Team project, HackAI

Bifocal

Flags abusive and manipulative messages with a multi-task classifier, and offers guidance through hybrid RAG with Gemini.

Python
PyTorch
FAISS
Gemini
Flask
React
TypeScript

Code

artifacts/metrics.json

# held_out accuracy, rounded
{
  "incident_type": 0.741,
  "emotional_state": 0.685,
  "severity_score": 0.481,
  "potential_crime": 0.824
}

Held-out accuracy for each task, rounded, from the checkpoint committed to the repo.

Problem

Manipulation and coercive control can be hard to see from inside a conversation, especially one message at a time. Our team wanted a tool that could flag those patterns and suggest a useful next step.

What I built

Paste a conversation and Bifocal classifies each message, gives the conversation a 0–100 risk score, recommends next steps for flagged messages and answers safety questions from a grounded knowledge base.

The model is served through Flask to a React and TypeScript app.

How

React and TypeScript appPaste a conversation
Flask APIanalyze, recommend, ask
MiniLM encoder and task headsIncident, emotion, severity, potential crime
Hybrid RAGDense and BM25 retrieval
Risk scoringMessage and conversation, 0–100
Gemini guidanceExtractive fallback without a key

A pretrained sentence encoder feeds four task heads, trained with a class-weighted multi-task loss on 540 labeled messages. Temperature scaling calibrates the probabilities the API returns.

Guidance comes from hybrid retrieval over a digital-safety knowledge base, with Gemini generating the answer and an extractive fallback when no API key is set.

Results

74% held-out accuracy across 9 incident types, and 93% held-out on the simpler concerning-versus-safe split.
Local CPU serving: single-message p95 of about 12 ms, and about 69 ms for a full recommendation with RAG.
A 33-test pytest suite and GitHub Actions CI cover the service.

What I learned

On a small, sensitive dataset, calibrated probabilities and honest held-out numbers matter more than a headline score.
Mapping a model output to a useful next step is a product problem as much as a modelling one.

Next projectWhodunit