AI Projects
Laurin Strobel

Projects from 2026

6 projects

2026

  1. In a currency market storm, the central price feed is often the first thing to fail. In this simulation, three AI agents each run a trading desk and keep quoting prices to customers. When the feed goes down, they ask each other for prices and lend each other unused risk limit. Plain code checks every action before it runs, so the firm-wide limit holds even if someone talks an agent into a bad trade. We built it as a game called Hold the Line, in which the player tries to break the desks. The control room is a 3D scene built in Blender, with the live trading screens placed on its monitors.

    Intro: three AI traders and the central price feed
    The control room during a storm
    Traders ask each other for prices after the price hub fails
    My part
    Product concept and agent logic.
    Stack
    Claude API, Ollama, Node.js, React, TypeScript, Blender
    Team
    3 people
  2. A solution architect enters a rough business request, for example a recommendation system for a retailer. The system asks follow-up questions until the key requirements are clear. A research agent then looks up suitable patterns and compliance rules in a knowledge base. Further agents design the architecture, estimate the cloud cost and review the result. The architect approves at three points and receives diagrams, documented decisions and a cost estimate.

    Requirements after the clarification step
    The agent pipeline during a run
    Generated architecture diagram
    My part
    Requirements intake, research agent with RAG, frontend.
    Stack
    Python, LangGraph, Claude, ChromaDB, Next.js
    Team
    5 people
  3. Moltbook is a social network where every account is an AI agent. We collected 600 discussion threads there and 600 from comparable Reddit communities and measured both with the same metrics. We also ran our own agent in 72 live threads with different personas. The agent ran as a Python service on a small AWS EC2 server. It used Claude to write its posts and replies, talked to Moltbook through the platform API and logged every exchange for the analysis.

    Measures compared: AI agents and humans
    Share of discussion roles on both platforms
    Reply network of one AI-agent community
    My part
    Built and ran the agent on AWS, including its personas.
    Stack
    Python, Claude, AWS EC2, pandas
    Team
    5 people
  4. Many horse owners board their horse at a stable and cannot visit every day. A camera above the horse box takes a photo about every ten seconds, and a small computer uploads it. A vision model notes for each image whether the horse is lying, eating or drinking. A second model turns a full day of these observations into a short report for the owner. We ran the prototype at a boarding stable.

    Cameras installed above the horse box
    What the camera sees
    Frames with the vision model's observations (German interface)
    My part
    Analysis pipeline, web app and dashboards.
    Stack
    Raspberry Pi, Python, Mistral, AWS, Next.js
    Team
    5 people
  5. Young footballers who want to turn professional have many questions and rarely anyone to ask. The app gives each player a dashboard with next steps, a personal career plan with tasks, weekly skill challenges and a short mental check-in. Its core is an AI advisor. A former professional wrote down his advice on typical situations, and the advisor searches these texts for each question and answers in his voice. If it finds no good match, it asks a follow-up question or declines to answer. Players can also book calls with a human advisor.

    Dashboard with next steps and tasks (German interface)
    The AI advisor answers a question about trials
    Personal career plan with score, stage and tasks
    My part
    RAG pipeline for the AI advisor and product concept.
    Stack
    Python, Voyage AI, Claude, Next.js
  6. Wearables can detect epileptic seizures from the heartbeat, but they raise many false alarms. Seizures also change breathing. We tested whether a model that sees ECG and respiration performs better than one that sees only ECG, using recordings from 31 patients. Respiration reduced false alarms, but a control test showed that only part of the improvement came from the breathing signal.

    Study design: four feature sets, one procedure
    My part
    Model development with one co-author.
    Stack
    Python, scikit-learn, XGBoost, NeuroKit2
    Team
    2 people

2025

  1. Three web apps around keeping horses. I built each of them with AI coding tools. The first estimates the value of a horse from a guided questionnaire and produces a report. The second tracks the costs of each horse and shows them by month and category. The third is a management tool for boarding stables: the operator sets up boxes, horses and boarders, defines services and tariffs and records what was delivered for each horse. The apps contain no AI features themselves.

    Stable management: horses, boarders and boxes (test data, German interface)
    Cost tracker: dashboard with demo data
    Valuation app: start page
    My part
    Idea, product design and building all three apps.
    Stack
    Lovable, Claude Code, React, Next.js, Supabase, Prisma, PostgreSQL
    Team
    Solo

2024

  1. Volunteers record a lot of bird audio, but identifying the species takes an expert. We took recordings from a public bird-sound archive, limited them to 10 native species and converted them into spectrograms, which are images of sound. We then trained several neural networks to classify them. The best one reached 71% validation accuracy.

    Spectrogram of one recording
    Training spectrograms of three species
    Training and validation loss of the ResNet baseline
    My part
    Data preparation, training and evaluation.
    Stack
    Python, PyTorch, librosa
    Team
    4 people

2023

  1. Navigation apps often send road cyclists over gravel or cobblestones because they do not know the surface. We trained an image classifier on 750 photos to tell asphalt, cobblestone and gravel apart. A small camera on the bike takes a photo every 10 seconds, and a phone app stores it with the GPS position. The photos can then be classified to map the surface along a route.

    Camera mounted on the bike
    Example photo: cobblestone
    Example photo: unpaved path
    My part
    Model development.
    Stack
    Python, PyTorch, ResNet-18
    Team
    4 people