0→1 radiology platform, multi-tenant, FDA SaMD-compliant. We pivoted the MVP from speed-first to interpretability-first, and adoption followed.
Post-scan, radiologists at partner hospitals were spending hours structuring narrative findings into reports that downstream clinicians, billing, and QA all needed. Existing AI assistants were fast, and ignored. The assumption was that speed was the win. Discovery told a different story.
We ran structured interviews across three groups: radiologists, referring physicians, and hospital administrators. The pattern was unambiguous: they didn't trust opaque outputs. A model that suggested a finding without showing why got dismissed. A model 15% slower but that surfaced source imagery regions, cited prior scans, and flagged uncertainty got adopted.
"I won't sign anything I can't defend. Show me what the model saw."
Before committing the team to a full rebuild, I tested the idea on a small group of radiologists first: an early version that showed its reasoning, not just its output. It held up. Only then did we pivot the MVP spec. Interpretability first, speed second, validated on a handful of clinicians before we bet the full build on it.
The rebuilt MVP shipped with region-level evidence highlighting, confidence intervals per finding, one-click model-feedback capture, and an audit log aligned to FDA SaMD requirements. We drove NLP accuracy from 68% to 88% through structured feedback loops: radiologists flagged errors inline, and those corrections flowed back into prompt tuning within the sprint.
The compliance decision: controlled automation over full autonomy. Every AI-generated report drafted, never submitted. Clinician sign-off was the spec, not an afterthought, a clear-cut call FDA SaMD requirements made obvious from the start.
The architecture decision underneath the pivot: retrieval-augmented generation, not a model answering from memory. Every finding had to trace back to retrieved source content, the current scan and cited prior scans, never asserted from general training. In a domain where a hallucinated finding is a patient-safety issue, that scoping was the entire point of choosing RAG over fine-tuning.
The platform launched across the client's partner network. ARR surpassed $500K within six months, faster than the original velocity-first spec projected, because adoption was the constraint, not feature count. Downstream, the structured data pipeline we designed eliminated 4,000+ manual hours per year of post-processing across partner sites.
20+ clinical interviews, 15+ hospital systems, across radiologists, referring physicians, administrators.
Reframed speed-first spec into interpretability-first. Pitched to exec team; authored revised PRD.
Mapped FDA SaMD + HIPAA requirements into acceptance criteria. Partnered with compliance.
Structured feedback loops, prompt tuning guidance, edge-case taxonomy. 68 → 88% accuracy.
Interpretability beats speed in clinical AI. Radiologists adopted the one they could defend to a patient, not the faster one.
Compliance and adoption aren't separate asks. The audit log and clinician sign-off FDA SaMD required were also what built the trust that drove adoption.
Feedback loops beat one-time accuracy pushes. 68% to 88% came from making every clinician correction traceable back into prompt tuning within the same sprint.
A $1M AI initiative stopped on purpose during discovery.