Finished - Backtesting Strategy with Autoresearch - An automated machine learning pipeline that predicts the daily directional movement of the S&P 500 ETF. The system uses an autonomous Large Language Model agent to iteratively rewrite its own PyTorch neural network architecture, optimizing for improved Sharpe Ratio and Strategy Return in a simulated trading environment.
Finished - Agentic Data Scientist - A fully hands-off AI agent that acts like a data scientist: it inspects data, cleans it, runs statistical analyses, creates visualizations, and compiles a final Markdown report — all without human intervention.
Finished - RAG Application in Java - A Java application for parsing XML protocols from the German Bundestag, loading them into a Neo4j graph database, and enabling natural-language querying via an AI-powered RAG pipeline.
Finished - RAG Application in Python - A Retrieval-Augmented Generation chatbot tailored for extracting quantitative metrics and qualitative insights from SEC 10-K financial filings using natural language.
Finished - AI Assistant for Ikea Instructions using Agents - A smart assistant that helps users interpret IKEA assembly manuals — making complex instructions easier to understand and act upon. Designed as a full-stack application with AI-driven components to analyze assembly instructions from PDFs and images.
Finished - Multimodal Full-Stack Application - Multimodal Parliament Explorer is a Java-based, MongoDB-backed platform designed to autonomously ingest, deeply analyze, and interactively present parliamentary data from the German Bundestag. By combining web scraping, advanced Natural Language Processing, and dynamic document generation, it provides a comprehensive tool for political text
Finished - ETL Pipeline with NLP and Visualization - A comprehensive platform for processing, analyzing, and visualizing parliamentary protocols from the German Bundestag with advanced NLP capabilities and multimedia processing.
Finished - Video Emotion Recognition - Multimodal emotion and identity recognition over recorded video: a video is run through a DUUI pipeline that transcribes speech, diarizes speakers, and scores emotion from the text, audio, and video modalities independently, plus resolves who is on screen.
Finished - Sensitvity of the Autoresearch Loop - Studied how the different components of the program.md file influence the search as well as output behavior of Karpathy's autoresearch loop
Early Stage - Städel MCP - This server interfaces with the museum's OAI-PMH API using the LIDO (Lightweight Information Describing Objects) format, allowing AI assistants to harvest records, retrieve rich multilingual metadata, and access high-resolution
Early Stage - GoetheBrain - A full-stack, retrieval-augmented conversational persona of Johann Wolfgang von Goethe — grounded in essentially everything relevant to him on Project Gutenberg
Early Stage - Anomaly Detection - Implementation of the multi-agent document anomaly detection system using neurosymbolic ai
Early Stage - Citation Worthy Sources Claude Skill - A source-validation skill for Claude Code: evaluate domains before fetching, refuse content farms and hyper-partisan sites, and halt rather than cite junk.
Discontinued - Fine Tuning LLM For Financial Sentiment - An end-to-end real-time data pipeline for ingesting financial social media text, analyzing its sentiment using a fine-tuned Large Language Model powered by Apple's MLX, and visualizing the results.
-
Dataformat for ai agents: webmcp
-
control loops for SWE
-
knowledge base for long term memory
-
jepa: Instead of predicting missing patches of high-resolution images, one builds a miniature JEPA from scratch to predict the future state of a 1D time-series—like a noisy sine wave, daily temperatures, or simulated stock prices.
-
Autoresearch benefits across domains