Tail Villain

My OpenAI Build Week hackathon submission story

Why GPT-5.6 roadmap generation kept failing after schema validation passed, and why the final fix required replaying the real application request

Connecting Study and Interview outcomes to FSRS so active recall quality drives the next review

Adding OpenAI text, speech, and Realtime paths without replacing the existing Gemini relay architecture

Adding privacy-safe structural telemetry for production RAG calls, admission, scores, latency, and source mix

Separating transcript scope, eligibility, and evidence states so partial interviews are not scored as complete sessions

Separating stable learner memory from query-specific retrieval after correct chunks ranked first but missed the score threshold

Building a source-scoped RAG pipeline and validating its effect on real Study responses

Stabilizing live interviews by defining reliable boundaries for user and assistant turns

Managing Gemini Live credentials and long-lived sessions through a backend WebSocket relay

Managing voice opt-in, waiting, playback, and cancellation as one product flow

Aligning Kovill’s reply style, question progression, and language with product-owned state

Adding standalone coding practice without breaking the existing session and dashboard flow

Removing software-interview assumptions to support interview practice across professions

Evaluating full interview conversations with a typed LLM-as-a-Judge pipeline

Building credits, consent, and operational boundaries before accepting payment

The first Tail Villain beta launch, from production deployment failures to preparing the product for its first users

Recovering interview control flow lost during a prompt and grounding refactor

Enforcing LLM cost limits in the backend and modeling their operational impact

Treating active sessions as resumable work while preserving truthful progress

Connecting spaced review to a landing-page story without letting RPG flavor obscure the product

A Tail Villain note on treating LLM completion signals as advisory instead of letting them end sessions unconditionally

A Tail Villain note on study telemetry and the move toward a roadmap-level interview hub

Turning study mode from a one-off chat into a session with state, progress, and a report

How interviewer personas moved from a selection UI into distinct interview and study flows

What bilingual support revealed about local UI state, account preferences, and dashboard structure

Connecting interview results to topic progress, review timing, and the practice loop

Provider boundaries, latency handling, and failure modes that surfaced while wiring Gemini and Ollama into the same product

The polling bug that made completed backend analysis keep looking like it was still running

I pressed the button and waited thirty seconds. Technically, nothing was wrong. That was the problem.

The starting point of Tail Villain, built around the gap between knowing something and retrieving it under interview pressure