ML infrastructure that automates neural architecture search. LLM-powered code generation (GPT-4, LLaMA, vLLM), real-time Dash/Plotly monitoring via WebSocket, and a 22-table PostgreSQL backend managed by FastAPI. 521 commits, 44.7K LOC.
ML researchers waste weeks on manual architecture experimentation with no infrastructure for systematic exploration. Each variation requires hand-coding, manual deployment, and ad-hoc tracking — limiting the search space and making results irreproducible.
End-to-end infrastructure from architecture vector encoding to training runs, with meta-learning feedback loop and real-time WebSocket monitoring.
FastAPI-based infrastructure for running, tracking, and managing ML experiments with a 22-table PostgreSQL backend
Automated neural architecture search using LLM code generation and meta-learning optimization
WebSocket-powered monitoring infrastructure for ML experiments
Comprehensive API infrastructure for programmatic access to the experiment platform
Each experiment pins its dependencies and stores generated code alongside results in PostgreSQL. This guarantees reproducibility — any experiment can be re-run months later with identical environment and configuration, without relying on containerization infrastructure.
Configuration-based NAS systems are limited to predefined building blocks. LLM code generation can produce novel architectural patterns that wouldn't exist in a predefined search space. The meta-learning loop then evaluates these novel architectures and feeds results back to guide future generation.
Use the live surface or the source as the next level of proof. The goal here is not to end on a marketing flourish, but to make the next inspection step obvious.