Enzo Incutti
I build systems that catch bugs before humans do. Currently obsessing over code review agents at Recurse — think git commit that knows you're about to ship a bug.
What I'm Building
At Recurse, I work on infrastructure that catches bugs before they ship. The vision: your IDE knows you're introducing a critical bug the moment you type the wrong keystrokes — not after three rounds of code review.
Agent Infrastructure
- Core agent architecture with context management, tool calling, structured outputs, and observability
- Evaluation pipelines integrated with Weights & Biases for continuous model improvement
- Elasticsearch-powered knowledge base with vector embeddings for semantic code search
Production Systems
- CI/CD on GitHub Actions, GCP CloudBuild, and CloudRun
- GitHub App with FastAPI webhooks for real-time PR monitoring
- Redis-based async processing for scalable background analysis
- Terminal UI with click for in-editor code review
Projects
GPT-4 bot that generates contextually-aware YouTube comments. Built to explore how LLMs handle conversational dynamics in adversarial environments.
Automated settlement system on BSV blockchain. Smart contracts for real-time collateral management and margin calls.
Artist royalty distribution on XRPL. Automatic splits and transparent on-chain payment tracking for creators.
Bayesian optimization for biological simulations. Efficient parameter search for agent-based bacterial models.
Research That Shapes My Work
-
On Usefulness of Deep-Learning-Based Bug Localization Models
The paper that convinced me automated code review was possible.
-
RepoBench: Benchmarking Repository-Level Code Auto-Completion
How I think about evaluating code understanding at scale.
-
Alibaba LingmaAgent: Automated Issue Resolution
State-of-the-art in agent-driven codebase exploration.
-
CoRNStack: Contrastive Data for Code Retrieval
Better embeddings = better semantic search.
-
CodeT5: Unified Pre-trained Encoder-Decoder for Code
Foundation for understanding code as structured data.