Samuel J. Baker IV

• Forecasting next-day realized volatility (risk, not returns) from open-licensed daily S&P 500 OHLC history (1927–2020), using OHLC-derived targets and robust lagged baselines.
• Ingesting and timestamp-aligning all major EDGAR filings (including 8-K for higher-frequency events) and earnings call transcripts; emphasizing domain-specific, context-aware sentiment/aspect signals over generic polarity.
• Training a regime-aware Temporal Fusion Transformer-style forecaster (regime indicators / mixture-of-experts) and benchmarking against HAR-RV and price-only baselines, with attribution back to specific disclosures and event windows.

Repo stays private until there is a proposal + baseline results (or an MVP).

• Developed a C++20 ECS game engine using SDL3 and OpenGL with an ImGui editor, asset pipeline and profiling tools.
• Implemented a real-time GPU path tracer via OpenGL compute shaders (progressive accumulation and A-Trous denoising), achieving 200-40 FPS at 1-8 samples per pixel on an RTX 4070 laptop GPU.
• Redesigned component storage from hash maps to dense arrays, reducing 100k entity add/remove operations from 99.8ms to 11.8ms (~88% reduction, 8.5x faster); and system-update time from 39.8ms to 6.93ms (~83% reduction, 5.7x faster); scaled to multi-million polygon scenes with ~27.8 ns component lookup.

Public repo. WIP; happy to walk through architecture and performance optimizations in detail on request.

• Built and operate a self-hosted homelab for local LLM inference and multiplayer game services, prioritizing privacy, transparency, and control over third-party cloud platforms.
• Implemented secure remote access over SSH and designed architecture for decentralized self-hosted services, with expansion paths for mesh VPN and DNS-layer filtering.
• Developed a custom UI for tokenization and model interaction to accelerate local inference, fine-tuning, and iterative model testing.

• Collaborated on a 5-member team to build a Dockerized full-stack parasite-screening app with HTML/CSS/JavaScript, Express.js, and Python/OpenCV, automating McMaster-slide analysis for sheep and goat samples.
• Designed a feature-based CV pipeline using grayscale blur, thresholding, contour detection, and ellipse/shape-color matching to estimate eggs per gram under iPhone 6 and low-connectivity constraints, avoiding CNNs due to limited labeled data.
• Improved threshold selection from a fixed value to modified Otsu with a translational shift, enabling correct health-condition classification on all test samples while reducing testing cost and turnaround versus the manual McMaster workflow.

• Built a chat-driven RPG arbiter that converts player dialogue into persistent state flags and event triggers, enabling branching narrative and adaptive encounters.
• Added lightweight memory and session persistence to maintain narrative continuity across multi-turn interactions in a text-adventure game loop.
• Prototyped agentic interaction patterns for games, focusing on stateful control flow and dialogue-to-action mapping.

• Ran controlled inoculation study on soybeans (n=60, 6 groups, 2 trials) under 70% / 100% / 130% water-holding capacity over 7-8 weeks.
• Found significant differences in shoot dry mass (p=0.007-0.023) and chlorophyll (p=0.014-0.032); presented at Chesapeake Bay Governor’s School symposium (VCU).