A small group(3) of students experimenting with LLMs, inference systems, RL, and CUDA. We build things to understand how they work, not to ship products.
We're focused on:
- LLM inference optimization and serving strategies
- Reinforcement learning for non-standard problems
- CUDA programming and GPU-level performance work
- Building end-to-end systems that combine multiple techniques
alphadesign — Hybrid AI framework combining reinforcement learning and genetic algorithms to optimize Formula 1 front wing aerodynamic designs. Features neural network-guided optimization, CFD analysis, structural constraint validation, and F1 regulatory compliance checking for accelerated design iteration.
frugalsot — An adaptive model selection system for efficient on-device NLP inference, enhancing speed, privacy, and resource use on edge devices.
phydra — End-to-end cargo management system with advanced 3D bin packing algorithms, A*/Dijkstra pathfinding implemented in C++ for ISS applications.
gideon — Emotion-aware LLM chat interface with Ollama, FastAPI, and React, featuring real-time image generation and project automation via AlphaStack agent. Accelerates full-stack and blockchain dev by 60%, blending local/cloud models for empathetic, context-rich interactions.
specquant — Scalable framework for adaptive LLM serving: classify prompt complexity → select quantized drafts → verify with FP16 target, no model retraining required.
alphastack — A universal, AI-powered development agent that supports any tech stack—battle-tested across 25+ full dev cycles. It intelligently scaffolds and iterates on complex projects with automated feedback loops to accelerate software delivery.
If you're working on similar problems or have ideas worth testing, we're open to collaboration. Code-first, lightweight experiments preferred.
