Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Shipping a Voice AI Interview Coach: Multi-Model Orchestration, and the Eval Loop I Thought Was Running (But Wasn’t)
Learn how Sage, a voice AI interview coach, uses GPT-4 and Claude for adaptive mock interviews. Discover multi-model orchestration and a debugging journey to fix a non-functional real-time evaluation feature.
Sage is a voice AI interview coach that runs adaptive mock interviews tailored to your actual resume and a specific job description. You upload your resume, paste the JD, pick the round (screening, technical, or final), and it interviews you like a real interviewer for that exact role.
Live, I’ll show the architecture: how GPT-4 runs the real-time voice layer through Vapi while Claude handles answer evaluation and the final assessment, how the interview plan is generated from the resume PDF and JD at setup time, and the actual route code. I’ll also show the audit that revealed a “real-time evaluation” feature I thought was live had zero callers, and how I traced live-vs-dead code across the repo to find it.
Sage is an adaptive, voice-based AI interviewer evaluating technical, behavioral, and communication skills.
Next.js TypeScript web application template running on port 3000.
- VapiVapi is the API platform for deploying advanced, human-like voice AI agents with sub-600ms latency across phone, web, and custom applications.Vapi is the core developer platform for building and scaling voice AI agents, handling complex infrastructure so you focus on the experience. The system orchestrates three core technologies (Speech-to-Text, LLM, Text-to-Speech) to enable real-time, natural conversations: achieving sub-600ms response times for human-like turn-taking. Agents offer full phone integration (inbound/outbound calls) and web embedding, plus Tool Calling to connect to your existing APIs for complex workflows (e.g., appointment scheduling, data retrieval). This is a modular, enterprise-grade solution designed for reliability and scale.
- GPT-4GPT-4 is OpenAI’s large multimodal model: it processes both text and image inputs, delivering human-level performance on complex professional and academic benchmarks.This is OpenAI’s latest milestone in scaling deep learning: a large multimodal model accepting both text and image inputs. It demonstrates a significant capability leap over its predecessor, scoring in the top 10% on a simulated bar exam (GPT-3.5 scored in the bottom 10%). The model handles nuanced instructions and long-form content, supporting context windows up to 32,768 tokens (32K model). This capacity allows processing up to 25,000 words in a single, complex prompt. GPT-4 is engineered for enhanced reliability, steerability, and advanced reasoning across diverse tasks.
- Claude HaikuClaude Haiku is Anthropic's fastest, most cost-efficient large language model, engineered for near-instant responsiveness and high-throughput AI applications.Haiku is the compact, high-speed model within the Claude family, optimized for latency-sensitive tasks like real-time customer service and agentic sub-agent orchestration. It delivers near-frontier coding quality, scoring 73.3% on SWE-bench Verified (Haiku 4.5), matching performance previously seen in larger models like Sonnet 4. Developers utilize Haiku for its exceptional value: pricing starts at $1 per million input tokens and $5 per million output tokens. This model supports a 200,000-token context window and includes multimodal vision capabilities, making it ideal for scalable, budget-conscious deployments that demand speed and accuracy.
- Claude SonnetClaude Sonnet 4.5 is Anthropic's premier model: state-of-the-art for agentic coding, computer use, and complex, long-horizon workflows.Claude Sonnet 4.5 is engineered for superior agentic performance, excelling in complex, multi-step workflows across coding, finance, and cybersecurity (e.g., achieving a 77.2% score on SWE-bench Verified). This model offers a powerful balance of speed and cost: it is priced at $3 per million input tokens and supports a massive 200,000-token context window. This capacity allows for sustained reasoning, with internal tests confirming the model maintains focus for over 30 hours on demanding tasks. It is available via the Claude API, Amazon Bedrock, and Google Cloud's Vertex AI, making it the top choice for developers building robust, production-ready AI agents.
- Anthropic SDKThe Anthropic SDK is your official, high-efficiency interface for integrating Claude models (like Claude 3.5 Sonnet) into applications, featuring native support for Tool Use and streaming.This is the direct conduit for deploying Anthropic's state-of-the-art models. The SDK provides robust, multi-language clients (Python, TypeScript) for the Claude API. It ensures streamlined integration with native support for critical features: synchronous and asynchronous client options, real-time response streaming, and advanced Tool Use (function calling) for external system interaction. Utilize models like Claude 3 Opus or the fast Claude 3.5 Sonnet in production with minimal setup.
Compose Email
Loading recent emails...