Voice of Experience.

Inspiration I chose this problem because, while working on the project, I often faced diagnostic questions with no one to guide me, and I realized how easily practical reasoning can be lost when it only exists in conversations. That experience made me want a way to preserve expert explanations so new engineers can access them when they need them. What it does The system records an expert’s voice, transcribes it, and turns it into a structured diagnostic card with five fields: problem, causes, checks, solution, and cautions. The code verifies only the extracted quotes against the transcript, the expert approves the card, and the search engine returns “I don’t know” when no matching card exists. How we built it I built the prototype using Python, FastAPI, SQLite, and a simple HTML/JS interface. Gemini and Groq handle transcription and field extraction, and TF‑IDF provides the search. I guided an AI coding agent (Antigravity) through the implementation, reviewed its reports, and tested the application myself. Challenges we ran into API quotas required a fallback chain of models, and the code had to verify the extracted quotes to avoid silent errors. A duplicated card caused a drop in search accuracy until it was removed. Accomplishments I ran 46 automated tests to validate core functions. The tuning score was 9/10 on the threshold‑tuning set and 7/7 on the held‑out set, though both sets are small and illustrative. What we learned I learned how to create API keys, connect multiple AI models, and verify extracted quotes instead of assuming they are correct. I also learned how small changes in the library can affect search behavior. What’s next If the project continues, I would collect recordings from real practitioners, improve handling of duplicate cards, and run larger evaluations. Three honesty notes All cards are demo examples, not field logs. The evaluation sets are small, written by me, and include one false positive. The system has not yet been tested with practicing engineers.

Built With

• api
• css
• fastapi
• gemini
• groq
• html
• javascript
• python
• scikit-learn
• sqlite
• whisper