Prateek Hitli

PilotKit

An insurance assistant that has to pass a test lab before it talks to customers. Jev decides and the LLM talks: every turn, Jev routes intent, catches prompt injection, flags PII and decides on human handoff before any tool runs, and gpt-4o-mini only extracts fields and writes the reply. Customers sign in to see their policies and file claims, shoppers compare 40 real plans, and voice input works too. In the lab, nine simulated customers (an angry claimant, a Hinglish speaker, a PII over-sharer, a prompt injector) chat with the bot and get graded against a rubric, and uncertain verdicts go to a human review queue. A turn costs $0.000359 with p95 latency of 4.23 s, and a full suite run costs about $0.02.

MKT·INTEL

An AI market research agent that sources every claim. A LangGraph team plans the research, searches and crawls the live web (Context.dev, then Tavily, then Firecrawl, with a credit breaker), extracts cited claims and has Jev make typed, confidence-scored decisions about each change, looping back when the evidence is weak. Eight playbooks, from pricing teardowns to TAM/SAM/SOM, export to PDF, PowerPoint, Excel, Markdown or JSON. Users can bring their own encrypted keys and models.

Jev Lab

Twenty-four decisions an AI product makes every day (classify, gate, route, filter, verify), each decided three ways: by Jev, by an LLM prompt and by plain code. Same labelled rows, real API calls, every mistake counted and charted as accuracy against cost per decision. Jev got 95.7% right across 486 decisions and lost 6 of the 24 projects. A full run of all 24 costs about $0.51.

I Tested a Decision Model 486 Times, Then Built a Product on It

Jev Lab runs 24 everyday AI decisions three ways: Jev, an LLM prompt, and plain code. Jev got 95.7% of them right and lost 6 of the 24 projects. MKT·INTEL is the market research agent I built on it afterwards, and it still runs without Jev.

September 29, 2026

I Built an AI Security Auditor, Then Audited It. It Was Wrong Twice.

Six agents read what three scanners found and grade a container image. Then I checked every finding against docker inspect and a raw Trivy run - and found one recommendation that would break your image and one that was three releases stale. Both are in the README.

August 30, 2026

Building the LEGO Speed Champions Fast & Furious Toyota Supra Mk4

Documenting my experience building the 319-piece LEGO Speed Champions Fast & Furious Toyota Supra Mk4 - from unboxing to the final iconic orange build.

August 12, 2026