AURA
Four fine-tuned classifiers across three datasets, and the evaluation harness that let 22 people work on them in parallel without stepping on each other. ML Project Lead at UTMIST.
The technical scope was well-defined — fine-tune classifiers on three datasets — but the coordination scope wasn't. Twenty-two contributors iterating on the same pipeline will silently produce four incompatible splits, three different metric definitions, and no way to tell whether a change helped.
The classifiers were the deliverable. The pipeline was what made a team of that size honest about them.
I owned the shared pipeline across all three datasets: ingestion, stratified splits that preserved class balance, per-class metrics rather than aggregate accuracy, and a versioned baseline comparison so every experiment had a reference point.
On top of that, an automated evaluation harness combining three signals: an LLM-judge scoring model outputs against a rubric, pairwise win-rate comparisons between checkpoints, and reference-free sanity metrics as a cheap regression tripwire. Any two of the three disagreeing was the flag to investigate.
The pipeline was the coordination surface. Once ingestion, splits, and evaluation were shared, sub-teams could iterate on individual classifiers independently and the results were comparable at the end.
Four classifiers shipped, evaluated against a shared baseline. Presented at MIT IEEE URTC in October 2025.
The honest takeaway isn't a benchmark number — it's that the coordination infrastructure held. Twenty-two people converged on comparable results because the pipeline enforced it, not because the process did.
Framing note: AURA is a support system, not a diagnostic or therapeutic tool. No clinical claims are made or intended.