cover_letter / art_HQmXB_7zOfI

Dear Roblox Compute Platform Hiring Team, Roblox sits at a genuinely rare intersection: a platform where 88 million people show up daily to create and connect, and where the infrastructure underneath — GPU fleets, Kubernetes control planes, AI training and inference pipelines — directly determines how fast that creative potential compounds. The recent launch of the Cube Foundation Model and real-time 4D generation makes clear that Roblox's next chapter is compute-bound, and that the Senior PM for Compute Platform will be setting the pace for everything that follows. My path from hand-coding backpropagation through time in C++ at Berkeley in 2004 to building production RL post-training workbenches that benchmark GRPO, DPO, PPO, and nine other algorithms across TRL, VeRL, OpenRLHF, and NeMo RL today has been oriented around exactly this kind of infrastructure-meets-AI challenge. **Technical and AI/ML Foundation** My credibility here is grounded in hands-on build experience, not just roadmap ownership. At Intuit, I owned the ICE platform — a distributed compute and communication layer spanning QuickBooks, TurboTax, Mint, Mailchimp, and Credit Karma — and scaled it from 6K to 50K transactions per second via an rSocket migration supporting approximately 1.5 million concurrent connections at sub-25ms TP99, while growing engagements 275% YoY to 675M+ in FY23. That required the same tradeoffs this role demands: utilization vs. latency vs. cost, with reliability as a non-negotiable constraint across a heterogeneous fleet. On the AI infrastructure side, I built an RL post-training workbench from scratch that covers the full RLHF pipeline: a Reward Lab for designing and A/B testing reward functions across GSM8K, MATH, HumanEval, and UltraFeedback; a live training Playground powered by TRL with real-time SSE metric streaming on Apple Silicon MPS and CUDA; and an Arena for head-to-head framework benchmarking across TRL, VeRL, OpenRLHF, and NeMo RL with GPU passthrough in Docker containers. I implemented 12 RL algorithms with algorithm-specific metric profiles and standardized throughput, memory, and convergence benchmarking across frameworks. This is the kind of compute-aware, GPU-scheduling-sensitive work that maps directly to what Roblox needs as it scales frontier model training and inference. My aeval platform — built on FastAPI, TimescaleDB, Redis, and Ollama — adds a production-grade evaluation layer with bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and automated safety gates integrated into CI/CD. And my NeurIPS 2014 paper on neural networks for protein secondary structure prediction, with a 2026 rewrite spanning 413 parameters to 8 billion (a 19-million-fold scale increase), reflects a research foundation that informs how I think about model architecture tradeoffs and compute requirements at scale. **Why This Role, Why Now** Roblox is at an inflection point where the Compute Platform is no longer just infrastructure — it is the product. The Cube Foundation Model, real-time world generation at 16fps, and the agentic Studio tools announced in May 2026 all run on the GPU fleet and Kubernetes substrate this role owns. That makes the Senior PM for Compute Platform one of the highest-leverage positions at the company, and it is exactly the kind of 0-to-1 infrastructure-to-product challenge I have been building toward. **Role-Specific Connection** What draws me most specifically to this role is the combination of fleet-wide GPU reliability and the abstraction layer that product teams build on. At Intuit, I built Asterias — a declarative asset lifecycle management platform with a GraphQL API — precisely to give product teams clean primitives over messy infrastructure reality, which mirrors the Fleet API and Managed Kubernetes Service work described here. I also initiated the MSaaS Drift Detection and Resolution program, writing a Java JAR library to scan Git repos for configuration drift and building a remediation roadmap using OpenRewrite — an analog to the driver and firmware management and fleet-wide health monitoring responsibilities in this role. The Roblox Studio Going Agentic initiative also resonates with my OpenClaw multi-agent orchestration framework, where I built a gateway protocol, subagent delegation, and session management — experience that informs how I think about agentic workloads placing novel scheduling and resource demands on compute infrastructure. **Selected Prior Experience** - **ICE Platform Scale (Intuit):** Scaled throughput from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections at sub-25ms TP99; grew platform to 675M+ engagements in FY23 across five major product lines. - **RL Workbench:** Built GPU-aware post-training infrastructure benchmarking 12 RL algorithms across TRL, VeRL, OpenRLHF, and NeMo RL with Docker GPU passthrough, standardized throughput/memory/convergence metrics, and live SSE streaming. - **ICE Self-Service DevPortal (Intuit):** Delivered GitOps config, ICE Playground, and DevPortal 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. - **MSaaS Drift Detection (Intuit):** Wrote Java JAR library to scan Git repos for configuration drift across microservices; built remediation roadmap using OpenRewrite and partnered with Design on DevPortal UI. - **Asterias Asset Lifecycle Platform (Intuit):** Built declarative asset lifecycle management platform with GraphQL API, using SQL and BigQuery telemetry to prioritize developer pain points across ~20 mobile apps and 30+ product SKUs. - **Search Orchestration (Splunk):** Owned Go microservices for Search Service and Search Catalog (PostgreSQL metadata), delivered Scheduler Service end-to-end in ~4 months, and achieved up to 10x query performance improvements for a Fortune 500 beta customer. - **aeval Evaluation Platform:** Built local-first model evaluation platform on FastAPI, TimescaleDB, Redis, and Ollama with CI/CD regression detection, automated safety gates, and statistical rigor including bootstrap confidence intervals and effect size measurement. **Closing** Roblox's mission — connecting a billion people with optimism and civility — scales only as fast as the compute infrastructure underneath it. Every millisecond shaved from GPU scheduling latency, every reliability improvement in fleet health detection, every clean abstraction that lets an AI research team spin up a training run without filing a ticket: these compound into the platform that makes Cube, real-time world generation, and agentic Studio tools possible at 88 million daily users and beyond. I would welcome the opportunity to bring my platform infrastructure, AI/ML, and developer tooling experience to that work. Thank you for your consideration. O. Felix Amoruwa famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info