Designs production agent infrastructure for regulated enterprise. Lead architect on every gopal client engagement; founder of Unstress.ai. Builds systems where the LLM is the last 5%, and the moat is in the rule engine beneath.
Puneet treats LLMs as a thin language surface over real reasoning. The interesting engineering — and the moat — sits in everything underneath: rule engines, SQL cross-relations across hundreds of tables, retrieval graphs, psychometric profiles, decision orchestrators. He'll cite the Layer 4 LLM as "the last 5% of the system" and mean it.
His engagements look more like systems architecture than prompt engineering. NV.AI was fine-tuning plus psychometric modelling on 500h of teaching corpus, not a wrapper. AstroAI is a 4-layer hybrid that handles crores of unique chart combinations no RAG system can replicate. PEP Compliance is a 5-source agentic researcher with confidence-scored output, not a GPT summarizer.
The throughline is replacing functions, not augmenting them. A control tower team. An astrologer roster. A compliance analyst day. A marketing pod. Each engagement removes a human function from the org chart and replaces it with infrastructure that runs at higher reliability, lower cost, and unbounded throughput.