cover_letter / art_SPKPfq4BR6c

Dear Databricks AI Platform Hiring Team, Databricks sits at the intersection of where enterprise data infrastructure and production AI converge — a problem space that matters because the gap between a model that works in a notebook and one that reliably serves millions of users is where most AI value gets stranded. That gap is precisely where I have spent the last several years: building platforms, frameworks, and tooling that move AI from experimentation into production. When I read about Mosaic AI's ambition to unify model training, serving, evaluation, and LLM infrastructure into a single governed platform, I recognized the same architectural instinct that drove my own work — and I want to help build it at Databricks' scale. ## Technical and AI/ML Foundation My technical foundation spans from the fundamentals to the frontier. In 2004, I hand-coded a neural network in C++ with custom backpropagation through time to predict protein secondary structure — work that was accepted at NeurIPS 2014. In 2026, I rewrote that system in PyTorch, scaling from 413 parameters to 8 billion across five architectures (feedforward, GRU, Transformer, ESM-2, multi-task), with MLflow experiment tracking, Optuna hyperparameter optimization, and FastAPI serving — a 19-million-fold parameter scale increase that reflects how much the field has moved and how closely I have tracked it. More directly relevant to Databricks' Mosaic AI model training platform: I built an RL post-training workbench that covers the full RLHF/DPO pipeline. It includes a Reward Lab for designing and A/B testing reward functions (RLVR, learned, and hybrid) across GSM8K, MATH, HumanEval, and UltraFeedback datasets; a Playground for real TRL-powered GRPO and DPO training with live 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 — PPO, GRPO, DAPO, REINFORCE, REINFORCE++, RLOO, DPO, SimPO, IPO, KTO, ORPO, and SPPO — with standardized throughput, memory, and convergence benchmarking across frameworks. This is the kind of infrastructure thinking that Mosaic AI's training and evaluation platform requires. 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, and statistical rigor including bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and saturation detection — integrated into CI/CD with automated safety gates. The stack — FastAPI orchestrator, TimescaleDB, Redis job queue, Next.js dashboard, Ollama — reflects the same architectural patterns Databricks uses across its serving and monitoring infrastructure. For retrieval and vector search, I architected a RAG pipeline at Fintellect AI with ChromaDB vector store, multi-provider LLM orchestration across Claude, GPT-4, and Gemini with fallback routing, structured output validation, and token budget optimization — directly analogous to Databricks' Vector Search and AI query products. ## Why This Role My arc from NeurIPS researcher to platform PM who shipped infrastructure at 675M+ engagements and 50K TPS points directly at what Databricks needs: someone who can go deep with ML engineers on distributed training architectures and real-time serving systems, and simultaneously own the roadmap, commercialization strategy, and enterprise adoption motion. The AI Platform Staff PM role is that exact intersection. What excites me specifically about this role is the mandate to connect MLflow, Unity Catalog, Model Serving, Vector Search, Feature Engineering, and Agent infrastructure into a cohesive experience. That integration problem — making the full ML lifecycle feel like one governed, performant system rather than a collection of tools — is the hardest and most consequential product problem in enterprise AI right now. I also want to work on the adoption bottleneck problem: identifying where enterprise ML teams stall between experimentation and production and building the platform capabilities that remove those blockers. ## Selected Prior Experience - **RL Post-Training Workbench (2026):** Built 3-phase workbench benchmarking GRPO/DPO across TRL, VeRL, OpenRLHF, and NeMo RL; implemented 12 RL algorithms with cross-tab workflow lineage tracking and standardized throughput/memory/convergence benchmarking — directly applicable to Mosaic AI's model training platform. - **aeval — AI Model Evaluation Platform (2025–2026):** Built evaluation platform with adversarial safety testing, statistical rigor (bootstrap CI, Welch's t-test, Cohen's d), and CI/CD regression detection — relevant to Databricks' model monitoring and evaluation infrastructure. - **Intuit ICE Platform — 675M+ Engagements, 50K TPS (2021–2024):** Delivered ICE Self-Service platform (DevPortal, GitOps config, ICE Playground), reducing developer onboarding from 2–3 weeks to minutes; achieved 275% YoY engagement growth; scaled throughput from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections with sub-25ms TP99 — demonstrating the ability to ship and scale data-intensive platform infrastructure at enterprise volume. - **Java and Python SDK Starter Kits (Intuit):** Extended SDK scaffolding with build configurations (Gradle/Maven), testing frameworks, and CI/CD integration — empowering developers to go from zero to production-ready microservice in minutes; directly analogous to Databricks' developer-facing ML platform tooling. - **Splunk Search Orchestration — 10x Query Performance (2019–2021):** Owned Search Service (Go microservices), Search Catalog (PostgreSQL metadata service), and SPL/SPL2; led query performance optimization achieving up to 10x improvements for beta enterprise customers — relevant to Databricks' query engine and data lakehouse performance work. - **Fintellect AI RAG Pipeline:** Architected retrieval pipeline with ChromaDB, multi-provider LLM orchestration with fallback routing, and token budget optimization — maps to Databricks' Vector Search and AI query products. - **NeurIPS 2014 — Neural Networks for Protein Structure Prediction:** Published research on artificial neural networks; original system hand-coded in C++ with custom BPTT — establishes credibility with Databricks' research-oriented ML platform engineering teams. ## Closing Databricks' mission — making it radically easier for enterprises to put AI into production — is the right problem to be working on. The organizations that figure out how to reliably operationalize AI at scale will define the next decade of enterprise technology, and the platform that enables them will matter enormously. I have spent 12 years building toward this intersection of deep technical credibility, platform product leadership, and enterprise scale, and I would welcome the opportunity to bring that to the AI Platform team. Thank you for your consideration. --- **O. Felix Amoruwa** famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info