Charles Patterson

• Built Penrose: open-source quant research engine with falsification core (30K+ LOC, 19 statistical gates, 10 executors, MCP server, UI)
• Developed Popper: qualitative thesis falsification framework
• Identified systematic mispricing in Kalshi weather tail markets (favorite-longshot bias 4.6x, fade strategy with OOS Sharpe +0.88)
• Implemented DSR (Deflated Sharpe Ratio), CPCV, parameter-fragility gates, and anti-mining deflation in a unified pipeline

• Engineered full-stack risk technology prototypes at the principal level — owning architecture decisions end-to-end and driving design pattern adoption in cross-team architecture reviews
• Used AI-assisted development tools (GitHub Copilot, Windsurf, Claude Code) daily and contributed to Capital One's shared body of autonomous/agentic development workflows
• Defined technical direction alongside senior engineers and product stakeholders as a regular contributor to principal-level design and architecture discussions

• Led architectural design for the company's next-generation SaaS platform — a portfolio performance comparison tool used by investment advisors to visualize and analyze risk
• Defined a multi-year technology roadmap through 2027, including strategic AI initiatives for the core product
• Led a small team of contractors alongside C-suite executives across early-stage development, research, and core product upgrades

• Founded and operated a multi-strategy digital asset fund; grew portfolio from $10K to a high-water mark of approximately $2M USD over 18 months through fundamental and quantitative analysis
• Designed and built production trading infrastructure on Solana: real-time execution engines, market-making bots, account watchers, multi-chain index construction algorithms, and on-chain/off-chain data ingestion pipelines
• Built pre-processing pipelines for MEV opportunity detection, sniping execution under latency pressure, and DeFi protocol monitoring
• Tech stack: Python (NumPy, Pandas, BeautifulSoup), Rust, TypeScript, PostgreSQL, WebSockets; deployed on AWS with Docker

• Led migration from monolithic architecture to microservices, increasing client prediction report fulfillment by 25% on monthly and weekly cadences
• Built a custom internal CLI for rapid non-Tableau reporting and prototyping, dramatically reducing time-to-insight for the business development team
• Expanded client-facing reporting capabilities through Vue.js UI development and REST API integration

• Developed VBA automation tooling for the IG credit desk serving sales, traders, and quants; flagship project saved 300+ hours annually through automated trade reconciliation and post-trade analysis
• Built and maintained KPI data reporting tools for management; led KYC onboarding and regulatory compliance for new and ongoing client relationships

• Open-source (Apache-2.0), v0.5.0, Python, ~30K lines across engine, UI, MCP server, and corpus management
• An LLM only proposes a strategy; deterministic code reconstructs it in a Docker sandbox and runs a power-aware robustness stack (deflated/probabilistic Sharpe, permutation and regime tests, walk-forward, parameter-fragility, tail-risk, cost/capacity) — the model proposes, the harness decides
• Validated the evaluator itself with a 5-arm control battery plus a multi-null specificity battery: 106/106 planted strategies detected, 0/300 false positives
• Engineered a discovery/confirmation firewall — a single-use locked holdout that burns after one confirmation, making benchmark contamination structurally impossible
• Extended to referee resolution-outcome prediction markets (Kalshi/Polymarket-style) via a dedicated event-market adapter; shipped a read-only MCP server exposing verdicts to LLM agents
• GitHub: PattersonResearch/Penrose

• Python, private project, the qualitative sibling to Penrose
• Enforces rigor on LLM-generated reasoning as hard write-time errors rather than after-the-fact warnings: the LLM proposes, code enforces every invariant
• Structural invariants: pre-registration (conditions locked and hashed before evidence), kill-first (no verdict without a refutation attempt), groundability (no accessible source → insufficient_data, never a fabricated verdict), and record-constrained rendering
• Decomposes a thesis into falsifiable claims, refutes each before confirming, and computes outcomes as a pure function of the verdicts — a grounded-generation / hallucination-guard pattern applied to analytical reasoning

• Founder — digitizes a user's closet (photo → structured inventory), builds an evolving taste profile, and powers wardrobe-aware outfit suggestions and shopping recommendations
• Shipped 50+ features across wardrobe digitization, AI styling, shopping discovery, social, and account management — pre-beta, invite-only launch
• Multi-model AI architecture: Gemini 2.5 Flash for garment vision/classification, Claude Haiku for conversational styling; Expo (React Native) frontend, Supabase backend, RevenueCat subscriptions
• Designed Twynn Connect — a B2B style infrastructure layer enabling retailers to embed wardrobe-aware "For You" feeds via OAuth-scoped REST APIs

• Designed and built on the Hermes agent runtime: a primary orchestrator routing to domain generals who delegate to 15+ specialist sub-agents — modeled on human prefrontal cortex hierarchy
• Built a persistent Neo4j graph memory system with semantic embeddings, trust decay mechanics, hit-count importance weighting, and tiered consolidation
• Dual-dispatch model routing (Claude Opus for reasoning/orchestration, GPT Codex for code generation) with git worktree isolation for safe parallel agent execution
• Designed an authority matrix governing autonomous vs. approval-required actions

Cyber-warfare and political interaction research for a joint MIT / Harvard / DoD project; maintained the Global System for Sustainable Development (GSSD); OEOEP community outreach mentor.