FOUNDER
VOICE.
An AI investor that pressure-tests startup pitches before the real room does.
WHY DO FOUNDERS FAIL IN VENTURE MEETINGS?
Founders spend weeks refining pitch decks with friends, advisors, and colleagues. But polite practice creates dangerous false confidence: people who care about you rarely ask brutal questions about market size, churn assumptions, or defensive moats.
When founders finally step in front of institutional venture partners, they get blindsided within the first three minutes by difficult unit-economics questions they have never rehearsed out loud.
THE NON-SYCOPHANTIC AI INVESTOR
Founder Voice removes sycophancy entirely. By tuning system prompts and reasoning temperatures, the AI partner behaves like a partner at a top-tier venture firm: polite, articulate, but aggressively probing every vulnerability in the founder's thesis.
“We are building an autonomous agent for hospital billing workflows that cuts billing cycle delay by 60%.”
“Healthcare billing has 9-to-12-month enterprise procurement cycles and draconian HIPAA audits. Your CAC will easily exceed $8,000 before your first pilot renewal. How does an early team survive that cash burn?”
SUB-400MS STREAMING PIPELINE
Human conversations break immersion if pauses between turns exceed 500 milliseconds. To solve this, we decoupled the architecture into concurrent streaming micro-pipes.
Streams 16kHz PCM audio chunks over low-latency WebSockets while parsing PDF slides via PDF.js into structured text and tabular data.
Performs real-time streaming speech transcription with sub-150ms word-level latency, punctuated sentence boundaries, and filler-word normalization.
Executes a multi-phase system prompt instructing the model to act as a seasoned, adversarial venture partner with zero sycophancy, seeking unaddressed risks at >300 tokens/sec.
Streams human-quality conversational audio directly back to the founder's earphones, delivering responses before natural human conversational turn-taking timeouts.
Stores pitch transcripts, extracted slide embeddings, question history, and outputs an audit scorecard ranking Market Size, Moat, and Team Conviction.
WHAT THE BUILDATHON TAUGHT ME
Waiting for the LLM to generate an entire paragraph before sending it to text-to-speech guarantees a 2-second delay. Feeding sentences to the TTS synthesizer as tokens arrive is essential for conversational reality.
If the model only knows general venture principles, its critiques become cliché. Extracting the real deck metrics (TAM, pricing, acquisition channel) gives the critic legitimate teeth.
Shipping in a 48-hour buildathon forces ruthless prioritization. Every feature that doesn’t directly serve the core audio loop gets cut, resulting in a cleaner, more focused product.
Selected as the Finalist (Runner-Up) in the OpenAI × NIAT Buildathon, validating both the technical execution and the emotional urgency of the user problem.