01 • LLM AGENTS / TEXT-TO-SQL • LIVE DEMO
SQL Analyst Agent
A multi-agent LangGraph system that answers plain-English business questions over a SQL database. It routes the question, picks tables, writes SQL, checks it with an AST safety guard, runs it read-only, fixes its own errors and explains the answer with a chart. Risky queries pause for human approval.
100%execution accuracy (lenient) on 40 questions
5 / 5malicious prompts blocked
176tests in CI
$0to run
LangGraphLangSmithGroqsqlglotFastAPIDockerStreamlit
02 • LLM FINE-TUNING / DISTILLATION • IN PROGRESS
Distill-SQL
Distilling text-to-SQL skill from large LLMs into a 1.5B model that runs on a laptop. Teacher models (Gemini, gpt-oss-120b, Qwen) generate SQL, which is kept only if it runs correctly on real databases. The verified data trains Qwen2.5-Coder-1.5B with LoRA SFT and DPO, then the model is quantized to 4-bit GGUF and served with FastAPI behind a small-model-first router.
63.4%base model on full Spider dev (1,034 questions)
0.36 smedian latency, local on a laptop
2.8 GBpeak memory
$0per query, vs $0.15–0.77 per 1k for teachers
Qwen2.5-CoderLoRA SFTDPOGGUFllama.cppFastAPISpider / BIRD
View Project on GitHub ↗
03 • AGENTIC RAG
Agentic RAG Research Assistant
Upload PDFs or notes and ask questions. A LangGraph agent retrieves from ChromaDB, decides whether to use a calculator or web search, and answers with page sources. Runs locally on Qwen3 through Ollama, or on Groq in the cloud.
LangGraphChromaDBSentence TransformersOllamaGroqStreamlit
View Project on GitHub ↗
04 • LLM EVALUATION / COGNITIVE SCIENCE
Do Language Models Learn Like People?
LLMs act as participants in psychology experiments: a two-armed bandit and the Cognitive Reflection Test. I fitted Q-learning models to their choices and compared them with 44 human participants. qwen2.5:7b reached 0.90 accuracy on the human task design, but stuck to choices far more than people do.
OllamaQ-learningMaximum LikelihoodStatisticsCI
View Project on GitHub ↗
05 • DEEP LEARNING / BRAIN SIGNALS
Motor Imagery EEG Decoding
Decoding imagined left vs right hand movement from EEG across 104 people. EEGNet reached 68.3% on people it had never seen, against 61.9% for the classic CSP + LDA pipeline, with subject-level splits and significance tests.
PyTorchEEGNetCSP + LDAMNEStatistics
View Project on GitHub ↗
06 • MACHINE LEARNING / CYBERSECURITY
Network Security ML System
An end-to-end machine learning project for network security: data preprocessing, model training, evaluation, reusable ML artifacts and a FastAPI service in Docker.
PythonScikit-learnFastAPIDocker
View Project on GitHub ↗