Applied ML Research Engineer

Bertil Braun

I build end-to-end machine learning systems across reinforcement learning, computer vision, LLM systems, and graph-based control, with a focus on problems where modeling, infrastructure, and evaluation all matter.

Projects

Selected work from independent research, applied ML systems, and LLM evaluation.

RL self-play, graph control, GPU systems
CV YOLO, tracking, video pipelines
LLM evaluation, agents, durable workflows
Systems deployment, observability, orchestration

A general AlphaZero-style system for learning board games from self-play. The project combines neural network training, Monte Carlo Tree Search, distributed data generation, and performance-oriented inference code.

  • General game framework spanning Tic-Tac-Toe, Connect Four, Checkers, Hex, and Chess.
  • Batched GPU inference and optimized C++ MCTS for self-play and evaluation.
  • Training and evaluation loop built around measurable playing strength rather than supervised labels.
GitHub repo
Chess engine example game animation Example game from the project documentation.

A computer vision product for turning raw windsurfing footage into usable rider views. The project covers the full path from training detectors to serving GPU jobs and shipping an interactive web player.

  • YOLO-based detection and pose models trained for windsurfing footage.
  • Multi-object tracking, stabilization, and focused playback for individual riders.
  • React/Firebase frontend with FastAPI backend and Modal-powered GPU processing.
GybeLock site GitHub repo
GybeLock processed windsurfing demo Processed demo from the Windsurf Analysis documentation.

A set of agentic LLM systems built around explicit service boundaries, retrieval, memory, tools, and observability. The largest system is an AI story narrator backed by specialized agents, tool servers, vector memory, and a knowledge graph.

  • Layered microservice architecture with React, FastAPI, Redis, agent services, tool servers, and provider services.
  • Retrieval and memory across PostgreSQL, Qdrant vector storage, and Neo4j knowledge graphs.
  • Durable coding runtime applies workflow orchestration to long-running agent tasks with tests and review gates.
DnD GM platform Coding runtime Temporal-Light
Temporal-Light dashboard Temporal-Light dashboard; the agentic systems use durable workflow orchestration as a control plane.

Research project on extracting structured competency profiles from unstructured documents and evaluating LLM-generated outputs without relying on a single fixed reference answer.

  • Multi-phase extraction pipeline for scientific and technical documents.
  • Fine-tuning workflow using synthetic preference data and DPO.
  • Automated evaluation framework based on pairwise comparisons and agreement thresholds.
GitHub repo
LLM competency extraction data flow System data flow from the thesis documentation.

Graph reinforcement learning system for traffic-signal control. The project focuses on learning policies that operate over road-network structure instead of hard-coded intersection layouts.

  • SUMO/TraCI environments with graph policies controlling junctions jointly.
  • Imitation-learning warm start followed by reinforcement learning refinement.
  • Movement-level action representation aimed at transferring across arbitrary OSM junctions.
GitHub repo

From-scratch Soft Actor-Critic implementation in pure JAX. The project demonstrates a fully GPU-resident RL pipeline with vectorized environments, GPU replay buffer, neural networks, and training loop.

  • Zero host-device transfers during training.
  • Up to 4096 parallel CartPole environments stepped as one batched operation.
  • GPU-resident training loop reaching thousands of parameter updates per second.
GitHub repo DeepWiki
GPU-resident RL performance benchmarks Performance benchmark plot from the project documentation.