cover_letter / art_CnvHcDoJgM4

Dear Hinge Health Intelligent Care Hiring Team, Hinge Health is doing something that most digital health companies only gesture at: using AI to substantively automate care delivery for the 1 in 2 Americans who will experience a musculoskeletal condition in their lifetime. The opportunity to build Robin — an always-on, clinically-informed AI care assistant — sits at exactly the intersection where I have spent the last several years: agentic AI systems, developer-facing platforms, and 0-to-1 product execution. My path from hand-coding backpropagation through time in C++ at UC Berkeley in 2004 to building production multi-agent orchestration frameworks and RL post-training workbenches today reflects a sustained, hands-on commitment to AI that goes well beyond product oversight. **Technical and AI Foundation** My AI work is applied and current. At Streamio AI, I designed and built OpenClaw — a multi-agent orchestration framework with a gateway protocol, subagent delegation, profile management, and session switching — enabling coordinated AI agent workflows across real estate, insurance, health, and financial services verticals. This is the same architectural pattern Robin requires: a central orchestration layer routing intent to specialized subagents, with context preserved across turns. I built this end-to-end, not through a vendor integration. On the evaluation side, I built aeval — a local-first LLM evaluation platform covering factuality, reasoning, instruction-following, safety, and code generation, with adversarial safety testing, refusal detection, and data contamination detection via SHA-256 hashing. Statistical rigor was non-negotiable: bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and automated safety gates in CI/CD. The stack — FastAPI orchestrator, TimescaleDB, Redis job queue, Next.js dashboard, Ollama — was built to support the kind of rigorous, iterative eval process that a clinically sensitive product like Robin demands. Hinge Health's requirement to own the LLM-as-a-judge and golden dataset eval process is work I have already done in production. My RL Workbench (2026) benchmarks GRPO, DPO, PPO, DAPO, REINFORCE++, RLOO, SimPO, KTO, ORPO, and SPPO across TRL, VeRL, OpenRLHF, and NeMo RL with live SSE metric streaming and GPU Docker passthrough — giving me direct, current fluency in post-training alignment methods that will shape the next generation of clinically safe conversational AI. My NeurIPS 2014 paper on neural networks for protein secondary structure prediction, and the 2026 rewrite of that system spanning 413 parameters to 8B, grounds this fluency in research, not just tooling. **Why This Role** I have spent three years building agentic AI products from scratch and four years before that scaling platform infrastructure at Intuit to 675M+ annual engagements. Robin is precisely the product that requires both: the technical depth to engage on prompt engineering, agent orchestration, and eval frameworks, and the platform instincts to transition Robin from a point solution to a company-wide capability layer that other pods build on. That platformization challenge is one I know well. **Role-Specific Connection** The JD's emphasis on multi-agent orchestration, proactive member engagement, and clinical triage maps directly to the architectural decisions I made building OpenClaw and the Fintellect Agents — domain-scoped conversational agents with context-aware routing and structured output validation. The requirement to own evaluation frameworks, including safety guardrails and human eval processes, is the exact scope of aeval. And the call to navigate regulatory and clinical complexity with Legal and Compliance partners reflects the kind of cross-functional leadership I exercised at Intuit, where I partnered with engineering, design, and executive stakeholders to ship developer infrastructure that touched every major Intuit product line. **Selected Prior Experience** - **OpenClaw multi-agent orchestration (Streamio AI):** Built gateway protocol, subagent delegation, profile management, and session switching enabling coordinated AI agent workflows — directly applicable to Robin's multi-agent architecture. - **aeval LLM evaluation platform:** Built 5 core eval types, adversarial safety testing with refusal detection, bootstrap confidence intervals, Welch's t-test, Cohen's d, and automated CI/CD safety gates — directly applicable to Robin's eval and clinical safety requirements. - **Fintellect Agents:** Built domain-specific conversational AI agents with RAG retrieval (ChromaDB), multi-provider LLM orchestration (Claude, GPT-4, Gemini) with fallback routing, structured output validation, and token budget optimization — applicable to Robin's personalized, context-aware care interactions. - **ICE Self-Service Platform (Intuit):** Delivered developer platform reducing onboarding from 2–3 weeks to minutes, mitigating $1M+ in projected opex growth; scaled ICE engagements 275% YoY to 675M+ — demonstrates platform-building and scaling at company-defining scope. - **RL Workbench:** Benchmarked 12 RL algorithms across 4 major frameworks with live metric streaming and GPU Docker passthrough — establishes current, hands-on fluency in post-training alignment methods relevant to improving Robin's response quality and safety. - **AutoEval — Automated Visual Evaluation for Robot Model Training:** Repurposed multimodal AI pipeline (Claude/GPT-4V) to score model outputs against natural-language rubrics, reducing evaluation cycles from 72 hours to ~4 minutes — demonstrates applied LLM-as-a-judge evaluation design. - **Splunk Search Orchestration (Senior PM):** Owned Go microservices, PostgreSQL metadata service, and SPL/SPL2 query language; delivered Scheduler Service end-to-end in ~4 months and achieved up to 10x query performance improvements — demonstrates technical depth in distributed systems and end-to-end delivery ownership. **Closing** Hinge Health's mission — making world-class MSK care accessible to anyone, anywhere — is one where AI can genuinely close a gap that the traditional care system has not. Robin, done well, is not a chatbot; it is a clinical companion that earns trust through accuracy, safety, and consistency. I have spent years building the technical foundations — multi-agent orchestration, rigorous evaluation, aligned post-training — that make that kind of trust possible. I would welcome the opportunity to bring that work to Hinge Health. Sincerely, **O. Felix Amoruwa** famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info