cover_letter / art_TFLCbif_TJU

Dear Anyscale Hiring Team, Anyscale sits at the foundation of how serious AI teams move from prototype to production — Ray's distributed computing model is increasingly the substrate on which the industry's most demanding ML workloads run, from OpenAI's training infrastructure to Uber's real-time prediction systems. That infrastructure layer is exactly where I have spent the last several years: scaling developer platforms to 675M+ engagements at Intuit, building post-training RL workbenches that benchmark distributed training frameworks head-to-head, and shipping ML evaluation platforms from scratch. When I saw the Ray Data PM role, the fit was immediate and specific. ## Technical Foundation My technical credibility in the distributed data and ML infrastructure space is grounded in hands-on work, not just roadmap ownership. At Intuit, I owned the ICE platform — a developer-facing infrastructure layer serving ~1.5M concurrent connections across QuickBooks, TurboTax, Mint, Mailchimp, and Credit Karma. I drove the rSocket migration that scaled throughput from 6K to 50K TPS with sub-25ms TP99 latency, and delivered 275% YoY growth in platform engagements to 675M+ in FY23. That work required the same balance Ray Data demands: keeping the open developer experience frictionless while building the proprietary runtime capabilities that justify enterprise adoption. I also reduced developer onboarding from 2–3 weeks to under 24 hours for production through the ICE Self-Service platform — a concrete example of developer experience as a product strategy, not a side concern. On the ML infrastructure side, I built an RL post-training workbench that benchmarks GRPO, DPO, PPO, DAPO, and eight other algorithms across TRL, VeRL, OpenRLHF, and NeMo RL — with GPU Docker passthrough and live SSE metric streaming on Apple Silicon (MPS) and CUDA. This is precisely the kind of distributed ML workload that Ray Data is designed to serve: multi-framework, compute-intensive, requiring reproducible throughput and memory benchmarking across environments. I built that system because I needed to understand the tradeoffs firsthand, not because a vendor told me what they were. I also built aeval, a local-first model evaluation platform with a FastAPI orchestrator, TimescaleDB for time-series metrics, Redis job queuing, and statistical rigor baked in — bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and automated safety gates for CI/CD integration. Building evaluation infrastructure from the ground up gave me direct exposure to the data preprocessing and batch inference pipelines that Ray Data is purpose-built to accelerate. My research foundation goes back further: a NeurIPS 2014 paper on neural networks for protein structure prediction, and the original hand-coded BPTT implementation in C++ at UC Berkeley in 2004 — rewritten in 2026 as a full PyTorch platform spanning 413 parameters to 8B (a 19M-fold scale increase), with MLflow, Optuna HPO, and Docker orchestration across six containers. ## Why Ray Data, Why Now The Ray Data PM role sits at the exact intersection of my background: distributed systems infrastructure, developer-facing platform strategy, and the ML data pipeline layer that determines whether AI workloads are feasible at scale. The challenge of balancing open source community growth with commercial differentiation in Anyscale Runtime is one I find genuinely interesting — it is the same tension I navigated at Intuit between platform openness and proprietary capability, and at Splunk between SPL's open query language and the commercial search infrastructure built on top of it. What excites me specifically about Ray Data is the architectural bet: building a scalable data processing library on Ray Core that handles both offline batch inference and ML training preprocessing in a unified model. The opportunity to define which parts of the ML/data lifecycle Ray Data should own — and which integrations create the most durable ecosystem lock-in — is a strategic problem I am well-positioned to work through, having done the adjacent work in distributed developer platforms and having built ML pipelines myself. ## Selected Relevant Experience - **ICE Platform, Intuit** — Scaled developer infrastructure to 675M+ engagements and 50K TPS; reduced onboarding from weeks to minutes; achieved 275% YoY engagement growth across 5 major product lines. - **rSocket Migration, Intuit** — Led throughput scaling from 6K to 50K TPS with sub-25ms TP99 latency supporting ~1.5M concurrent connections — directly analogous to Ray Core's distributed execution model. - **RL Post-Training Workbench** — Built end-to-end benchmark platform for 12 RL algorithms across TRL, VeRL, OpenRLHF, and NeMo RL with GPU Docker passthrough and live metric streaming — a Ray Data target workload built from scratch. - **aeval Evaluation Platform** — FastAPI + TimescaleDB + Redis architecture with CI/CD integration, regression detection, and automated safety gates; direct experience with the batch inference and evaluation pipelines Ray Data serves. - **Splunk Search Orchestration** — Owned Search Service (Go microservices), Search Catalog (PostgreSQL), and SPL/SPL2; delivered Scheduler Service end-to-end in ~4 months; led query performance optimization achieving up to 10x improvements for enterprise customers. - **Enterprise Service Language Assessment, Intuit** — Conducted analysis across 9 languages and 20+ mobile apps to inform CTO-level strategic investment decisions — the kind of ecosystem-wide thinking required to position Ray Data against competing data processing frameworks. - **SDK Starter Kits, Intuit** — Extended Java and Python SDKs with scaffolding, build configs, testing frameworks, and CI/CD integration; developer experience as a first-class product deliverable. ## Closing Anyscale's mission — making distributed AI infrastructure accessible and production-ready for the teams building the next generation of AI systems — is one I find worth working on. The Ray ecosystem is already the substrate for some of the most consequential ML workloads in the industry, and Ray Data is the layer that determines whether those workloads are fast, reliable, and developer-friendly at scale. I would bring to this role a combination of platform PM experience at scale, hands-on ML infrastructure work, and the technical depth to engage credibly with both the open source community and enterprise customers. I would welcome the opportunity to discuss how my background maps to what you are building. Sincerely, **O. Felix Amoruwa** famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info