|
|
|
|
COURTSIDE, mid-rally. Pose skeletons, the shuttle trail, the shot call with its confidence,
the score read straight off the scoreboard, and a 2D replay animated from the tracks.
Nothing here was labelled by hand.
How COURTSIDE scores against human annotators
Thresholds were tuned on one match and tested untouched on a second, so the held-out column is true out-of-distribution performance. Ground truth is ShuttleSet22.
| Stage | Tuned | Held out |
|---|---|---|
| Hit detection | F1 87.9 | F1 85.8 |
| Rally segmentation | F1 97.6 | F1 94.0 |
| Player → court metres | 0.57 m median | 0.64 m |
| Landing position | 0.55 m median | 1.12 m |
| Score OCR | 95.2% | 97.3% |
End to end the label-free chain reproduces 84.5% / 79.5% of annotated strokes. Player detection and shot classification come from pretrained third-party models; hit detection, landings, rally segmentation, score OCR and the analytics are mine. Full breakdown in the repo.
The AI Lab page publishes the gap between the pipeline and the human labels, per stage, rather than only the wins.
More things I've built
| Clash Royale Analytics | Scheduled ETL that has auto-committed 1,300+ times and redeployed itself since launch, unattended. Live ↗ |
| OneCall | Voice healthcare navigation. A deterministic red-flag screen runs ahead of any model call, then specialty triage over semantic retrieval. FHIR R4, real CMS provider data. |
| FoldEx | Finalist (top 5), Cornell Claude Builder Club Hackathon. Async FastAPI + RQ pipeline orchestrating Claude over Ensembl, ClinVar, gnomAD and AlphaFold. Live ↗ |
| Veto | MCP budget-guardrail server. Agents with receipts. |
Where I've worked
Cisco · Software Engineer Intern, Duo Security · Summer 2026
Migrated Duo onto Cisco's internal multi-tenant incident platform as its third tenant org, designed
and owned the end-to-end incident-response pipeline replacing the outgoing third-party tool, and
built the retrieval layer of an internal SRE assistant on AWS Bedrock.
NSF CROPPS · Software Engineer · Mar 2026 – present
React/Flask control dashboard for 57 networked research plants, 6 IoT device systems, live camera
feed at 15 FPS over Flask.
CUAUV · Software Engineer · Oct 2025 – present
Bin-search mission logic in Python/asyncio on the team's RoboSub vehicle. 7-channel RGB + surface-normal
- depth YOLO detector across 10 classes, ~20% detection improvement over RGB-only.
Stanford AIMI · Research Intern · Jun 2024
12,549 chest X-rays, RadGraph-parsed reports. Data-level downsampling took the model from 0.56 to
0.86 macro-AUROC and fixed three classes that had collapsed to 0.00 F1.
Research
- Multilabel Classification for Lung Disease Detection · arXiv:2412.11452 (2024)
- A Novel Hybrid Post-Hartree-Fock and Monte Carlo Algorithm · IJCEA / ICCME 2025. Led an 8-person research team.
Toolbox
