sessional

Inspiration

Therapists often see patients once a week (or less), but emotional state can change every day. Between sessions, important signals get lost in text check-ins or never get shared at all. We built Sessional to close that gap: give patients a low-friction way to send voice updates, and give therapists a fast, structured view of what needs attention before the next session.

What it does

Sessional is an asynchronous voice check-in platform for mental health care.

Patients record a short voice memo in a guided flow.
The system analyzes both:
- what is said (semantic sentiment), and
- how it is said (prosody / vocal affect).
It detects possible divergence moments (e.g., “I’m fine” words with stressed tone).
Therapists get a clinician-facing brief with:
- 60-second snapshot
- risk level
- key themes
- divergence timeline with timestamps
- suggested opening questions
- trend indicators vs prior check-ins

How we built it

Frontend: Next.js + React + TypeScript + Tailwind
Backend: FastAPI + Pydantic + SQLAlchemy
Database: PostgreSQL (including audio storage for local MVP)
Auth: Email/password + JWT + role-based access (patient/clinician)
AI pipeline:
- Hume AI for utterance-level prosody + language sentiment
- Gemini for structured, evidence-grounded brief synthesis
Quality controls: schema-constrained outputs, guardrails, snippet normalization, fallback behavior when external APIs fail

Challenges we ran into

Designing a clinically useful output without over-claiming or hallucinating
Balancing summary richness with strict evidence-grounding
Handling real-world audio variability and short/non-ideal utterances
Integrating multi-step async processing (upload → extraction → synthesis) with clear UX
Keeping role-based routing/auth secure while moving quickly in MVP mode
Making divergence explanations understandable to therapists, not just technically correct

Accomplishments that we're proud of

Built a full end-to-end MVP from recording to therapist dashboard
Implemented true semantic-vs-prosody divergence logic (not just keyword matching)
Improved report quality with grounded prompts + post-generation guardrails
Added trend logic across check-ins (risk/frequency/intensity direction)
Shipped practical therapist UX touches: quick-view modal, full brief page, report deletion
Added tests around normalization, divergence logic, and guardrail behavior

What we learned

In mental health tooling, trust comes from clarity and restraint, not flashy generation
Structured outputs + repair layers significantly reduce hallucination risk
Small UX details (wording, color semantics, snippet context) strongly affect clinician confidence
Voice carries critical emotional information that text-only check-ins can miss
Building modular provider layers early makes future model swaps much easier