cover_letter / art_eb64aKCx7ME

Dear Perplexity Hiring Team, Perplexity is doing something rare: building a product that treats curiosity as infrastructure. The shift from search-as-lookup to Computer-as-agent — where knowledge becomes action — is exactly the kind of platform inflection I have spent the last several years working toward. When I built the OpenClaw multi-agent orchestration framework at Streamio AI, coordinating subagent delegation, session switching, and profile management across real estate, insurance, and financial-markets workflows, I was solving the same core problem your AI Capabilities team owns: how do you turn nondeterministic intelligence into repeatable, composable experiences that users can actually trust and control? **Technical and AI Foundation** My engagement with AI systems is not recent. In 2004 I hand-coded a backpropagation-through-time neural network in C++ for protein structure prediction at UC Berkeley — work that led to a NeurIPS 2014 acceptance. The 2026 rewrite of that same system spans 413 parameters to 8 billion (a 19-million-fold scale increase), built in PyTorch with five neural architectures, MLflow experiment tracking, Optuna hyperparameter optimization, and FastAPI serving across six Docker containers with 823 automated tests. That arc — from first principles to production-scale — shapes how I think about AI product work: rigor at the model layer is a prerequisite for reliability at the product layer. More directly relevant to this role, I built a full RL post-training workbench covering the complete RLHF/DPO pipeline: a Reward Lab for designing and A/B testing reward functions across GSM8K, MATH, HumanEval, and UltraFeedback; a Playground running real TRL-powered GRPO and 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 — PPO, GRPO, DAPO, REINFORCE, REINFORCE++, RLOO, DPO, SimPO, IPO, KTO, ORPO, SPPO — with standardized throughput, memory, and convergence benchmarking. This is the kind of evaluation infrastructure that lets a team steer nondeterministic models toward high-value outcomes, which maps directly to what the AI Capabilities team needs to do with Skills and Computer. I also 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, bootstrap confidence intervals, Welch's t-test, Cohen's d effect size, and CI/CD regression gates — the statistical rigor required to make capability improvements measurable rather than impressionistic. **Why This Role** The through-line in my work is turning AI primitives into products that feel coherent to users who did not build them. At Intuit, that meant scaling a developer platform to 675M+ engagements and reducing onboarding from weeks to minutes. At Streamio, it meant building multi-agent workflows that non-technical users in real estate and financial markets could actually operate. The AI Capabilities role at Perplexity is the natural next step: owning the product structures — Skills, Projects, artifact generation, collaboration — that determine whether Computer becomes a platform people build on or a feature people occasionally use. **Role-Specific Connection** The JD's framing of Skills as teachable, reusable behaviors that compose into larger workflows is precisely the design problem I find most interesting. The question of how users discover, customize, and share capabilities — and how the platform surfaces the right defaults while remaining extensible — is one I have been working on concretely through OpenClaw's gateway protocol and Vantage's 40+ agent tool library with enforced mutation approvals. I am particularly drawn to the artifact generation and Projects surface areas: the challenge of turning an ongoing, context-rich conversation into a polished, shareable output with clear provenance is something I built toward with Vantage's ProofReel verified candidate profiles and Fintellect's structured AI advisory outputs. **Selected Prior Experience** - Built OpenClaw multi-agent orchestration framework (gateway protocol, subagent delegation, profile management, session switching), coordinating AI agent workflows across real estate, insurance, health/dental, and financial-markets industries — directly analogous to the composable Skills and connectors architecture the AI Capabilities team is scaling. - Shipped Vantage on iOS and macOS with a shared FastAPI backend, 40+ agent tools, enforced mutation approvals, streaming agent chat via SSE, and per-role artifact generation (interviewer questions, prep canvas) backed by async job polling and idempotent endpoints. - Achieved 275% YoY growth in ICE platform engagements at Intuit, scaling to 675M+ in FY23; reduced developer onboarding from 2–3 weeks to minutes; scaled throughput from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections at sub-25ms TP99. - Built aeval evaluation platform with adversarial safety testing, statistical significance testing (bootstrap CI, Welch's t-test, Cohen's d), and CI/CD regression detection — the measurement infrastructure required to steer model and agent behavior toward reliable outcomes. - Architected RAG retrieval pipeline at Fintellect with multi-provider LLM orchestration (Claude, GPT-4, Gemini), fallback routing, structured-output validation, and token-budget optimization across 13 specialized AI advisors. - Extended Java and Python SDK Starter Kits at Intuit with scaffolding templates, build configurations, testing frameworks, and CI/CD integration — empowering developers to reach production-ready microservices in minutes; built Asterias, a declarative asset lifecycle management platform with GraphQL API. - Built AutoEval, an automated visual evaluation system for robot model training that reduced evaluation cycles from 72 hours to ~4 minutes using a zero-integration screen-capture architecture with multimodal AI scoring against natural-language rubrics. **Closing** Perplexity's mission — powering curiosity through a continuous cycle of learning, building, and integrating — is one I take seriously as a practitioner. The people who drive change are the ones who can move fluidly between understanding a system deeply and making it accessible to others. That is the work I have been doing across developer platforms, AI frameworks, and agentic products for over a decade, and it is the work I want to do at Perplexity. I would welcome the opportunity to discuss how my background maps to the specific challenges the AI Capabilities team is navigating. Sincerely, **O. Felix Amoruwa** famoruwa@berkeley.edu · 909-731-9011 · felixamoruwa.info