HackStack

Project Info

AId Hominem

Devpost

What it does

AId Hominem is a live debate coach. It: Listens and transcribes speech in real time via streaming speech-to-text Flags logical fallacies as they're spoken — ad hominem, strawman, false dilemma, slippery slope, circular reasoning, emotional appeals, and more; each with the offending quote, a severity/confidence rating, an explanation, and a suggested follow-up question Fact-checks claims in real time by extracting a checkable claim, searching the web, reading sources, and returning a verdict (supported / contradicted / unclear) with citations

How we built it

Backend (server.js) — a Node.js HTTP server with a hand-rolled WebSocket proxy. Speech-to-text — Deepgram streaming STT over a WebSocket proxy. Fallacy analysis — Gemini 3.1-Flash-Lite calls with a structured JSON output schema. Fact-checking — Gemini extracts one checkable claim, Browserbase searches and fetches pages, and a second LLM call produces the verdict with sources.

Challenges we ran into

Real-time without latency death — running STT, fallacy analysis, and web fact-checking simultaneously meant carefully throttling work (analysis every ~15s, fact-checks every ~30s) so the UI stayed responsive. Writing a WebSocket proxy from scratch — to keep the dependency footprint tiny we implemented the WebSocket framing protocol over raw TCP ourselves, which meant debugging frame parsing by hand.

What we learned

How to architect a single codebase that runs in both browser and Node/Electron contexts. The nuts and bolts of the WebSocket framing protocol. How to coax reliable structured JSON out of LLMs and combine multiple model calls into a search-and-verify pipeline.

What's next

Speaker diarization — attribute flags to who said what in multi-person debates. Post-debate reports — a scored summary of each participant's argument quality and fallacy count. Rebuttal suggestions — go beyond flagging and propose how to respond. Mobile and meeting integrations — Zoom/Meet/Teams plugins for live calls. Tunable strictness — let users dial sensitivity up for casual chats or down for formal debate. Personal coaching mode — analyze your own arguments over time to help you debate better.

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only
  • CSSIn code
  • HTMLIn code
  • JavaScriptIn code
  • Google GeminiClaimed
  • Node.jsClaimed

3 of 5 appear in the indexed code. 2 claimed on Devpost could not be matched to code, which may simply mean the tool leaves no trace in the repository.

AI coding agents

  • Claude CodeCommits

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

55 KB

Source files

6

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.

0 stars