PitchRoast: The Brutal AI VC

Inspiration

Every day, brilliant founders fail to raise capital because they choke during actual VC calls. There is no safe, high-pressure space to practice an elevator pitch. Existing AI tools are mostly text-based or rely on slow, expensive telecom APIs (like Twilio) that don't mimic the fast-paced, interruptive nature of a real Silicon Valley investor. We wanted to change that by building an AI that doesn't just listen, but aggressively pushes back.

What it does

PitchRoast AI is a real-time, browser-based mock-VC simulation.

1. The founder clicks "Start Pitch" and speaks into their microphone.
2. Powered by Amazon Nova 2 Sonic, the AI acts as an impatient investor. If the founder rambles or uses too many buzzwords, the AI interrupts them in real-time with tough questions (e.g., "Stop throwing buzzwords, what is your CAC?").
3. To mimic a real elevator pitch (and control token costs), the session is hard-capped at 2 minutes.
4. Post-call, Amazon Nova 2 Lite analyzes the transcript to generate a structured "Roast Report," grading the founder from 0-100 on clarity, confidence, and objection handling.

How we built it

We architected a highly lean, real-time streaming pipeline to bypass expensive telecom setups:

• The Voice Engine: We used Amazon Nova 2 Sonic via Amazon Bedrock to handle the speech-to-speech generation.
• The Pipeline: Audio is captured via the browser's MediaRecorder API, chunked into Base64, and sent over a Node.js WebSocket connection directly to AWS Bedrock.
• The VAD Hack: Instead of using heavy WebAssembly libraries for Voice Activity Detection, we utilized the browser's native AudioContext and AnalyserNode to detect 1.5 seconds of silence, automatically triggering the AI's barge-in response.
• The Brain & Storage: Amazon Nova 2 Lite handles the post-call JSON evaluation. We used Next.js for the frontend, Node.js for the WS server, and a self-hosted MySQL database with Prisma ORM to ensure zero-latency cold starts.

Challenges we ran into

Handling streaming audio responses using the AWS SDK v3 InvokeModelWithResponseStreamCommand was our biggest hurdle. Managing the audio buffers and piping the Uint8Array byte chunks back through the WebSocket to the Next.js frontend without audio distortion required tight engineering. Additionally, balancing the VAD threshold so the AI doesn't interrupt too early (or too late) took significant fine-tuning.

Accomplishments that we're proud of

• Sub-second Latency: Achieving near-instantaneous voice interactions completely in the browser without third-party WebRTC providers.
• Cost-Control Architecture: Implementing the strict 2-minute cutoff and frontend VAD ensures we don't waste AWS tokens on background noise or dead air.
• End-to-End Execution: Going from concept to a fully deployed, database-backed Voice AI SaaS in under 48 hours.

What we learned

We learned the sheer power of Amazon Nova 2 Sonic. Building an interruptible, low-latency voice agent is no longer a massive infrastructure problem; with Bedrock's streaming APIs, indie developers can build enterprise-grade voice experiences over a weekend.

What's next for PitchRoast: The Brutal AI VC

The next step is leveraging Amazon Nova's Multimodal Understanding. We plan to allow founders to upload their Pitch Deck (PDF) before the call. The AI will physically "look" at the slides during the voice conversation and ask targeted questions about specific charts or financial projections on Slide 5!

Built With

• amazon-bedrock
• amazon-nova-2-lite
• amazon-nova-2-sonic
• mysql
• next.js
• node.js
• prisma
• tailwindcss
• typescript
• websockets