STF Copilot

Inspiration

The Somali Family Safety Task Force (SFSTF) helps Somali women many of them recent immigrants, often on tenuous immigration status navigate domestic violence, sexual assault, and trafficking. Caseworkers face an impossible trade-off: keep a laptop screen between them and the survivor and capture a structured record, or stay present in the room and lose the data. On top of that, every cloud-based AI tool introduces real risk for this population, PII exposure, and a chilling effect on disclosure if the survivor knows her words are leaving the device.

What it does

A caseworker opens the tool on a laptop, confirms consent with the survivor, presses record, and has a normal conversation in Somali, English. Afterwards, the tool produces a complete review surface:

• Bilingual annotated transcript. Somali source on the left, English translation on the right, with sentence-level hover linking and inline paralinguistic markers ([whispering], [long_pause], [voice_tremor], [pace_slowed]) drawn from prosody analysis. Descriptive only, never interpretive.
• Structured intake form auto-extracted by a local LLM, with per-field confidence levels, evidence links back to the transcript, and mandatory manual confirmation on sensitive PII (names, DOBs, addresses, immigration status).
  
• Ranked resource recommendations matched against a curated directory of real Seattle-area organizations (Somali-speaking shelters, immigration legal aid, protection-order help) with per-card for reasoning computed from the case attributes.
  

Built With

• ctranslate2
• fastapi
• faster-whisper
• ffmpeg
• javascript
• llm
• ollama
• pydantic
• python
• qwen
• silero-vad
• sqlite