cover_letter / art_N8VOthqq1Z4

Dear Anthropic Hiring Team, Anthropic's mission — building AI systems that are reliable, interpretable, and steerable — is not a tagline I encountered for the first time when I read this job description. It is the frame I have used to evaluate every technical decision I have made over the past several years, from designing reward functions in my RL post-training workbench to architecting the multi-provider LLM fallback routing in Fintellect AI. When I saw that Anthropic is now deploying Claude across 276,000 KPMG employees, standing up Agents for Financial Services, and acquiring Stainless to deepen its developer platform, I recognized the exact inflection point where my background becomes directly useful: the moment a frontier model needs to be translated into production architectures that real enterprise engineering teams can trust and operate. ## Technical Foundation My ML work began in 2004 at UC Berkeley, where I hand-coded a backpropagation-through-time neural network in C++ for protein secondary structure prediction — work that was accepted at NeurIPS 2014. In 2026 I rewrote that system as a full PyTorch platform spanning five architectures (feedforward, GRU, Transformer, ESM-2, multi-task), MLflow experiment tracking, Optuna hyperparameter optimization, and FastAPI serving across 413 to 8B parameters — a 19-million-fold scale increase from the original. That arc, from hand-rolled BPTT to modern transformer stacks, gives me genuine fluency at both the algorithmic and systems layers of ML. More directly relevant to Anthropic's current work: I built a three-phase RL post-training workbench covering the full RLHF/DPO pipeline. The Reward Lab module supports A/B testing of reward functions (RLVR, learned, and hybrid) across GSM8K, MATH, HumanEval, and UltraFeedback. The Playground runs real TRL-powered GRPO and DPO training with live SSE metric streaming on Apple Silicon (MPS) or CUDA. The Arena benchmarks TRL, VeRL, OpenRLHF, and NeMo RL head-to-head with GPU passthrough in Docker containers, standardizing throughput, memory, and convergence metrics across 12 algorithms including PPO, GRPO, DAPO, DPO, SimPO, KTO, and ORPO. When an enterprise customer asks how Claude's post-training choices affect their specific task distribution, I can speak to that from implementation experience, not just documentation. On the evaluation side, I built aeval — a local-first model evaluation platform with five core eval types (factuality, reasoning, instruction-following, safety, code generation), adversarial safety testing with refusal detection, bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and CI/CD regression gates. Helping enterprise customers develop evaluation frameworks for their specific Claude use cases is listed as a core responsibility in this role; I have built the tooling that underlies exactly that workflow. ## Why This Role The Applied AI Architect role sits at the intersection of deep technical credibility and enterprise customer success — translating Claude's capabilities into architectures that survive contact with production infrastructure, compliance requirements, and multi-stakeholder buying cycles. That is the work I have been doing, from different angles, for twelve years. What specifically draws me to this role now is the combination of Anthropic's enterprise acceleration (KPMG, PwC, the new Blackstone/Goldman/H&F services company) and the Stainless acquisition signaling serious developer platform investment. I spent three years at Intuit as Staff PM for Developer Frameworks and Platform Infrastructure, where I owned the SDK Starter Kits, the ICE Self-Service DevPortal, and the GitOps configuration layer — the exact surface area Anthropic is now building out. I understand what it takes to reduce enterprise developer onboarding from weeks to hours at scale, and I understand the telemetry, documentation, and tooling gaps that create friction in that journey. ## Role-Specific Connection The JD asks for someone who can move fluidly between engineering deep dives and executive business-value conversations. At Intuit I conducted an enterprise-wide Service Language Assessment across nine languages, synthesized usage data and developer feedback, and presented strategic investment recommendations directly to the CTO — while simultaneously writing Java JAR libraries for configuration drift detection and shipping rSocket migrations that scaled throughput from 6K to 50K TPS. At Splunk I owned Search Service (Go microservices) and SPL/SPL2 roadmaps, built RICE-based prioritization frameworks for three microservice backlogs, and demoed at .conf19. I am comfortable in both rooms. The vertical industry dimension of this role also aligns with my hands-on work: Fintellect AI gave me direct experience architecting RAG pipelines with multi-provider LLM orchestration (Claude, GPT-4, Gemini), fallback routing, and structured-output validation for financial services — precisely the domain where Anthropic is now pushing Agents for Financial Services. ## Selected Relevant Experience - **RL Post-Training Workbench:** Implemented 12 RL algorithms (PPO, GRPO, DAPO, DPO, SimPO, KTO, ORPO, and others) with cross-tab workflow lineage tracking and standardized benchmarking across TRL, VeRL, OpenRLHF, and NeMo RL — directly applicable to helping enterprise customers evaluate and configure Claude's behavior for their task distributions. - **aeval Evaluation Platform:** Built adversarial safety testing with refusal detection, statistical rigor (bootstrap CIs, Cohen's d), and CI/CD regression gates — the infrastructure pattern for the evaluation frameworks this role asks me to help customers build. - **OpenClaw Multi-Agent Orchestration:** Designed and implemented a gateway protocol with subagent delegation, profile management, and session switching across real estate, insurance, health/dental, and financial-markets domains — maps directly to Anthropic's agentic product direction and the integration architecture questions enterprise customers will bring. - **Intuit ICE Platform — 675M+ Engagements, 50K TPS:** Achieved 275% YoY growth in ICE engagements, scaled throughput from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections with sub-25ms TP99 — demonstrates the ability to architect and ship at the scale Anthropic's enterprise customers require. - **Intuit Developer Onboarding — 2–3 Weeks to Minutes:** Delivered ICE Self-Service platform (DevPortal, GitOps config, ICE Playground), reducing developer onboarding from 2–3 weeks to minutes in pre-prod and under 24 hours for production, while mitigating $1M+ in projected opex growth. - **Fintellect AI RAG Pipeline:** Architected a RAG retrieval pipeline (ChromaDB) with multi-provider LLM orchestration, fallback routing, structured-output validation, and token-budget optimization — directly relevant to guiding enterprise customers integrating Claude into existing data and compliance infrastructure. - **NeurIPS 2014 Publication:** Accepted paper on artificial neural networks for protein secondary structure prediction, establishing peer-level research credibility with Anthropic's technical teams. ## Closing Anthropic's commitment to building AI that is safe and beneficial is not a constraint on commercial ambition — it is the reason the enterprise deals with KPMG, PwC, and Goldman Sachs are possible at all. Customers at that scale are not buying capability alone; they are buying trust, interpretability, and the confidence that the system will behave predictably under production conditions. That is the value I want to help communicate and architect. I bring twelve years of enterprise platform experience, hands-on depth across the RL post-training and evaluation stack, and a track record of shipping developer-facing infrastructure at scale — and I would welcome the opportunity to put that to work for Anthropic's customers. Thank you for your consideration. --- **O. Felix Amoruwa** famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info