Work
Things I built, and why.
Most of these started as a question I couldn’t answer by reading. The company is the one that got out of hand.
Athlitix
CTO & co-founderSports analytics without the sensor rig.
Athlitix exists to make performance analysis something anyone can do from a phone — no wearables, no markers, no instrumented facility. RinkUp is our first product, and figure skating turned out to be an unreasonably hard place to start.
RinkUp on the App Store ↗athlitix.com ↗
The hard parts
- A jump lasts under 0.7 seconds. The skater opens, pulls tight, and lands — and nobody at the boards can honestly tell you whether that was three rotations or an optimistic two and a half.
- Practice ice is crowded. Twenty skaters, near-identical dress, and exactly one of them is yours.
- Mid-air, pose confidence collapses. A tracker left to its own judgement will cheerfully switch to whoever is standing in the background.
- Ice has no markings. No pitch lines, no stadium geometry, no reference frame to calibrate against — so the measurements come out of a physics model instead.
Open source
github.com/agirishkumarDCT-VisionImage processing that never decodes the image. Operates directly on JPEG DCT coefficients, skipping the pixel round-trip that every other library pays for.2026 · PyPIAutoDataPreprocessAutomatic preprocessing for tabular data — the tedious 80% of every project, packaged and published.2024 · PyPIMedScriptDiagnostic report generation over MIMIC-4 with retrieval, built as an MLOps course project at Northeastern. Airflow, Qdrant, PubMedBERT embeddings, Llama3-Med42 and Gemini on Vertex AI — with bias analysis and 97% test coverage, because a medical model that isn’t audited isn’t finished.2024 · Group projectTransformer From ScratchAttention, positional encoding, and the training loop, written out longhand.2024 · From scratch