Vishal Kumar — AI Engineer
Founding engineer at a stealth AI startup, building multi-agent systems in production. I design LLM platforms with RAG, evals, and guardrails — earlier work cut inference latency 20% and lifted retrieval accuracy 28%.
How I build AI systems
01 · Embed
Text → vectors. The right embedding model and chunking so retrieval has something good to find.
02 · Retrieve
RAG over a vector store — hybrid search and reranking. Lifted retrieval accuracy 28% on private corpora.
03 · Agent
Multi-agent orchestration: planning, tool use, and routing between specialist agents in production.
04 · Guardrail
Input/output validation, jailbreak mitigation, and schema + policy checks before anything reaches a user.
05 · Eval
Offline and online evals with regression gates, so changes are measured — not vibed.
06 · Output
Grounded, low-latency responses — 20% latency cut with strict data segregation.