Under the Hood
Behind-the-scenes posts on how Interview Me was built — architecture decisions, AI integration, and the tradeoffs we made building a real-time interview tool from scratch.
Why Does the Smallest Model Probe the Hardest?
Across 90 interviews, Claude Haiku 4.5 probes deeper and adapts more than Opus 4.7 and Sonnet 4.6 — and turn budget doesn’t explain it. Given three times the room to press, Opus stops anyway. A controlled turn-cap experiment, the judges’ own rationales, and the disposition-versus-capability hypothesis behind the result.
Evaluating Claude as an AI Interviewer: Opus, Sonnet & Haiku
We built an eval harness that grades the AI interviewer itself — synthetic personas with ground-truth caliber, two LLM judges across Claude, GPT and Gemini, and a three-way calibration decomposition — then ran Claude Opus 4.7, Sonnet 4.6, and Haiku 4.5 through it, twice each. The results, why a single run tells you almost nothing, and why we’d split the job between models.