cover_letter / art_XVMqZweYa4g

Dear Tavus Hiring Team, Tavus is building something genuinely novel — not another chatbot wrapper, but a new computing paradigm where machines can see, hear, and respond with the presence of a real person. The gap between human connection and machine reliability is one of the more consequential problems in applied AI, and closing it has direct implications for healthcare access, education, and how people navigate complex decisions. I've spent the last year building AI platforms that sit at exactly this intersection — real-time multimodal pipelines, multi-agent orchestration, and developer-facing infrastructure — and the work Tavus is doing with PALs is the logical frontier of where those threads converge. **Technical Foundation** My credibility here is built from first principles, not job titles. In 2004, I hand-coded backpropagation through time in C++ to train a neural network for protein structure prediction — work that was accepted at NeurIPS 2014. In 2026, I rewrote that system in PyTorch spanning 413 parameters to 8 billion, a 19-million-fold scale increase, with MLflow experiment tracking, Optuna hyperparameter optimization, and FastAPI serving across six Docker containers. That arc — from bare-metal implementation to production ML platform — shapes how I think about AI product development: I understand what's actually happening under the hood, which means I can make better tradeoffs and earn the trust of engineering teams who can tell the difference. More recently, I built a full RL post-training workbench covering the complete RLHF/DPO pipeline — implementing 12 algorithms (PPO, GRPO, DAPO, DPO, SimPO, KTO, and six others) with live SSE metric streaming, cross-framework benchmarking across TRL, VeRL, OpenRLHF, and NeMo RL, and GPU Docker passthrough for Apple Silicon and CUDA environments. I built aeval, a local-first model evaluation platform with bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and automated safety gates — because shipping AI products without rigorous evaluation infrastructure is how you get surprised in production. These aren't portfolio pieces; they're the tools I built because I needed them. On the infrastructure side, I shipped the OpenClaw multi-agent orchestration framework — a gateway protocol with subagent delegation, profile management, and session switching enabling coordinated AI agent workflows across multiple industry verticals. I also built a real-time HLS livestreaming pipeline with multi-stream canvas compositing at 30fps, FFmpeg transcoding, and WebSocket communication — the kind of latency-sensitive, stateful system that requires genuine engineering judgment to get right. **The Bridge** At Intuit, I scaled developer platform infrastructure to 675M+ engagements in FY23, drove 275% YoY growth in ICE engagements, and led a throughput migration from 6K to 50K TPS supporting ~1.5M concurrent connections at sub-25ms TP99. I reduced developer onboarding from 2–3 weeks to under 24 hours for production environments. That experience — owning the full arc from developer experience to production outcomes at enterprise scale — is directly relevant to what Tavus needs as it moves from Series B into the phase where platform reliability and developer trust become the competitive moat. **Why This Role** Tavus's framing of PALs as a new application category — not a feature, not an API, but a new computing primitive — is the kind of 0-to-1 problem I find most interesting. The JD is intentionally sparse, which signals that you're looking for someone who can define the problem space, not just execute against a pre-written roadmap. The combination of real-time human simulation models, multimodal AI, and developer-facing platform infrastructure maps directly onto the work I've been doing, and the mission — a therapist anyone can afford, a medical assistant for every patient — is the kind of outcome worth orienting a product career around. **Selected Prior Experience** - Built production multi-agent orchestration framework (OpenClaw) with gateway protocol, subagent delegation, and session management — enabling coordinated AI workflows across real estate, insurance, health/dental, and financial markets verticals. - Architected RAG retrieval pipeline with ChromaDB vector store, multi-provider LLM orchestration (Claude, GPT-4, Gemini) with fallback routing, structured output validation, and token budget optimization. - Delivered ICE Self-Service platform (DevPortal, GitOps config, ICE Playground), reducing developer onboarding from 2–3 weeks to minutes in pre-prod and <24 hours for production, while mitigating $1M+ in projected opex growth. - Achieved 275% YoY growth in ICE engagements, scaling to 675M+ in FY23; scaled throughput from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections with sub-25ms TP99. - Built automated visual evaluation system (AutoEval) for robot model training using screen capture and multimodal AI, reducing evaluation cycles from 72 hours to ~4 minutes with zero-integration architecture. - Implemented ICE Presence in async chat, generating $480K/month in additional invoicing; deployed Background-to-Foreground Messaging on iOS/Android with <100ms latency. - Led 0-to-1 product strategy, AI development, and go-to-market execution for two AI platforms (StreamIO AI, Fintellect AI) from customer discovery through App Store launch. **Closing** Tavus's mission — machines that truly understand people — is not a product category I'm approaching from the outside. It's the direction my technical and product work has been pointing for some time. I build things that work, I earn engineering trust by operating with rigor and clarity, and I move fast without losing the thread of what actually matters to users. I'd welcome the conversation. Sincerely, **O. Felix Amoruwa** famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info