cover_letter / art_lKhbwZYXAUs

Dear Adobe Research and AI Hiring Team, Adobe's Research and AI team sits at a genuinely rare intersection: foundational model research that ships into products used by hundreds of millions of creators, from a photographer retouching in Photoshop to a marketing team generating campaign assets in Express. That breadth — research-to-product, professional-to-business — is exactly the surface I have been operating on. When I built the RL post-training workbench that benchmarks GRPO, DPO, and ten other algorithms across TRL, VeRL, OpenRLHF, and NeMo RL, I was not doing it as an academic exercise; I was doing it because I needed a rigorous, reproducible way to evaluate model quality tradeoffs before committing to a production direction. That instinct — evidence-based quality judgment at the model layer — is what this Principal PM role is asking for, and it is how I have worked for the last several years. **Technical and AI/ML Foundation** My technical credibility in generative AI and model evaluation runs from first principles to production systems. In 2004, I hand-coded backpropagation through time in C++ for a protein structure prediction system at UC Berkeley; that work was accepted at NeurIPS 2014. In 2025–2026, I rewrote that system in PyTorch spanning 413 parameters to 8 billion — a 19-million-fold scale increase — with five neural architectures (feedforward, GRU, Transformer, ESM-2, multi-task), MLflow experiment tracking, Optuna hyperparameter optimization, and FastAPI serving across six Docker containers with 823 automated tests. More directly relevant to Adobe's generative AI work: I built aeval, a local-first model evaluation platform covering factuality, reasoning, instruction-following, safety, and code generation. The platform applies bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and saturation detection — the kind of statistical rigor required to make defensible quality/cost/latency tradeoff decisions rather than relying on anecdote. I also built an RL post-training workbench implementing 12 algorithms (PPO, GRPO, DAPO, REINFORCE, REINFORCE++, RLOO, DPO, SimPO, IPO, KTO, ORPO, SPPO) with standardized throughput, memory, and convergence benchmarking across frameworks — exactly the comparative evaluation discipline needed when a research team is deciding which training approach to commit to at scale. On the multi-modal and agentic side, I architected a RAG retrieval pipeline with multi-provider LLM orchestration (Claude, GPT-4, Gemini) including fallback routing, structured-output validation, and token-budget optimization for Fintellect AI. I also built the OpenClaw multi-agent orchestration framework — a gateway protocol with subagent delegation, profile management, and session switching — coordinating AI agent workflows across multiple industry verticals. These systems required exactly the kind of platform thinking Adobe needs: individual capabilities that compose into a coherent, scalable architecture. **Why This Role** My arc has moved consistently toward the layer where research meets product: building the evaluation infrastructure, the quality benchmarks, and the tradeoff frameworks that let model researchers and surface teams speak the same language. The Principal PM, Research and AI — Foundations role is precisely that layer at Adobe, and Firefly's scope — image, video, Photoshop, Express, and beyond — means the quality decisions made here propagate across the entire creative ecosystem. What specifically draws me to this role is the mandate to work alongside model researchers to guide research direction and set quality benchmarks, then translate early-stage capabilities into market-ready requirements. That translation work — from research artifact to product specification to customer outcome — is where I have spent the most deliberate energy. At Intuit, I scaled the ICE platform to 675M+ engagements in FY23 and drove a 6K-to-50K TPS throughput increase; the technical credibility required to make those infrastructure decisions with engineering teams is the same credibility required to engage meaningfully with model researchers on architecture tradeoffs. I also built evaluation frameworks (aeval, the RL workbench) that operationalize quality measurement rather than leaving it implicit — which maps directly to Adobe's need for scalable, evidence-based model quality assessment. **Selected Prior Experience** - **RL Workbench (2026):** Implemented 12 RL algorithms with algorithm-specific metric profiles and standardized throughput/memory/convergence benchmarking across TRL, VeRL, OpenRLHF, and NeMo RL — establishing a reproducible framework for model quality and efficiency tradeoffs. - **aeval (2025–2026):** Built a model evaluation platform with five core eval types, adversarial safety testing with refusal detection, bootstrap confidence intervals, Welch's t-test, and Cohen's d effect size — CI/CD integrated with regression detection and automated safety gates. - **Fintellect AI — Multi-provider LLM orchestration:** Architected RAG pipeline with Claude/GPT-4/Gemini fallback routing, structured-output validation, and token-budget optimization; built 13 domain-specific AI advisors with context-aware advisory. - **Intuit ICE Platform:** Achieved 275% YoY growth in engagements, scaling to 675M+ in FY23; drove throughput from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections with sub-25ms TP99 — evidence of platform-scale product leadership with deep infrastructure engagement. - **Intuit Developer Frameworks:** Extended Java and Python SDK Starter Kits with scaffolding templates, build configurations, testing frameworks, and CI/CD integration; delivered ICE Self-Service platform reducing developer onboarding from 2–3 weeks to minutes — translating technical complexity into developer-facing product clarity. - **NeurIPS 2014 — Protein Structure Prediction:** Published research on artificial neural networks for secondary structure prediction; original 2004 system hand-coded in C++ with custom BPTT, rewritten in 2026 spanning 413 to 8B parameters. - **AutoEval — Automated Visual Evaluation for Robot Model Training (2025):** Built automated visual evaluation system using multimodal AI (Claude/GPT-4V) to score model outputs against natural-language rubrics, reducing evaluation cycles from 72 hours to ~4 minutes — a direct precedent for scalable, automated quality evaluation in generative model workflows. **Closing** Adobe's mission — empowering everyone to create — is meaningful precisely because the foundational capabilities built by the Research and AI team determine what is possible for every creator in the ecosystem. The work of defining quality benchmarks, guiding research direction, and translating model capabilities into product requirements is not peripheral to that mission; it is the mechanism by which research becomes creative empowerment at scale. I would welcome the opportunity to bring my model evaluation frameworks, platform-scale product experience, and technical depth in AI/ML to that work. Thank you for your consideration. O. Felix Amoruwa famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info