Inspiration
1 billion people live with untreated mental illness. The psychiatrist-to-patient ratio in low-income countries is 1:500,000. And the therapists who do exist spend 30–40% of their clinical time not with patients — writing notes.
The bottleneck isn't compassion. It's bandwidth.
Structured assessments get filled from memory, not from what the patient said. Subtle risk cues — a passing mention of hopelessness, a reference to means — get missed not from negligence, but from cognitive overload. And in India specifically, an entire layer of clinical signal gets lost when Western diagnostic frameworks meet patients who express distress through somatic complaints, family-system language, and cultural idioms that don't map cleanly to DSM criteria.
We built TherapAI Scribe to fix the infrastructure, not the clinician.
What It Does
TherapAI Scribe is an AI clinical co-pilot. Upload or paste a session transcript — the system returns:
Traceable structured assessments. Every PHQ-9 and GAD-7 item rated and anchored to the exact transcript passage that drove it. Click any item — see the source. No opaque AI outputs.
SOAP notes with direct-quote Subjective sections. Formatted for clinical records, editable before approval, with provisional diagnostic impressions and severity markers in the Assessment section.
Real-time risk flagging. Columbia Protocol-aligned detection of suicidal ideation, self-harm intent, and acute distress — graded by severity (passive ideation → active plan → means access), each linked to the triggering passage. A mandatory disposition workflow means no flag can be silently dismissed.
India-specific cultural adaptation. Somatic distress expressions mapped correctly to PHQ items. Supernatural attributions handled with clinical nuance, not pathologized. Risk flags calibrated for Indian-specific precipitants. Hinglish and code-switched transcripts processed natively.
Session analytics. Talk-time split, emotional valence arc, theme tracking session-over-session, therapeutic technique detection, and a supervision-ready summary.
Group therapy & substance abuse mode. One transcript → individual SOAP notes per member, anchored only to that member's utterances. AA/NA support group mode generates meeting summaries without inappropriate clinical records.
Pharmacological signal extraction (psychiatrists only). Prescribing-relevant clinical signals surfaced from transcript, filtered against India's NLEM formulary. No prescriptions generated. Every decision stays with the clinician.
Nothing is finalized until the therapist clicks Approve & Save.
How We Built It
Built in 3 hours. Architecture kept deliberately lean.
- Frontend: Vite + React + Tailwind CSS — split-screen review UI, transcript left, AI outputs right
- Transcription: Groq Whisper — fast audio-to-text with speaker diarization for group sessions
- Intelligence: Claude claude-sonnet-4-20250514 via Anthropic API — SOAP generation, PHQ/GAD mapping, risk detection, cultural adaptation, analytics, pharmacological signal extraction via structured chained prompts with JSON-schema outputs
The core engineering insight: treat citation as a constraint, not a feature. Prompts instruct Claude to refuse to rate any assessment item without quoting its transcript anchor verbatim. This makes traceability enforceable — not aspirational.
Challenges
Traceability at output level. Free-text AI outputs couldn't be audited.
Switching to structured JSON with mandatory transcript_anchor fields on every
rating made accountability enforceable.
Cultural nuance without stereotyping. The supernatural belief handler required a three-mode system — document / cultural context / MSE concern — rather than a binary flag. The line between culturally normative and clinically significant is not a toggle.
Risk detection without liability exposure. Flagging suicidality without a structured response pathway creates liability, not safety. Every flag requires explicit therapist disposition before session close. The flag cannot be dismissed silently.
Drug suggestion scope discipline. "Suggest a drug" is not a feature — it's practicing medicine without a license. Rebuilt as pharmacological signal extraction for psychiatrists only, with NLEM formulary filtering. Medication decision notes, not prescriptions.
What We Learned
The hardest part of building for healthcare is not the AI. It's the accountability architecture around the AI — who sees what, who must act before what, what can never be auto-approved.
Claude made it possible to hold complex, multi-constraint clinical reasoning within a single structured prompt without losing coherence. The mental health capacity crisis won't be solved by more therapists alone. It'll be solved by making the ones we have twice as effective — and by systematically catching what falls through the cracks of human attention.
Log in or sign up for Devpost to join the conversation.