Inspiration
The past few years made one thing painfully clear:
Information trust is now an infrastructure problem.
Deepfakes, miscaptioned videos, AI-generated clickbait, and sensationalized reporting blur truth faster than humans can process it. Traditional fact-checking is:
Slow
Centralized
Dependent on trust in third-party servers
Often too late
Meanwhile, devices are suddenly powerful enough to run models locally.
Epistemiq was born from a simple question:
What if anyone could fact-check media privately, instantly, and offline-first?
Not as a cloud service, but as a browser app you can take anywhere — like a pocket truth lab.
What it does
Core experience
Paste text / upload media / link a video
App cleans content locally
Extracts claims
Validates claims against cached model outputs
Retrieves academic sources
Generates a readable verdict + report
Exports results to PDF
Privacy design: Speech data never leaves the device. Claims/verification only use cloud when not cached.
How we built it
Browser AI Layer
window.ai.prompt() → Gemini Nano summarization & prep
Whisper tiny.en via transformers.js (WebAssembly)
Manual 16kHz resampling to avoid silent 48k upsampling (Chrome/Safari quirk)
Graceful fallback: Whisper-only mode when Nano unavailable
Backend (minimal, PythonAnywhere)
Flask
yt-dlp + ffmpeg for safe media extraction
SQLite cache (SHA-256 text hashing)
Fact-checking pipeline User text → SHA256 hash → Check database → If cached → return cached claims → Else call cloud LLM → store & reuse
Caching Objects
Article text hash → claims
Claim hash → verdict & research questions
Claim + question hash → full report
Design ethos
Local > cache > cloud (in that order).
Challenges we ran into
Browser model headaches
WebAudio upsampling to 48kHz → Whisper failure → Solved with forced AudioContext + soxr resample
WASM memory limits during long audio
Nano API UX inconsistencies across Chrome channels
Deployment gymnastics
Vercel static frontend + PythonAnywhere backend
CORS + streaming SSE + media blob piping
Fallback behavior when Nano is missing
Database integrity
Preserving consistent hashing logic while refactoring
Ensuring cross-session reuse under rate limits
Human challenge
Hackathon + dev fatigue + model rate limits + deadlines = 💥 (Made it anyway.)
Accomplishments that we're proud of
End-to-end local speech-to-text in browser
Chrome Nano pipeline for text prep working reliably
Deterministic hashing + smart cache layer
YouTube → local STT pipeline without cloud STT
Usable in low-connectivity situations
Maintained privacy guarantees
Full interface polished enough for real users
And honestly — staying sane while debugging 48kHz audio issues at 3AM.
What we learned
Technical
Nano + WASM can power hybrid reasoning pipelines
Device-first AI requires careful UX fallback
ffmpeg resampling quirks are not for the weak
Browser SSE + LLM streaming is surprisingly elegant
Human
Timeboxing matters
Simplicity > “correctness on paper”
Big insight
People don’t just need answers — they need transparent reasoning they can export, trace, and trust.
What's next for Epistemiq
Phase 2 — journalist/security features
Source provenance scoring
PDF cite-trace (auto-link evidence to sources)
Offline evidence cards
Chrome extension mode
Mobile full-offline WASM model pack
Federated cache sharing between team devices
Plugin API for newsroom workflows
Long term vision:
An offline-capable epistemic OS for journalists, researchers & educators.
Epistemiq should feel like a personal research assistant that never tattles on you.
Log in or sign up for Devpost to join the conversation.