Interview Gym

Inspiration

Interview prep is noisy, passive, and non-adaptive. We wanted a system that behaves like a strict interviewer: asks, evaluates, adapts difficulty, and tracks weaknesses over time.

What it does • Simulates real interview rounds (DSA, systems, behavioral). • Asks multi-step, follow-up questions. • Evaluates answers with rubric-based scoring. • Tracks topic-level gaps (e.g., DP state design, graph traversal pruning). • Adapts next questions based on performance. • Generates focused drills from mistakes.

How we built it • LLM-driven agent with role-constrained prompting (interviewer persona). • Structured state: {topic, difficulty, weaknesses, history}. • Evaluation pass: rubric + reasoning trace + score normalization. • Feedback generator separated from question generator (reduces leakage). • Memory layer for longitudinal tracking. • Lightweight UI + session replay.

Challenges we ran into • Overly generous grading by the model. • Leakage: feedback contaminating future prompts. • Difficulty calibration drift. • Hallucinated “expected answers” for open-ended system design. • Keeping evaluation deterministic enough for tracking progress.

Accomplishments that we’re proud of • Adaptive loop that converges on user weaknesses in <5 sessions. • Clear, surgical feedback (not generic tips). • Realistic follow-ups that pressure-test assumptions. • Consistent scoring via rubric anchoring. • Reduced token cost with staged evaluation passes.

What we learned • Separation of roles (ask vs grade vs coach) improves quality. • Explicit rubrics outperform “grade this answer” prompts. • Memory must store abstractions, not raw transcripts. • Agents need guardrails to avoid becoming a tutor mid-question. • Determinism matters for perceived fairness.

What’s next for Interview Gym • Multi-agent panel interviews (e.g., DSA + hiring manager). • Timed whiteboard mode with incremental hints. • Code execution sandbox + test case grading. • Behavioral story scoring using structured frameworks. • Long-term skill trajectory visualization.

Built With

• python