All case studies
ResearchML Infrastructure2025 — present
SynapseML
Evaluation-first toolkit for training, comparing, and shipping small models.
Role · Lead engineer
Overview
A workflow library that treats evaluation as the primary artifact of a training run, not an afterthought.
Problem
Small-model iteration cycles waste time on bespoke training scripts and shallow eval suites that miss regressions.
Solution
Opinionated config-driven training with first-class evaluation, comparison reports, and reproducible artifacts.
Tech stack
- Python
- PyTorch
- Weights & Biases
- Hydra
Engineering notesWhat's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
Engineering notes
What's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
Design goals
- Evaluation reports are the primary run artifact.
- Every run is reproducible from a config file alone.
- Comparing two runs is a first-class operation, not a spreadsheet.
Philosophy
You can't improve what you don't measure. The toolkit's job is to make measurement the default, not the ceremony.
Planned architecture
Hydra-driven config surface; PyTorch training loop with pluggable eval callbacks; artifacts and comparison reports pushed to Weights & Biases.
Current stage
Design phase — API surface and eval contract.
Links
Repository and demo links will appear here once the project is ready for public review.