"We had great ML engineers but no strategic direction. Apoorve gave us the AI product roadmap and the governance framework that turned our prototype into something we could actually take to market — and to regulators."
The Challenge
A digital health startup had built an impressive NLP model for extracting clinical insights from physician notes. The model showed strong research performance, but the founding team — talented ML researchers without a commercialization background — faced several critical blockers:
- No clear regulatory strategy for FDA 510(k) or Software as a Medical Device (SaMD) classification
- Patient data governance was informal, creating HIPAA exposure
- Model versioning was non-existent: production models were being updated without systematic validation
- Product-market fit was undefined — the model solved a technical problem, but the business use case was under-specified
- Investor conversations stalled because the team couldn’t articulate a defensible AI moat
The Engagement
This was a pure advisory engagement: 12 hours per month over 10 months, combining strategic direction, technical architecture review, and investor preparation.
AI Product Strategy
The first 60 days focused on use case prioritization. Through structured discovery sessions with target clinical buyers, we identified the highest-value applications:
- Clinical documentation assistance (primary) — reducing physician documentation burden
- Care gap identification (secondary) — flagging patients with unaddressed chronic condition indicators
- Retrospective analytics (de-prioritized) — too far from core workflow, deferred to v2
This clarified the product roadmap and created a defensible wedge: documentation assistance had clear ROI, fast feedback loops with physicians, and a more straightforward regulatory path than diagnostic-adjacent use cases.
Regulatory Architecture
Worked with a regulatory counsel to map the product to FDA SaMD guidance. Key decisions:
- Positioned as Class II Clinical Decision Support (not diagnostic software) to qualify for 510(k) pathway
- Designed a human-in-the-loop architecture that kept AI recommendations advisory, not autonomous — a critical distinction for both regulation and clinical trust
- Established labeling requirements and post-market surveillance processes required for FDA registration
ML Governance and Infrastructure
Built the governance layer the team lacked:
- Implemented MLflow for experiment tracking and model registry
- Established validation protocol: every model update requires holdout evaluation on a demographically stratified test set before promotion to production
- Designed differential privacy approach for model training to reduce re-identification risk in compliance with HIPAA Safe Harbor
- Created model cards for each production model documenting training data, performance metrics by subgroup, known limitations, and intended use
Performance Improvement
Reviewed the model architecture and training data strategy. Key findings:
- Training data was skewed toward one EHR system, creating distribution shift in deployment
- Entity recognition was being handled with a generic NLP model not fine-tuned on clinical text
Recommended fine-tuning on ClinicalBERT with a curated, multi-site training set. Over three model iterations, F1 score improved from 71% to 89%.
Results
The company successfully launched their clinical documentation product with 6 hospital pilot customers. The FDA registration was achieved without a full 510(k) submission (exempt under current CDS guidance), significantly de-risking the regulatory timeline. The Series A closed at $8M with the AI platform cited as a primary differentiator by the lead investor.
Key Lessons
Regulatory strategy is a product design constraint: The human-in-the-loop architecture wasn’t just a regulatory concession — it was the right product design for clinical adoption. Clinicians trusted a tool that showed its reasoning and invited their judgment.
Use case clarity unlocks everything: The founding team’s instinct was to build a broad clinical AI platform. Narrowing to one high-value use case first gave them focus, faster feedback, and a cleaner story for investors.
AI governance is a commercial asset: What felt like compliance overhead became a selling point with hospital procurement teams, who had been burned by unaccountable AI vendors before.