cover_letter / art_OJ_IrpiPBrU

Dear Adobe CX Enterprise Hiring Team, Adobe is building the control plane that turns LLMs into enterprise coworkers — grounded in first-party data, governed by enterprise contracts, and accountable to real customer outcomes. That framing is precise, and it matches how I think about the problem. My interest in this role is grounded in a specific trajectory: from hand-coding BPTT in C++ at Berkeley in 2004, to shipping a production multi-agent orchestration framework (OpenClaw) with gateway protocol, subagent delegation, and session management in 2024, to building an RL post-training workbench that benchmarks GRPO, DPO, PPO, and nine other algorithms across TRL, VeRL, OpenRLHF, and NeMo RL today. I have been living in this stack, not reading about it. **Technical and AI/ML Foundation** The work most directly relevant to this role falls into three areas. First, agent orchestration and the harness layer. OpenClaw is a multi-agent orchestration framework I designed and built from scratch — gateway protocol, subagent delegation, profile management, and session switching — enabling coordinated AI agent workflows across distinct industry verticals (real estate, insurance, financial markets). I also built and shipped MCP server integration in StreamIO, exposing screen capture tools to AI coding assistants via the Claude MCP SDK. I have debugged MCP servers, written skills, and have direct opinions about where the protocol is well-specified and where it leaks abstraction. I understand why "just use the model" is not an answer at the platform layer. Second, evaluation infrastructure. My aeval platform is a local-first model evaluation system with five core eval types (factuality, reasoning, instruction-following, safety, code generation), adversarial safety testing with refusal detection, and statistical rigor: bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and saturation detection. It integrates into CI/CD with regression detection and automated safety gates. The stack is FastAPI, TimescaleDB, Redis, Next.js, and Ollama. I built this because I needed evals that predict behavior, not evals that flatter models. That distinction — vanity evals versus evals that predict customer outcomes — is exactly the framing in the JD, and I have shipped infrastructure around it. Third, RL post-training and the feedback loop. My RL Workbench covers the full RLHF/DPO pipeline: Reward Lab for designing and A/B testing reward functions (RLVR, learned, hybrid) across GSM8K, MATH, HumanEval, and UltraFeedback; a Playground for real TRL-powered GRPO/DPO training with live SSE metric streaming on Apple Silicon (MPS) or CUDA; and an Arena for head-to-head framework benchmarking with GPU passthrough in Docker containers. Twelve algorithms implemented with algorithm-specific metric profiles and standardized throughput/memory/convergence benchmarking. The override and feedback loop is the moat — I have built the tooling to study it empirically. This work is grounded in a research foundation: NeurIPS 2014 publication on neural networks for protein secondary structure prediction, and the BRAIN platform rewrite in PyTorch spanning 413 parameters (original 2004 C++ system) to 8B parameters, with MLflow, Optuna HPO, and FastAPI serving. **Why This Role** My arc — from platform infrastructure at scale (Intuit ICE: 675M+ engagements, 50K TPS, ~1.5M concurrent connections) to founding AI-native products with real agentic architecture — has consistently been about the same problem: building the contracts and control planes that make complex systems composable and trustworthy. Adobe's CX Enterprise agent runtime is that problem at a scope and data depth I want to work on. What specifically draws me to this role is the combination of owning the agent loop contracts (tool protocols, skills, knowledge grounding) *and* the developer surface that practitioners build against. At Intuit, I delivered the ICE Self-Service platform — DevPortal, GitOps config, ICE Playground — reducing developer onboarding from 2–3 weeks to minutes in pre-prod. I understand that the developer surface is not documentation; it is a product with its own contracts, and local dev/prod parity is a first-class requirement, not an afterthought. Adobe's position — grounding agents in RTCDP, AJO, and CJA data with enterprise governance — means the control plane work is genuinely defensible, and the evaluation loop closes against real customer data. That is the interesting problem. **Selected Prior Experience** - **OpenClaw multi-agent orchestration framework** — designed gateway protocol, subagent delegation, profile management, and session switching; production deployment across real estate, insurance, and financial markets verticals. - **MCP SDK integration in StreamIO** — built MCP server exposing screen capture tools to AI coding assistants via Claude MCP SDK; hands-on experience debugging MCP servers and writing skills in production. - **aeval evaluation platform** — local-first eval system with adversarial safety testing, refusal detection, statistical rigor (bootstrap CI, Welch's t-test, Cohen's d), CI/CD regression detection, and automated safety gates. - **RL Workbench** — 12-algorithm post-training benchmark platform (PPO, GRPO, DAPO, DPO, SimPO, and seven others) with live metric streaming, cross-framework Arena benchmarking (TRL, VeRL, OpenRLHF, NeMo RL), and GPU Docker passthrough. - **Intuit ICE Self-Service platform** — DevPortal, GitOps config, ICE Playground; reduced developer onboarding from 2–3 weeks to minutes in pre-prod and under 24 hours for production; mitigated $1M+ in projected opex growth; scaled to 675M+ engagements in FY23. - **Java and Python SDK Starter Kits at Intuit** — extended with scaffolding templates, build configurations (Gradle/Maven), testing frameworks, and CI/CD integration; zero-to-production-ready microservice in minutes. - **Splunk Search Service and SPL/SPL2** — owned Go microservices, PostgreSQL metadata service, and Splunk Processing Language roadmap; delivered Scheduler Service end-to-end in four months; achieved up to 10x query performance improvements for beta enterprise customer. - **NeurIPS 2014** — published research on artificial neural networks for protein secondary structure prediction; establishes depth in ML research, not just product. **Closing** Adobe's stated mission — empowering everyone to create — extends in the enterprise context to empowering practitioners to build reliable, governed, data-grounded agents at scale. The control plane work your team is doing is the infrastructure that makes that possible. I want to own that layer, hold the line on the contracts that matter, and build the evaluation loop that makes agent quality measurable against customer outcomes rather than benchmark artifacts. I would welcome the opportunity to go deeper on any of the technical specifics above. --- **O. Felix Amoruwa** famoruwa@berkeley.edu · 909-731-9011 · felixamoruwa.info