Poise.ai

Live interview agent

Inspiration

Interview prep is often unstructured: you grind LeetCode-style problems but rarely rehearse how you sound under pressure, pacing, clarity, confidence, and storytelling. Poise.ai exists because the uncertainty of “am I ready?” is as much about communication and presence as it is about getting the right answer.

What it does

Poise.ai helps users practice both technical interview skills and soft skills in one flow: structured prompts, realistic scenarios, and feedback that goes beyond correctness to include how you explain tradeoffs, handle ambiguity, and present yourself. Our model mimics a real life interview using the Gemini 3.1 Flash model.

How we built it

We designed a practice loop: prompt → response (typed and/or spoken) → structured evaluation → targeted drills. The product layer focuses on repeatable sessions (warm-up, deep-dive, behavioral), while the intelligence layer scores clarity, structure, and technical depth—optionally with rubrics tailored to role type (e.g., frontend vs. backend vs. leadership).

Challenges we ran into

Making feedback feel coach-like, not “LLM-y.” We iterated on rubrics, few-shot examples, and guardrails so critiques stayed actionable and interview-realistic. Balancing technical vs. soft evaluation without one dominating the score or confusing the user. Latency and cost if using multimodal or long sessions; we had to tune session length, caching, and summarization. Trust & safety: avoiding “perfect answer” leakage for real company questions; steering toward general frameworks and transferable skills.

Accomplishments that we’re proud of

A unified practice experience that treats communication skills as first-class, not an afterthought. A feedback system users can apply immediately in the next attempt (clear “do this next” guidance).

What we learned

Interview readiness improves fastest with tight feedback loops and specific rehearsal (STAR stories, system design narration, debugging out loud). We also learned that product success here depends on calibration: users need to understand what “good” looks like for their target role, not a generic bar.

What’s next for Poise.ai

Deeper personalization: role/company track, weakness detection over time, spaced repetition. Richer modalities: voice stress practice, timed drills, mock panels. Community templates and mentor-reviewed rubrics (optional human-in-the-loop).