A selection of my work in agentic AI, RAG systems, LLM evaluation, NLP, computer vision, and quantitative analysis. Every project links to its source code on GitHub.
Four LangGraph agents (Planner, Searcher, Synthesizer, Writer) produce citation-backed research reports. A local Qwen 3 model handles planning and synthesis while Gemini writes only the final report, cutting API requests ~74% and input tokens ~80%.
A reasoning agent routes sub-queries across a vector DB, SQL database, knowledge graph, and live web search, fuses the evidence, and runs a Verifier/Critic node that automatically retries low-confidence answers.
A benchmark suite that tests the Agentic RAG Engine against a single-pass RAG baseline using 24 ground-truth questions, an auditable LLM-as-judge rubric, and statistical analysis of groundedness, correctness, latency, and cost.
A multi-factor stock ranking pipeline that computes technical signals (RSI, MACD, EMA, ATR, OBV, VWAP), scores stocks across the universe, and measures signal uncertainty with Monte Carlo simulation and Bayesian updating.
Can an LLM beat traditional ML on structured data? A quantized Mistral 7B classifies whether NYC taxi riders tip over 20% by reasoning over trip data converted into natural-language descriptions, reaching ~92% validation accuracy.
A domain-adapted machine translation model, fine-tuned from Helsinki-NLP's opus-mt-de-en on legal and university terminology so domain terms translate correctly (e.g. “Matrikelnummer” → “Student ID”).
A fully local RAG pipeline for chatting with PDF documents without sending data to the cloud, using Gemma 3 via Ollama, a Dockerized Weaviate vector DB, and OCR support for scanned files.
Generates training labels automatically with no manual annotation: a detection model finds objects, SAM 2 refines each one into a pixel mask, and geometric fitting turns the masks into YOLO oriented bounding boxes.