PsychConnect telepsychology platform

The problem

Two common assessment inputs are slow to score by hand: projective drawing tests, where a patient draws a house, a tree and a person, and the PHQ-9 and DASS-21 questionnaires for depression, anxiety and stress. The aim was to give the psychologist a structured first draft, so their time goes on judgement, not transcription.

Data

Patient drawings as images, questionnaire scores, a short patient context, and clinical reference material. The reference material is embedded in FAISS and pgvector for retrieval. The drawing service takes a different route: the House-Tree-Person manual is short enough to pass to Gemini in full on every call, which removes the risk of retrieving the wrong passage.

Architecture

1. Extract featuresGemini Flash describes the house, tree and person drawings as structured JSON, all three in parallel.
2. Interpret each drawingGemini Pro reads each drawing with the manual and the extracted features, again in parallel.
3. SynthesiseOne final Gemini Pro call combines all three reports with the questionnaire scores into a draft for the psychologist.

Running the first two phases in parallel brings total time to about three model calls instead of seven. Around the AI service, the platform runs on Next.js and Supabase, with row-level security and role-based access on patient records, plus a Python service that matches patients with psychologists.

Clinical safety rules

Every prompt enforces nine rules. Among them: never diagnose, use cautious language, anchor every claim to a specific feature visible in the drawing, always include a limitations section, never recommend treatment, and keep patient identifiers out of the report.

Results

Results

Measure	Result
Assessment time	20–30% faster
Latency, three-phase pipeline	≈3 model calls, not 7
Validation	pilot testers

Where it breaks

• Projective drawing tests have contested validity in the research literature. The output is a set of hypotheses for a psychologist, never a finding.
• Interpretation manuals are Western-normed. The prompts flag interpretations that need cultural checking for Pakistani patients rather than applying them as written.
• A multimodal model can describe a feature that isn't in the drawing. Requiring every claim to cite a visible feature makes that checkable, not impossible.
• Validated with pilot testers, not in a clinical study.