cover_letter / art_XiTxeL4kppU

Dear Gusto Hiring Team, Gusto's mission — to grow the small business economy by handling the hard stuff so owners can focus on their craft — is one of the most concrete expressions of AI's potential to create real economic value for real people. That framing resonates directly with work I've done building AI platforms where mistakes have immediate consequences: from a trading copilot managing live Alpaca market data to an AI financial-education platform guiding first-time investors through a $100k virtual portfolio. The stakes are real, and the trust bar is high — which is exactly the environment where I do my best product work. ## Technical and AI Foundation My technical foundation spans the full arc from first principles to production systems. In 2004 I hand-coded a neural network in C++ with custom backpropagation through time for protein structure prediction — work that became a NeurIPS 2014 accepted paper. In 2026 I rewrote that same system in PyTorch, scaling from 413 parameters to 8B (a 19-million-fold increase), with MLflow experiment tracking, Optuna HPO, and FastAPI serving across six Docker containers. That arc — from low-level algorithm implementation to modern ML infrastructure — gives me a working mental model of what's actually happening inside the systems I ship. On the RL and post-training side, I built a three-phase RL workbench covering the full RLHF/DPO pipeline: a 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 (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, DPO, SimPO, KTO, and others) with standardized throughput, memory, and convergence benchmarking — the kind of eval infrastructure that turns model quality from a gut feeling into a measurable signal. For evaluation more broadly, 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 via bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and saturation detection. This maps directly to what Gusto's JD describes: writing evals, holding a real point of view on AI quality, and owning the responsibility of a product customers rely on day-to-day. On the multi-agent and orchestration side, I designed and implemented OpenClaw — a multi-agent gateway protocol with subagent delegation, profile management, and session switching — coordinating AI agent workflows across real estate, insurance, health/dental, and financial-markets domains. I have direct experience reasoning about API design and MCP layers: Vantage exposes screen-capture and job-search tools to AI coding assistants via an MCP server, and Fintellect's trading copilot uses an App-Store-safe agent control bridge with a typed, allow-listed command bridge to drive embedded TradingView charts. ## Why This Role The CoreX AI charter — building a conversational assistant that earns customer trust while completing real work across payroll, benefits, HR, and compliance — is the hardest version of the AI assistant problem. It's not a demo; it's a product where a wrong action has immediate financial consequences for a small business owner. That's the same constraint I've operated under building Fintellect's live trading features and Vantage's mutation-approval architecture, where enforced approval gates prevented irreversible agent actions. I'm drawn to this role specifically because it requires holding both platform thinking (API design, MCP layer, tool/domain coverage strategy) and UX judgment (how customers feel about AI acting on their behalf) within the same scope — that's the combination I've been building toward. ## Selected Prior Experience - **Conversational assistant with 40+ agent tools and enforced mutation approvals (Vantage):** Shipped a streaming agent chat platform over a shared FastAPI backend where all state-mutating actions required explicit user approval — directly analogous to the trust and safety bar Gusto's assistant must meet when acting on payroll or benefits data. - **OpenClaw multi-agent orchestration (StreamIO AI):** Designed and implemented a gateway protocol with subagent delegation and session switching, coordinating AI agent workflows across multiple industry domains — foundational architecture for expanding tool and domain coverage across Gusto's payroll, time, benefits, and compliance surface area. - **RAG retrieval pipeline with multi-provider LLM orchestration (Fintellect AI):** Architected a ChromaDB-backed RAG pipeline with Claude, GPT-4, and Gemini, including fallback routing, structured-output validation, and token-budget optimization — directly relevant to Gusto's need for a scalable, reliable API strategy for agent-accessible systems. - **275% YoY growth in ICE engagements, scaling to 675M+ in FY23 (Intuit):** Scaled platform infrastructure from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections with sub-25ms TP99 — evidence of platform-level product ownership at the scale Gusto's assistant will need to reach. - **Developer onboarding from 2–3 weeks to minutes (Intuit ICE Self-Service):** Delivered DevPortal, GitOps config, and ICE Playground, reducing time-to-production for developers while mitigating $1M+ in projected opex growth — the same "narrow slices that get customers value early" philosophy the JD calls out explicitly. - **aeval — AI model evaluation platform with adversarial safety testing:** Built CI/CD-integrated regression detection and automated safety gates, establishing a repeatable quality bar for model behavior — the infrastructure needed to hold quality as Gusto's assistant expands domain coverage. - **RL Workbench — 12-algorithm post-training benchmark:** Implemented standardized benchmarking across TRL, VeRL, OpenRLHF, and NeMo RL with cross-tab workflow lineage tracking — demonstrates the depth of AI systems reasoning the JD identifies as a differentiator for this role. ## Closing Gusto's bet is that small businesses deserve the same quality of financial and HR infrastructure that large enterprises take for granted — and that AI is the mechanism to close that gap at scale. That's a mission worth building toward carefully and quickly. I'd welcome the opportunity to discuss how my experience building trusted AI products, scaling developer platforms, and owning end-to-end AI system design maps to what the CoreX AI team is building. Thank you for your consideration. **O. Felix Amoruwa** famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info