
"I build the infrastructure layer that turns messy tech into real life solutions within production-grade systems."
software and machine learning @ German Aerospace Center engineering future-ready road and rail vehicles
building unicorn @ brikli (inbound vc interest, pre-seed), the agentic infrastructure layer for real estate
training semantic ingredient state tracking models @ wat.ai
applied Stanford's RSI agent harness to a coding-agent domain in 48 hours
creating a personal brand to document my travels across europe (50k+ views)
seeking winter 2027 internships
click into a project for code snippets, detailed system design, and the personal story behind each build
Projects



Self-improving infrastructure layer for coding agents that autonomously proposes, evaluates, and rewrites agent harnesses based on real execution traces. Built as composable state machines, it transforms harness optimization from a linear iteration loop into a branching search process with sandboxed evaluation, deterministic checkpointing, and time-travel forking.
Hack the Valley X Winner

Rainbolt.ai is an AI-powered geolocation system that uses CLIP embeddings, Pinecone vector search, Gemini reasoning, and Mapillary verification to predict where a photo was taken. It streams its reasoning live, highlights visual clues directly on the image, and displays the final prediction with sourced reference imagery on an interactive globe.
uOttaHack 7 Winner
.0v~dhedd9qh0y.jpg&w=3840&q=75)
An AI-powered defense intelligence platform that combines interactive 3D military equipment analysis with Canadian Arctic maritime patrol simulations. It uses VLMs to identify components, generate tactical assessments, and visualize threats, trade routes, mission events, and strategic infrastructure on a real-time globe.
DeltaHacks XII Winner
VibeTrade is a real-time BTC trading terminal powered by an agentic LLM, live market data, and a voice-first interaction pipeline. It ingests multiple data streams, computes a risk score every second, and uses a LangGraph agent to inspect market conditions, analyze sentiment, and execute paper trades through Alpaca.
Ignition Hacks VI Winner
A crime-aware navigation platform that evaluates multiple Toronto routes using geospatial crime data, OSRM pathfinding, and dynamic risk scoring. It visualizes incidents through clusters and heatmaps, compares routes by safety, distance, and duration, and recommends the lowest-risk path in real time.

Hack the Valley X Winner

Rainbolt.ai is an AI-powered geolocation system that uses CLIP embeddings, Pinecone vector search, Gemini reasoning, and Mapillary verification to predict where a photo was taken. It streams its reasoning live, highlights visual clues directly on the image, and displays the final prediction with sourced reference imagery on an interactive globe.
DeltaHacks XII Winner
Ignition Hacks VI Winner
A crime-aware navigation platform that evaluates multiple Toronto routes using geospatial crime data, OSRM pathfinding, and dynamic risk scoring. It visualizes incidents through clusters and heatmaps, compares routes by safety, distance, and duration, and recommends the lowest-risk path in real time.


Self-improving infrastructure layer for coding agents that autonomously proposes, evaluates, and rewrites agent harnesses based on real execution traces. Built as composable state machines, it transforms harness optimization from a linear iteration loop into a branching search process with sandboxed evaluation, deterministic checkpointing, and time-travel forking.
uOttaHack 7 Winner
.0v~dhedd9qh0y.jpg&w=3840&q=75)
An AI-powered defense intelligence platform that combines interactive 3D military equipment analysis with Canadian Arctic maritime patrol simulations. It uses VLMs to identify components, generate tactical assessments, and visualize threats, trade routes, mission events, and strategic infrastructure on a real-time globe.
VibeTrade is a real-time BTC trading terminal powered by an agentic LLM, live market data, and a voice-first interaction pipeline. It ingests multiple data streams, computes a risk score every second, and uses a LangGraph agent to inspect market conditions, analyze sentiment, and execute paper trades through Alpaca.
