Tim Luka Horstmann

IP Paris Teaching & Research Assistant is a local-first, multi-agent LangGraph system built at Hi! PARIS. It blends course material RAG with live web research to deliver comprehensive answers and transparent sourcing for students, professors, and admins.

Highlights

• Local-first approach, leveraging open-source LLMs via Ollama.
• Router agent orchestrates RAG and web research paths with reflection and synthesis.
• Document pipeline with OCR, vector search, and inline citations for trustworthy responses.
• Role-aware dashboards featuring ratings, analytics, and real-time streaming UX.

GitHub Repository Explore Architecture & Setup

Pamel.ai is an AI‑powered yoga coach that gives you instant posture feedback while following any yoga video. Built in just 24 hours at the {Tech: Europe} Hackathon in Paris — where it won 1st place — it's now available as a free public beta.

Visit Pamel.ai Join Waitlist

T-REX is a live, interactive tool for verifying whether a textual claim is supported or contradicted by tabular data using state-of-the-art language models. It highlights relevant cells and streams reasoning steps in real-time to ensure transparency and accessibility.

This project is an advanced continuation of a course project from Language Models and Structured Data, where we explored early-stage table fact-checking systems. The paper has been accepted for publication at ECML PKDD 2025!

🚀 Original Prototype: Before T-REX became a published tool, I built an early version using transformer.js to perform table fact-checking directly in the browser. This prototype demonstrated the feasibility of client-side AI inference for table reasoning and laid the groundwork for the current system.

Try the live T-REX demo Watch the video demo Read Paper (arXiv - Coming Soon!) View the source on GitHub Try the Original Prototype (transformer.js)

My MPhil thesis project at the University of Cambridge, supervised by Professor Dr Paula Buttery, in collaboration with RegGenome and Cambridge Judge Business School.

About:
AI research project in regulatory technology (RegTech), utilizing advanced machine learning and natural language processing techniques. The project focused on improving regulatory compliance in financial institutions through innovative automated analysis of thousands of legal documents containing financial regulatory information.

A dedicated platform for large-scale subjective TTS listening studies with randomized pair presentation, inline playback, and a frictionless rating workflow secured behind user/admin passwords.

• Admin dashboard tracks progress in real time and exports clean CSV summaries in one click.
• Automated SendGrid email backups preserve every submission for audit-ready data integrity.
• Optimized for rapid study launches directly from Hugging Face Spaces.

Launch on Hugging Face

The Ultimate Personal CRM: ContactSphere is your private, all-in-one solution for mastering your network. Seamlessly integrate your Google Contacts and LinkedIn connections to visualize, organize, and leverage your relationships like never before.

View Code

An interactive chatbot integrated into this website, allowing users to ask questions about my CV and experience. Fully hosted on HF spaces.

• Powered by a quantized Qwen3 - 1.7B model (via llama-cpp-python) served via a FastAPI backend (currently running Gemini as HF spaces free tier resources are limited).
• Features streaming responses and conversation history management.
• Includes infrastructure for Retrieval-Augmented Generation (RAG) using FAISS/Sentence Transformers (experimented, but currently using full CV context for faster performance and full context).
• Optimized for CPU-only inference within Hugging Face free tier hardware constraints.

Try the Chatbot View Backend Space (HF)

This project was developed as part of the Machine and Deep Learning Data Challenge (CSC_51054_EP) at IP Paris. Our goal was to detect sub-events (like goals or red cards) from tweet activity during FIFA World Cup matches.

Approach: We implemented and compared three different strategies: traditional ML models (e.g., LightGBM, SVM), graph-based methods analyzing adjacency matrices, and LLM-based approaches. Our final model, the EnhancedPeriodClassifier, enriched BERTweet embeddings with handcrafted features and reached an accuracy of 77%.

Key Features:

• Preprocessing pipeline adapted from SOTA research
• Graph construction and deviation analysis
• Custom neural architecture combining pre-computed embeddings and multi-head attention
• Extensive hyperparameter tuning with Optuna

View Code
Read Report

A voice-enabled advanced AI assistant combining local LLMs, ASR (Whisper), TTS engines and external tool integration (MCP) for a seamless conversational experience.

• Wake-word detection & streaming text + audio
• Local inference via llama-cpp-python
• Multiple TTS engines (Kokoro, Coqui, FastTTS)
• Tool integration: Maps, Search, Weather, Custom functions

View on GitHub

Conducted a comparative analysis of two state-of-the-art constituency parsers, the Charniak-Johnson parser and the Berkeley Neural Parser, as part of Cambridge’s L95, "Introduction to Natural Language Syntax and Parsing" course. The project focused on evaluating performance and identifying error patterns, highlighting the advantages of modern neural parsing techniques. Implemented in Python using bllipparser and benepar libraries.

View PDF
View Code

Research project focusing on generating natural-sounding French speech using advanced deep learning techniques during my internship at Hi! PARIS, supervised by Eric Moulines.

A machine learning pipeline designed to analyze cardiac audio signals and classify heart murmurs. The project leverages advanced audio processing techniques such as MFCCs and spectral contrast, coupled with sensitivity-focused models, to enhance clinical decision-making in mobile health applications.

• Implemented audio preprocessing with bandpass filtering and resampling.
• Extracted key features (MFCC, spectral contrast, RMS energy) for classification.
• Trained multiple ML models emphasizing sensitivity for medical contexts.

View PDF
View Code

Development of a transformer-based machine learning model capable of providing abstractive summarizations to news articles as part of Cambridge's L90, "Overview of Natural Language Processing", course. Implemented in PyTorch.

View PDF
View Code

Worked with stakeholders across different organizations and drove resolution of the defect by setting up a QuickSight dashboard on top of Redshift ETL pipelines and using machine learning.

Developed a novel machine learning algorithm to optimize customer demand planning during the global semiconductor crisis. The system disaggregates high-level volume forecasts into granular material-specific predictions, replacing a heuristic SAP macro.

Key Achievements:

• Outperformed Baseline: Achieved a 45% higher R² and 43% lower MSE compared to the existing SAP logic.
• Big Data Processing: Engineered features from ~11 million data records spanning 102 months of history using Python (Pandas/NumPy) on AWS.
• Robust Validation: Implemented a Walk-Forward Validation (Rolling Evaluation) framework to ensure temporal consistency and prevent data leakage.
• Impact: The solution was handed over for production implementation in Q4 2022 to secure component availability.

Development of a computer vision model using YOLO for detecting fire-related objects in images. This project was part of a technical test for the Machine Learning Research Intern position at Hi! PARIS.

View PDF
View Code
View Experiments on wandb