About Me
Hi, I’m Torin. I’m a computer science graduate student at the University of Colorado Boulder with interests in computational complexity, theoretical machine learning, and optimization. I’m particularly interested in understanding which computational requirements are fundamental and which can be reduced by exploiting underlying structure. I also build software and machine learning systems, with projects spanning video processing, data analytics, and scientific computing. When I’m not programming, you can usually find me training Brazilian Jiu Jitsu or gaming.
Professional Projects
HDR to SDR
Visit Site →Desktop GUI application with over 2,500 downloads for converting HDR video to SDR using FFmpeg. Features a live side-by-side frame preview, multiple tonemappers (Reinhard, Mobius, Hable), and static or dynamic conversion modes. The Pro tier adds GPU-accelerated tonemapping via libplacebo (Vulkan), hardware encoding (NVENC / AMF / QSV), batch queue conversion, and a node-locked licensing system backed by hardware fingerprinting and HMAC-signed tokens.
AI Code Bot
GitHub →Open-source AI coding assistant. Contributed as a software engineer by integrating Ollama as a model provider, enabling offline coding assistance with any locally hosted LLM (Llama 3, Mistral, etc.). Abstracted the model interface for dynamic configuration and expanded the test suite to validate Ollama API responses.
Sankaku Analytics Closed Source
ML-powered retail analytics platform for CPG companies. Ingests point-of-sale data from Nielsen, SPINS, IRI, and Whole Foods, then clusters SKU portfolios via K-Means, visualizes commercial DNA through radar and PCA charts, and surfaces prescriptive trade-spend and distribution recommendations through a rules-based strategy engine.
NASA L'SPACE PDR
Download PDF →240-page Preliminary Design Review delivered to NASA as part of the L'SPACE Mission Concept Academy. Covers mission architecture, CDH subsystem design, trade studies, and a revised scope under a simulated 22% budget cut. Co-led as Deputy Project Manager and CDH Systems Lead.
Resume
Skills
Python, C, C++, PyTorch, scikit-learn, NumPy, pandas, Machine Learning, Deep Learning, NLP, Quantitative Analysis, Scientific Computing, Vulkan, libplacebo, FFmpeg, Hardware Acceleration (NVENC, AMF, QSV), Git, Linux, REST APIs, Test-Driven Development, Technical Documentation
Education
Projects
- Sole developer of a Windows desktop application for HDR-to-SDR video conversion and mathematical tone mapping.
- Designed a ground-up HDR-to-SDR color-processing pipeline implementing BT.2020 to BT.709 gamut mapping via a custom-generated 3D LUT with tetrahedral interpolation.
- Built GPU video processing pipelines using Vulkan and libplacebo with robust CPU fallback paths.
- Integrated automatic hardware encoder detection and acceleration for NVIDIA NVENC, AMD AMF, and Intel QSV.
- Engineered a thread-safe, asynchronous processing architecture with real-time tracking and rendering control.
- Developed offline hardware-locked licensing using HMAC-signed node-locked tokens, alongside automated installer, auto-updater, and code-signing release infrastructure.
- Maintained a 19-module automated test suite using unittest.mock to enforce a 90% code coverage floor across core processing paths.
- Shipped 26 production releases, with the application surpassing 2,500 downloads.
- Developed an Attention U-Net model using publicly available satellite imagery to estimate electricity infrastructure needs and map underserved regions in Sub-Saharan Africa.
- Achieved 75% segmentation accuracy on satellite remote sensing data.
- Applied explainable AI (XAI) techniques downstream of the prediction pipeline to interpret and visualize spatial feature representations influencing predictions.
- Received course-level "Future Prospects Award" in CS 178 (Machine Learning).
- Built an automated real-time statistical analysis system in Python for ingesting, scoring, and filtering high-frequency probability signals.
- Implemented a historical variance filtering model to analyze probability distributions and eliminate non-systematic noise signals.
- Developed a multithreaded polling architecture with variance estimation, slippage estimation, and consensus weighting for signal confidence scoring.
Experience
- Developed the MVP for an early-stage retail analytics startup project, building high-performance data processing pipelines and backend systems.
- Implemented and optimized K-means clustering algorithms for automated retail data segmentation and pattern discovery.
- Architected system backends and data architecture from the ground up, making technical decisions on software stack selection and data integrity.
- Analyzed the client's proprietary retail data and presented cluster analysis findings and actionable technical insights directly to the client's CEO, Head of Retail, and Data Analyst.
- Co-led a 12-person student engineering team simulating NASA Phases A and B for a lunar mission concept to analyze rock-ice samples in lunar Permanently Shadowed Regions (PSRs).
- Co-authored and delivered a 240-page Preliminary Design Review (PDR) and led mission scope revisions to fit a simulated 22% budget reduction.
- Defined CDH subsystem requirements and variable data-transmission rates (30s science payload / 1s telemetry) for critical mission phases.
- Conducted trade studies for avionics and mechanical components (IMUs, motors), selecting hardware based on thermal operating ranges (-193°C to 111°C) and power constraints.
- Analyzed market trends, financial and ESG information, and client and IRA-related data.
- Prepared reports and data presentations, assisted with office organization, and participated in professional and networking activities.
- Organized client data into a readily available physical resource to streamline client interactions.
Personal Projects
Open Source
OpenTurbo
GitHub →Experimental CUDA prototype for 3-bit KV-cache compression in LLM inference. Implements a TurboQuant-style FWHT encoder with QJL residual sign correction, GQA-aware scoring, and a downstream bridge targeting llama.cpp integration.
ShrinkWrap
GitHub →CLI tool that compresses AI agent instruction files (CLAUDE.md, .cursorrules, system prompts) into a token-optimized format. Keeps security rules and architecture constraints byte-identical via SHA-256 checksums while aggressively deduplicating volatile status and todo sections.
Quantitative Analysis Betting Bot
GitHub →24/7 automated trading bot for the Polymarket prediction platform. Implements a Smart Copy system that analyzes historical trader performance and filters luck-based variance to identify consistently profitable whales. Built on a multithreaded polling architecture with slippage calculation and consensus weighting. Simulated trades only; real execution is not implemented. Project is deprecated.
Heart Disease Prediction
GitHub →Binary classification project predicting heart disease presence from clinical features. Compares Logistic Regression and Random Forest on the Kaggle heart disease dataset, with full EDA, feature importance analysis, and ROC curve evaluation. Random Forest achieved 100% test accuracy on this corpus.
Closed Source
Discord YouTube Bot
Self-hosted Discord music bot with YouTube playback, queue management, and playlist support. Uses yt-dlp with browser cookie fallback for age-restricted content and backs off automatically when YouTube rate-limits a session.
Electricity Access Mapping
Attention UNet trained on IEEE GRSS 2021 satellite imagery to identify underserved settlements in Sub-Saharan Africa lacking electricity access. Achieved 75% accuracy with GradCAM-based explainability. Received the Future Prospects and Recognition Award for adapting a medical imaging architecture to satellite data.