Inspiration
In 1960, Kennedy and Nixon held the first televised presidential debate. More than sixty years later, the format is the same: two podiums, a moderator, and an audience left to work out on their own what was true, what was spin, and what was a dodge. Debates stood still while political literacy slipped. Fact-checks show up the next morning, long after the clip has gone viral. We wanted the truth to show up while the words are still on screen.
What it does
The BS Subtitle is live captioning that fact-checks a debate as it happens. While the candidates speak, it transcribes them, pulls out each checkable claim, and flags:
- False or misleading claims, checked against a curated knowledge base and live web sources
- Logical fallacies, such as straw men, false dilemmas and ad hominem attacks
- Contradictions, when a speaker goes against something they said earlier in the same debate
Every flag cites its evidence, and the moderator is never flagged. An operator view lets a human kill a flag before it goes on air. The audience view runs a few seconds behind live, like a broadcast delay, so each flag lands under the exact line it's about. After the debate, a report collects every claim, verdict and source.
How we built it
- Speech to text: Deepgram live streaming, which reconnects automatically if the stream drops
- Backend: FastAPI with an async session runner. Every caption segment fans out to parallel analysis lanes (claims, stance, fallacy, consistency), and each lane works from a frozen snapshot of the debate so it can't see the future.
- LLMs: Groq is the primary model for speed, with Gemini as the fallback. Each lane has a hard budget of 1.2 seconds. Groq gets most of it, and Gemini only runs if enough time is left to finish.
- Evidence: a sentence-transformers knowledge base of reviewed source passages, plus web retrieval, with MongoDB for sessions and a claim ledger
- Reliability: every LLM output is validated against a strict schema (Pydantic). A bad output gets one retry, and after that the line is marked "pending" instead of being shown wrong.
- Frontend: Vite, React and TypeScript, with operator, audience and report views kept in sync over websockets
- Quality: contract tests in CI, SonarCloud on every PR, and Playwright end-to-end tests
Challenges we ran into
Our bugs can talk for us. The worst ones:
- The 1.2-second wall. One slow Groq call could eat the whole budget, so the Gemini fallback never got a chance. We had to split the time down to the millisecond. At one point a timing error of a fraction of a second skipped the fallback without any error at all.
- Free-tier reality. Groq's free reasoning models take about half a second per call, so we had to pick between smarter verdicts and on-time verdicts on almost every lane.
- Two clocks, one show. The AI analyzes a trimmed clip while the audience watches the untrimmed original, so every caption and flag has to be mapped onto the original video's timeline, including the parts we cut.
- Never flag the moderator. It sounds simple, until speaker labels from transcription move around in the middle of a sentence.
Accomplishments that we're proud of
The whole website so far: a working live pipeline from audio to captions to flags to evidence, with operator, audience and report views that stay in sync, a real broadcast delay, and a human kill switch. All of it runs on a budget that fits inside a sentence.
What we learned
We learned how models really behave under pressure. Spotting a fallacy means reading intent, and that's much harder than checking a fact. Even "true" needs a definition: what counts as proof, which source wins, and how confident is confident enough to put on screen? We also learned to treat Groq and Gemini as a trade-off between quality and speed, and to design for what viewers expect rather than for a perfect answer that shows up too late.
Built With
- deepgram
- gemini
- groq
- mongodb
- python
- typescript
- uvicorn