Engineering Projects
A selection of research and production projects focused on robotics, autonomous systems, and scalable AI infrastructure.
AMR-Agent — F-PLC-Supervised AGV Fleet Control
A production-grade warehouse automation system where a Siemens S7-1500 safety PLC supervises autonomous forklifts — ISO 13849 functional safety, VDA 5050 fleet management, and a planned LLM supervision layer, all demonstrated in Gazebo simulation.
- —5-layer architecture: Safety (ISO 13849, F-CPU) → Teleoperation (S7-1500, OPC UA) → Autonomy (waypoint router, pure pursuit, LiDAR guard) → Fleet (VDA 5050, MQTT) → LLM Supervision (planned)
- —Siemens S7-1500F fail-safe PLC — coded-processing F-CPU, two-channel input evaluation where required, PL d (ISO 13849-1) / SIL 2 (IEC 62061) target, 13 locked architecture invariants
- —In-house tricycle forklift model with 3 safety laser scanners + 360° navigation LiDAR, driven autonomously by an onboard waypoint router and pure-pursuit follower
Autonomous Delivery Vehicle - Area99
Last-mile autonomous delivery vehicle engineering — mechanical prototyping, steer-by-wire chassis kinematics, embedded vehicle control, ROS 2 integration, sensor fusion, simulation, and cloud infrastructure.
- —Chassis prototype designed and manufactured, with advanced steer-by-wire chassis kinematics
- —Contributed to the company's core mechanical IP generation
- —Embedded Vehicle Control Units bridging low-level actuation with the ROS 2 autonomy stack
Language-Conditioned Humanoid Manipulation on the Unitree G1
An end-to-end imitation-learning system that drives a Unitree G1 humanoid from natural-language commands and vision. Action-chunking transformer policies reach 100% task success in high-fidelity MuJoCo simulation, backed by a rigorously validated evaluation pipeline.
- —100% success on single-arm pick-and-place (50/50), replicated across 3 independent training seeds
- —100% success on bimanual box lifting with friction-only grasping and full contact physics — no attachment shortcuts
- —88–96% under domain randomization (lighting, distractors, camera jitter) with graceful out-of-distribution degradation
MuJoCo Robotics Lab
A 13-chapter blog series distilled into 9 core labs — from planar arms to VLA-controlled humanoid manipulation, building every layer of a robotics stack from first principles in MuJoCo.
- —13-chapter blog series organized into 9 core labs — from 2-link planar arms to VLA-controlled humanoid manipulation
- —Dual-engine architecture: Pinocchio (analytical brain) + MuJoCo (physics simulator)
- —Sub-millimeter tracking accuracy with computed torque and impedance control
BelegPilot: AI Receipt & Invoice Extraction API
Production document extraction API using multimodal models, orchestration, and observability from file upload to structured JSON.
- —Multi-provider LLM integration through LiteLLM
- —AWS deployment pipeline with containerized delivery
- —OpenTelemetry traces for production debugging and monitoring
Energy & EV Forecasting
Forecasting systems for charging-station occupancy, imbalance pricing, and solar production in the German energy market.
- —Production-deployed time-series models
- —MLflow experiment tracking and versioned model promotion
- —Containerized deployment workflows for energy-market forecasting
Normacore Hackathon: AI-Powered Robot Control
SmolVLA fine-tuning pipeline for autonomous ElRobot manipulation — teleop recording, Nebius GPU training, and hybrid Claude+SmolVLA deployment at CIC Berlin.
- —47 teleop demos recorded across 6 waffle positions with 2-camera setup
- —SmolVLA fine-tuned on Nebius L40S GPU — loss converged to ~0.05
- —Hybrid architecture: Claude orchestrates + verifies, SmolVLA executes
AI for Industry Challenge: Autonomous Cable Insertion
Solo robotics competition entry for the Intrinsic AIC — a hybrid learned + classical pipeline for sub-millimeter cable insertion on a UR5e arm.
- —Solo entry to Intrinsic's $180K AI for Industry Challenge — cable insertion on a UR5e with sub-mm precision
- —Hybrid architecture: ACT/Diffusion Policy for visual approach + classical impedance + spiral search for insertion
- —Custom LeRobot dataset pipeline — 1,200 demos, 274 MiB on disk, direct-to-parquet recording (no rosbag)
RAG Configurator: Low-Code Platform for Custom RAG Pipelines
Design, deploy, and test custom RAG pipelines through a visual 9-step wizard. Four microservices, two Vue 3 frontends, 9 agent architectures, full local operation with Ollama.
- —9-step visual wizard for full RAG pipeline configuration — no code required
- —9 agent architectures from Naive RAG to Graph RAG, all built on LangGraph
- —5 chunking strategies and 5 retrieval methods with fine-tuned parameters
Rookie — Personal AI Assistant
A 24/7 personal AI assistant running on a secure VPS, managing Obsidian notes, portfolio website content, emails, calendar, and daily tasks — all through Telegram.
- —24/7 Telegram interface with multilingual support (Turkish, English, German)
- —Automated portfolio content publishing with ISR revalidation
- —Obsidian vault management synced to iPhone via iCloud
Reinforcement Learning for Autonomous Robot Navigation
A TurtleBot3 that learns to navigate from scratch using Q-Learning — no pre-programmed paths, pure trial-and-error learning with LIDAR sensing and 20 hyperparameter configurations.
- —Q-Learning agent learns autonomous navigation from scratch — 144 discrete states, 3 actions, zero prior knowledge
- —20 experiment configurations: Softmax vs ε-greedy exploration, random vs fixed start, 5 hyperparameter scenarios
- —~85% crash-free navigation after 200 training episodes with smooth obstacle avoidance