Shield

Inspiration

Social posts shape mood and self-worth long before likes show up. We wanted a sandbox where creators could see how a message might land emotionally — and how it could spread — before hitting publish.

What it does

SHIELD stress-tests text or video against 10,000 diverse synthetic viewers. It predicts reactions and five affect dimensions (empathy, relation, inspiration, curiosity, joy), maps them on a brain-style view, simulates emotional contagion through a share network, and lets you chat with simulated viewers in first person.

How we built it

Frontend: React + Vite — glass UI, brain sim, network graph, wellbeing analytics, persona chat Backend: FastAPI + local Whisper for video transcripts Micro ML model: PersonaPathPool (~56 KB weights) — a tiny persona-conditioned path-pool head on BGE-small segment embeddings. It pools transcript tokens with persona-aware attention, then outputs 5 reaction classes (incl. share) and 5 emotion scores per viewer Simulation: Cosine affinity picks who sees content first; sharers pass it to similar peers across contagion waves Chat: Ollama for empathy-style first-person interviews

Challenges we ran into

Making persona-specific predictions from aggregate training labels (solved with affinity-tilted targets) Keeping inference fast enough for thousands of agents in one run Getting persona chat to stay grounded in the actual transcript, not generic deflections Balancing a rich results UI without overwhelming the hero

Accomplishments that we're proud of

A sub-100 KB model that still captures “this line hits this person differently” Full pipeline: paste/upload → embed → predict → network spread → inspect → chat Wellbeing-first analytics (harm signals, polarization, segment attention) instead of pure virality chasing What we learned Small, structured models on good embeddings can beat throwing a giant LLM at every viewer. Persona vectors in the same space as transcript segments make semantic affinity a useful knob for both training and simulation.

What's next for Shield

Fine-tune PathPool on real creator feedback loops Stronger harm/risk classifiers and age-segmented cohorts Exportable “wellbeing report” for teams reviewing campaigns before launch Optional cloud personas + faster batch inference for 100k+ runs