Greptile co-founder on building an AI code review agent: model inversion catches 10% more bugs
AI Engineer · youtube · 2026-10-26
AI Engineer interview with Daksh Gupta, co-founder of Greptile (which reviews 1% of all GitHub PRs), on building a code review agent:
- Baseline problems: a simple sandbox-plus-webhook setup misses bugs, costs $3-25 per PR, and floods users with nitpicks.
- Fix 1: targeted prompts to cut nitpicks.
- Fix 2: a dataset of past incident PRs.
- Fix 3: model inversion — reviewing Claude-written code with GPT and vice versa catches 10% more bugs ("don't let AI grade itself").
- Fix 4: codebase context plus docs, tickets and a living wiki.
- Fix 5: autonomous testing in sandboxes using computer use.
- Fix 6: route cheaper models to most subtasks to control cost.
More from coding & agent
- You can just ask Claude Code to write a plugin for smart compaction and indefinite session continuity — repligate · 2026-10-26
- $10-12K build for local agentic coding: 1x 5090 vs 4x 3090 vs Mac Studio M5 Ultra — Shadmelor · 2026-10-26
- 4 problems hit when MCP tools started charging per call via x402 — SWARM_operator · 2026-10-26
- I counted a week of x402 payments on Base: 431K calls, only 18 sellers with 10+ repeat buyers — SWARM_operator · 2026-10-26
- Dev hot take: if you're still using PRs, you're behind — TAbrodi · 2026-10-26
- Blogger ditches Codex for Claude to rebuild personal blog after 'hallucinating' fixes — sujingshen · 2026-10-26