cover_letter / art_QE_mTwJrrJg

Dear Character.AI Hiring Team, Character.AI is doing something genuinely rare: building a new medium from scratch, not optimizing an existing one. The ambition to define what interactive entertainment looks like for the next decade—where AI characters feel alive, distinct, and emotionally present—is the kind of product challenge that demands both deep technical fluency and genuine craft. I've spent the last two years building AI-native products from zero, and the last twelve building developer platforms at scale, and this role sits precisely at the intersection of those two arcs. **Technical and AI Foundation** My engagement with AI is not recent or surface-level. 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 same 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. That 19-million-fold scale increase in a single research thread reflects how I think about AI: longitudinally, architecturally, and with a bias toward building real systems rather than integrating APIs. More directly relevant to Character.AI's product challenge, I built an RL post-training workbench covering the full RLHF/DPO pipeline—implementing 12 algorithms (PPO, GRPO, DAPO, DPO, SimPO, KTO, ORPO, and others), benchmarking across TRL, VeRL, OpenRLHF, and NeMo RL with live SSE metric streaming. This is the infrastructure layer that sits beneath the frontier models Character.AI deploys. Understanding how reward functions shape model behavior, how GRPO diverges from DPO in convergence characteristics, and how to benchmark across frameworks is exactly the kind of fluency that lets a PM push research teams with the right questions—not just consume their outputs. 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, and statistical rigor including bootstrap confidence intervals, Welch's t-test, and Cohen's d effect size. When I partner with AI research, I can read the eval results, interrogate the methodology, and translate findings into product decisions. **The Bridge** Building AI products at the frontier requires two things that rarely coexist: the ability to understand what models can actually do (not what the paper claims), and the taste to translate that capability into an experience a person wants to return to. My path—from NeurIPS researcher to platform PM scaling 675M+ engagements at Intuit to founding AI-native consumer products—has been building toward exactly that combination. **Why This Role** What excites me about the Core Product role specifically is the framing around emotional resonance and session depth as the primary success metrics—not funnel conversion or short-term DAU spikes. The hardest product problem in AI entertainment is not making the model respond correctly; it's making the character feel present, consistent, and worth returning to. The conversation experience, character creation tooling, and creator ecosystem are the three surfaces where that problem lives, and they are the three areas I most want to work on. The explicit call for a PM who values craft and taste over dashboard optimization is the kind of role definition that signals a team building something they actually care about. **Selected Prior Experience** - Built production Electron + React + TypeScript desktop application with 100+ components and Redux Toolkit state management, providing real-time AI analysis using Claude (MCP SDK) and enabling contextual conversations across multiple industry verticals—shipping the full product from 0 to production, including code signing, notarization, and native macOS ScreenCaptureKit integration via Swift. - Implemented OpenClaw multi-agent orchestration framework with gateway protocol, subagent delegation, profile management, and session switching—enabling coordinated AI agent workflows across distinct character-like agent personas scoped to specific 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—building domain-specific conversational agents embedded in mobile and web. - At Intuit, delivered ICE Self-Service platform reducing developer onboarding from 2–3 weeks to minutes, achieved 275% YoY growth in ICE engagements scaling to 675M+ in FY23, and scaled throughput from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections at sub-25ms TP99—demonstrating the ability to own a platform product end-to-end at significant consumer scale. - Led customer discovery efforts at Fintellect AI, refining the platform based on trader feedback, establishing influencer partnerships, and executing go-to-market through App Store launch—closing the loop between user motivation research and shipped product. - Built AutoEval, an automated visual evaluation system for robot model training that repurposed a multimodal AI pipeline to score model outputs against natural-language rubrics, reducing evaluation cycles from 72 hours to ~4 minutes—an example of translating new model capabilities (multimodal spatial reasoning) into a concrete, measurable product workflow. - Conducted enterprise-wide Service Language Assessment at Intuit across 9 languages, analyzing usage data and developer feedback to inform strategic investment decisions presented to the CTO—demonstrating comfort operating in high ambiguity with incomplete data and cross-functional stakeholder complexity. **Closing** Character.AI's mission—empowering people to connect, learn, and tell stories through interactive entertainment—is one I take seriously as both a builder and a researcher. The characters people create and return to on this platform are not features; they are relationships. Getting that right, at the scale of tens of millions of monthly users, requires a PM who understands the model layer, respects the craft of the experience layer, and can hold both simultaneously. I would welcome the opportunity to discuss how my background maps to the specific challenges the Core Product team is navigating. Sincerely, **O. Felix Amoruwa** famoruwa@berkeley.edu | 909-731-9011 | felixamoruwa.info