I built PerPitch as a voice-first investor simulation: a founder can pitch naturally, receive selective questions, and get structured analysis grounded in the session, startup context, and supporting evidence.
RULE 01 The founder speaks first.
RULE 02 A pause is not permission to interrupt.
A founder may know the deck but still need to discover where the reasoning is weak, which assumptions invite challenge, and whether the evidence supports the claim. Access to that kind of conversation is uneven and often depends on a warm introduction.
Speech detection can notice a pause, but it cannot decide whether the founder has finished an argument. Retrieval and enrichment may improve analysis while making a live reply too slow to feel conversational. The product needed an explicit policy for when the AI should speak—and when it should wait.
I led the product and AI work across turn policy, live voice orchestration, session state, evidence retrieval, structured scoring, objection handling, and IC-style memo generation. The live path keeps the exchange moving; deeper research continues after the turn or session.
The default story stays at the level of the person and the work they are trying to finish.
The founder brings the startup, deck, and current reasoning into one session.
The conversation policy protects longer founder turns instead of replying to every pause.
Questions respond to the phase and evidence rather than a fixed interviewer script.
Retrieval, enrichment, scoring, and contradiction checks move outside the latency-critical turn.
The session becomes a grounded memo and feedback surface rather than a disappearing chat.
Founders feel conversational delay immediately, while research and scoring become more useful when they can inspect the whole session. Splitting those clocks made the simulation more responsive without reducing the evidence available afterward.
Earlier versions placed retrieval, enrichment, and heavier analysis inside the live loop. Some stages took roughly 15–25 seconds, turning a simulated conversation into a wait state.
Only speech state, a bounded amount of context, question policy, and the next response had to complete before the founder could continue.
I narrowed the synchronous path and moved deeper research, scoring, contradiction checks, and report generation into background work.
Later architecture reached roughly 2–4 seconds at p50 in the documented performance context: keep synchronous intelligence narrow and let post-session intelligence be thorough.
Recorded product-use context across PerPitch's founder-evaluation work.
Discovery conversations informed the product, scoring rubrics, objections, and report workflow.
Documented later performance context after removing heavier work from the live loop.
The technical route explains the application-level turn policy, live/async split, evidence grounding, and the difference between a useful simulation and an overactive voice model.
Latency-critical dialogue stays narrow while evidence-heavy analysis runs after the turn or session.
A pause does not automatically grant the AI permission to interrupt.
Evaluation combines conversation evidence with startup context instead of relying on persona performance.
The evidence supports interaction volume and product learning—not predictive validity.
Have a useful AI idea trapped in an unclear brief or brittle prototype?
Send me the rough version →