Most problems aren't model problems — they're system design problems in disguise.
CS student building systems that don't just work, but make sense. I care about the gap between "it runs" and "it scales" — and I spend most of my time in that gap.
SpendSight · Financial Transaction Intelligence
4-stage hybrid classification pipeline (regex → heuristic → MiniLM → LLM) for financial NLP.
Cut LLM inference costs by ~76% without sacrificing accuracy on ambiguous transactions.
Python NLP HuggingFace Cost-Aware ML
Topology-aware MCTS with a custom DifficultyAllocator using uncertainty signals for adaptive compute routing.
Exploring how board topology interacts with tree search efficiency.
RL MCTS Python Research
VMARO · Vectorless Multi-Agent Research Orchestrator
8-stage autonomous research pipeline with a Streamlit frontend. No RAG — pure LLM-native retrieval and synthesis.
Built a preprocessing layer (Stage 00 Topic Normalizer) and relevance scoring for PubMed paper ranking.
Multi-Agent Streamlit Groq Moonshot AI
- Hybrid inference systems — confidence-threshold routing between cheap and expensive models
- Cellular automata as computation — maximal-cycle CAs, ΔG decomposition, unified cycle theory
- Efficient MCTS — topology-aware search without blowing compute budgets
- Reasoning benchmarks — what they actually measure vs. what we think they measure
My default stack is whatever fits the problem. I'm not religious about tools.
Primary: Python · JavaScript · Go · Bash
ML/AI: PyTorch · HuggingFace · LangChain · Anthropic SDK
Infra/Other: Streamlit · FastAPI · Docker · Linux
I'm open to research collaborations, ML systems roles, and interesting problems — especially anything at the intersection of efficiency, inference, and system design.