cover_letter / art_LBU0TwkLoR4

Dear Supabase Hiring Team, Supabase has quietly become the default backend for developers who need to move fast — and the bet you're making on agent-native infrastructure is the right one at exactly the right moment. The shift you're describing, where AI coding tools account for a meaningful share of the development happening on your platform, is not a future state I'm theorizing about. It's the environment I've been building in and for. When I shipped StreamIO's MCP server exposing screen capture tools to AI coding assistants, and when I implemented the OpenClaw multi-agent orchestration framework with gateway protocol and subagent delegation, I was solving exactly the class of problems this role owns: what do the right abstractions look like when an agent is the primary consumer of your API? ## Technical Foundation My technical grounding runs from the infrastructure layer up through the AI layer. At Intuit, I owned developer-facing platform infrastructure at genuine scale — 675M+ ICE engagements in FY23 across QuickBooks, TurboTax, Mint, Mailchimp, and Credit Karma, with throughput scaled from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections at sub-25ms TP99. I extended Java and Python SDK Starter Kits with scaffolding templates, build configurations, and CI/CD integration, and I delivered the ICE Self-Service platform that compressed developer onboarding from 2–3 weeks to minutes. That work taught me how developer tooling decisions compound: a bad default in a starter kit propagates across hundreds of services before anyone notices. On the AI side, I've built and operated systems that put real pressure on the same primitives Supabase is working on. The OpenClaw multi-agent orchestration framework I built for StreamIO implements a gateway protocol with subagent delegation, profile management, and session switching — the kind of coordination layer that breaks in interesting ways when context is wrong or tool schemas are ambiguous. I also built aeval, a local-first model evaluation platform with factuality, reasoning, instruction-following, safety, and code generation eval types, statistical rigor via bootstrap confidence intervals and Welch's t-test, and CI/CD integration with automated safety gates. The question of whether your agent tooling is actually working — which is precisely what the evals portion of this role addresses — is one I've built infrastructure to answer. My RL Workbench covers the full RLHF/DPO post-training pipeline with 12 algorithms (PPO, GRPO, DAPO, DPO, SimPO, and others) benchmarked across TRL, VeRL, OpenRLHF, and NeMo RL with GPU Docker passthrough. That's not background color — it's evidence that I track what Anthropic, OpenAI, and the research community ship and form opinions about the primitives underneath. ## Why This Role The Supabase PM role is specifically about what happens when agents are the primary users of your product decisions — defaults, error messages, naming, deprecations, docs all behave differently when an AI coding tool sees them before a human does. That lens is one I've developed through first-hand use, not observation. I connect to infrastructure through MCP servers, I've built agent skills that call external APIs, and I've watched agents fail in ways that are invisible until you look at the tool schema. What excites me about this specific role is the combination of ownership scope and measurement discipline. Owning the MCP server, agent skills, and the evals that tell you whether any of it is working is a complete feedback loop — you define the abstractions, you ship them, and you have the instrumentation to know if agents are actually succeeding. The requirement to define metrics before engineering starts, tracking developer activation and task completion alongside AI eval scores, is exactly the rigor I applied at Intuit when I used SQL and BigQuery telemetry to prioritize developer pain points across 20 mobile apps and 30+ product SKUs. ## Selected Relevant Experience - **MCP server and agent tooling (StreamIO):** Shipped MCP server exposing screen capture tools to AI coding assistants; implemented OpenClaw multi-agent orchestration with gateway protocol, subagent delegation, and session switching — direct experience with the abstractions this role owns. - **Developer platform at scale (Intuit):** Delivered ICE Self-Service platform reducing developer onboarding from 2–3 weeks to minutes; scaled to 675M+ engagements in FY23; extended Java and Python SDK Starter Kits with scaffolding, build configs, and CI/CD integration. - **AI eval infrastructure (aeval):** Built evaluation platform with 5 eval types, adversarial safety testing, refusal detection, bootstrap confidence intervals, and automated safety gates — directly applicable to the eval ownership this role requires. - **RAG and multi-provider LLM orchestration (Fintellect AI):** Architected RAG retrieval pipeline with ChromaDB, multi-provider LLM orchestration (Claude, GPT-4, Gemini) with fallback routing, structured output validation, and token budget optimization. - **Search platform ownership (Splunk):** Owned Search Service (Go microservices), Search Catalog (PostgreSQL metadata service), and SPL/SPL2 — built roadmaps, PRDs, and acceptance criteria; delivered Scheduler Service end-to-end in ~4 months and achieved up to 10x query performance improvements for a beta customer. - **Enterprise-wide language and tooling strategy (Intuit):** Conducted Service Language Assessment across 9 languages analyzing usage data and developer feedback; findings presented to CTO. Built Asterias, a declarative asset lifecycle management platform with GraphQL API. - **NeurIPS-published ML research:** Accepted paper on artificial neural networks for protein structure prediction; original 2004 system hand-coded in C++ with custom BPTT, rewritten in 2026 spanning 413 to 8B parameters. ## Closing Supabase's mission — giving developers the fastest path from idea to production — is being stress-tested by a new class of developer that doesn't read docs, doesn't tolerate ambiguous error messages, and fails silently when the abstractions are wrong. Getting the MCP server, agent skills, and eval infrastructure right is not a feature request; it's the foundation that determines whether agents can reliably build on Supabase at all. That's a problem I've been working toward from multiple angles, and I'd welcome the chance to own it directly. Thank you for your consideration. --- **O. Felix Amoruwa** famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info