA/QLoRA, Unsloth, DeepSpeed ZeRO · Infra Python, TypeScript, Rust, FastAPI, Next.js, Postgres, OpenSearch, Qdrant, AWS
B.Tech CS, Navrachana University · IEEE Xplore: deepfake detection · 2nd place, Tinkerthon 3.0