cover_letter / art_EwGwqiA_Ez4

Dear Five9 Hiring Team, Five9 sits at a genuinely consequential intersection: cloud infrastructure that handles millions of real customer conversations daily, now being reshaped by applied AI that can make those conversations faster, more accurate, and more human. That combination — enterprise-grade reliability meeting frontier AI capabilities — is exactly the kind of product challenge I have spent the last several years building toward. My path from hand-coding backpropagation through time in C++ at UC Berkeley in 2004 to shipping production RL post-training workbenches and multi-agent orchestration frameworks today reflects a consistent orientation: I build AI systems that work in the real world, not just in research notebooks. **Technical and AI Foundation** My AI/ML work is applied and production-oriented. I built aeval, a local-first model evaluation platform with five core eval types — factuality, reasoning, instruction-following, safety, and code generation — incorporating adversarial safety testing with refusal detection, bootstrap confidence intervals, Welch's t-test, and Cohen's d effect size for statistical rigor. The stack runs FastAPI orchestration, TimescaleDB, Redis job queuing, and a Next.js dashboard with CI/CD regression detection and automated safety gates. This is the kind of infrastructure that answers the exact questions a product team shipping LLM-powered agent assist needs to answer: does the model actually work, under what conditions does it fail, and how do we detect drift before customers do? My RL Workbench covers the full RLHF/DPO post-training pipeline — Reward Lab for designing and A/B testing reward functions across GSM8K, MATH, HumanEval, and UltraFeedback; a Playground for real TRL-powered GRPO/DPO training with live SSE metric streaming on Apple Silicon 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 standardized throughput, memory, and convergence benchmarking. Understanding these tradeoffs at the implementation level — not just the product level — is what allows me to have credible conversations with data science teams about latency constraints, accuracy-cost tradeoffs, and model selection decisions. At Intuit, I owned the developer platform infrastructure that scaled to 675M+ engagements in FY23 across QuickBooks, TurboTax, Mint, Mailchimp, and Credit Karma. I drove a rSocket migration that scaled throughput from 6K to 50K TPS supporting approximately 1.5M concurrent connections at sub-25ms TP99. That is the kind of real-time performance envelope that agent assist products live and die by — a suggestion that arrives 800ms late is a suggestion that gets ignored. **Why This Role** The Agent Assist product domain is precisely where my technical depth and product leadership converge. Real-time AI assistance for contact center agents is a latency-constrained, accuracy-critical, human-in-the-loop problem — and it requires a PM who can navigate model training tradeoffs, evaluate LLM outputs rigorously, and translate that understanding into roadmap decisions that engineering and data science teams can execute against. That is the work I have been doing. **Role-Specific Connection** The JD's emphasis on navigating complex data environments, ML parameter tradeoffs, and enterprise AI deployment realities maps directly to what I built at Intuit and what I continue to build in my own AI projects. I am particularly drawn to the beta testing and customer engagement dimension of this role — at Intuit, I ran enterprise-wide service assessments across nine languages, synthesized developer feedback at scale, and translated that into strategic investment decisions presented to the CTO. Doing that same work with contact center operators and agents, in a domain where AI assistance has immediate, measurable impact on customer outcomes, is a compelling next chapter. **Selected Relevant Experience** - Delivered ICE Self-Service platform (DevPortal, GitOps config, ICE Playground) at Intuit, reducing developer onboarding from 2–3 weeks to minutes in pre-production and under 24 hours for production, while mitigating $1M+ in projected opex growth — a 0-to-1 platform launch with measurable enterprise adoption. - Scaled ICE platform to 675M+ engagements in FY23 with 275% YoY growth; drove rSocket migration achieving 50K TPS throughput and sub-25ms TP99 latency across ~1.5M concurrent connections — directly relevant to the real-time performance requirements of agent assist. - Built aeval evaluation platform with adversarial safety testing, statistical significance frameworks, and CI/CD safety gates — providing the measurement infrastructure needed to ship LLM-powered features responsibly. - Built RL Workbench benchmarking 12 algorithms across TRL, VeRL, OpenRLHF, and NeMo RL with standardized convergence and throughput metrics — enabling data-driven model selection decisions. - Implemented OpenClaw multi-agent orchestration framework with gateway protocol, subagent delegation, and session management — production experience coordinating multiple AI agents across distinct task domains. - Architected RAG retrieval pipeline at Fintellect AI with ChromaDB vector store, multi-provider LLM orchestration (Claude, GPT-4, Gemini) with fallback routing, structured output validation, and token budget optimization — directly applicable to the LLM orchestration challenges in real-time agent assist. - At Splunk, led query performance optimization for a beta enterprise customer, building a mirrored topology for benchmark testing and achieving up to 10x performance improvements — experience translating customer-specific performance problems into systematic engineering solutions. **Closing** Five9's mission — bringing joy to customer experience — is not a soft aspiration. It is an engineering and product challenge: make AI assistance fast enough, accurate enough, and contextually aware enough that agents can focus on the human dimension of every interaction. I have spent twelve years building toward exactly that kind of work, from NeurIPS-published neural network research to production platform infrastructure serving hundreds of millions of engagements. I would welcome the opportunity to bring that foundation to the Agent Assist portfolio. Thank you for your consideration. Sincerely, **O. Felix Amoruwa** famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info