ΑΙ Software Engineer

What the candidate will build in priority: - Restructure the Gemini pipeline from one long monolithic prompt into a chain of focused calls, each with structured JSON output. - Enforce JSON compliance using Gemini's response_schema plus post-hoc validation and retry logic. No more silent hallucinations landing in Airtable. - Ground every scored field in video timestamps. Any claim without a valid timestamp gets flagged for human review. - Add an LLM-as-judge critic pass and self-consistency scoring to catch hallucinations before they reach production. Potentially: - Build a small evaluation harness on a 30–50-video calibration set so every prompt or model change is measured, not guessed. - Automate the current manual glue: Airtable ↔ Gemini ↔ Frame.io, including auto-posting Frame.io annotations from the timestamped JSON output. - Route audio analysis out of Gemini to dedicated tools (Shazam/AcoustID for track ID, librosa/Essentia for BPM and key). Must have: - Solid Python. You will be writing the first real codebase for this project - Hands-on experience with the Google Gemini API, including video input and structured output - Working knowledge of the Airtable API beyond the no-code UI: auth, rate limits, upserts, idempotent writes. - Comfort with JSON Schema and post-hoc validation. - Systematic debugging. - Clear written communication: updates, honest about blockers