Engineering Projects
Voxhelm — Screen-Aware Voice Assistant for Windows (Shipped — Founder)
Founder · voxhelm.de · Windows 10/11, macOS build in progress · Mar 2026 — Present.
"An AI that lives in your PC." You speak, and Voxhelm reads what is on your screen, then drives the real cursor, clicks and types — opening apps, controlling the browser, sorting files, managing windows, music and settings across 40+ actions. It narrates each step as it works and always asks first before anything it cannot undo.
It also teaches: ask how something works and Voxhelm draws directly on your screen — arrows, circles, step by step — instead of only describing it in words.
Shipped solo, end to end: a real-time push-to-talk voice loop, screen understanding, native Windows automation, an on-screen annotation overlay, Google sign-in with hosted inference (no user API keys), and subscription billing. The mic is live only while the user holds the shortcut, and no session content is retained server-side.
Site: voxhelm.de.
AI for Insurance — Automated Claims Handling Agent (Hackathon)
TUM.ai × EHL Hackathon · QuantCo "Claim to Fame" track · Munich · 22 — 23 August 2026 · 17 teams · 100 live rounds.
Built an AI agent that does an insurance claims handler's job end to end: it reads the policy, the damage report and the repair invoice, then rules on every invoice line the way a real adjuster does — is it covered, is it related to this damage, is the price fair — paying legitimate work in full while rejecting inflated bills, fully autonomously inside a 60-second deadline per claim.
Ran unattended for all 100 live rounds against 16 competing teams, ruling on ~18,500 incoming invoice lines at 73% correct accept/reject decisions. Mined the settlement record of already-closed claims into a 378-class price reference table that overrides the language model wherever a rate is proven, cutting per-line pricing error from ~60% to ~9%.
Unitem — Cross-Platform Design-Consistency Agents (Hackathon)
RAISE Summit Hackathon · Cursor track (Cerebral Valley) · Paris · 7 — 9 July 2026.
Teams shipping the same product on iOS and Android drift apart screen by screen, and a plain diff cannot tell a real inconsistency from each platform correctly being itself. Built an AI agent system that scans paired SwiftUI (iOS) and Flutter (Android) codebases and judges every design difference as propagate, hold, or flag.
Each verdict is grounded in an Apple HIG vs. Material convention base and in replayed human overrides, so the system learns which differences a team has already chosen to keep instead of re-raising them.
Code: GitHub.
Tactical Position Intelligence — Threat Zones on a 3D Battlefield (Hackathon)
European Defense Tech Hackathon (EDTH) · Munich · 26 — 28 June 2026.
Analyses the battlefield terrain directly from a georeferenced 3D point cloud and computes the risk zones an enemy position creates — line-of-sight, sectors of fire, and kill zones derived from each unit type and its weapon and observation constraints. The operator marks enemy positions from intel and the zones recompute automatically as those positions change, with a notification the moment one of our own soldiers steps into a threatened zone.
Demo: YouTube.
THIRD HAND — Voice-Commanded Assistive Robot Arm (Hackathon)
Robotics × AI Hackathon — Berlin (first edition) · NormaCore track host · ~29h · 20 — 21 June 2026.
Built a voice- and chat-commanded assistive robot arm: an agent (Claude on the Claude Agent SDK) that looks through the arm's cameras, plans the physical task, and executes it (pick, place, stack, drag, point), correcting itself when it fails. NormaCore's Station API is wrapped as an MCP server so the agent drives the real hardware in a perceive → decide → act → verify loop, not a scripted demo. Landed the first full autonomous pick-and-place on the physical arm, and fine-tuned a SmolVLA vision-language-action policy (released on Hugging Face).
Links: GitHub, SmolVLA model, demo.
StealthDetection — Digital Twins of Solar Plants (Hackathon — 1st Place)
Energy Hack Munich · ENERPARC × Invertix track · 1st place · Munich · 24h · 12 — 13 June 2026 · CDTM.
Won 1st place on the ENERPARC track. Solar plants are judged by a single Performance Ratio: it tells you THAT you are losing energy, never which inverter, why, or how much. We gave each of 65 inverters its own digital twin, trained on its healthy first year and run across 10 years of data. The gap between expected and actual output is hidden money: €245K of underperformance (11% of the plant's potential revenue), pinpointed to the inverter, explained by cause (downtime / hidden faults / grid curtailment), and priced in euros, with accuracy ~40% better than the industry-standard methods.
On top sits a grounded AI agent that does the cross-referencing analysts do by hand today, answering in plain language with every number tied to a real event, never guessed. It flagged 41 of 45 faults before the service ticket was even raised (median 10 days early) and uncovered €131K of losses with no ticket at all: money the plant was bleeding without knowing.
Live: stealthdetection-liart.vercel.app.
P2P Agentic Energy Sharing on Algorand (Hackathon — 2nd Place)
Algorand Builders Berlin · Agentic Commerce x402 Hackathon · 42 Berlin · 2nd place · 36h · 6 — 7 June 2026 · $21K+ prize pool.
Built a peer-to-peer energy-sharing platform where an autonomous agent transacts over x402 on Algorand mainnet: households sell their surplus solar to neighbours for any electricity use and pay only for what they consume, settled as real per-kWh micropayments in a MiCA-regulated digital euro (Quantoz EURD), with no intermediary.
Energy is priced dynamically by live solar output: the more sun (and the fuller the battery), the cheaper the kWh, and the buyer agent only purchases while the price stays under a user-set cap and budget, so you charge when energy is cheap and wait for surplus otherwise. Demoed end-to-end across a Raspberry Pi solar producer, the autonomous buyer agent, and a mobile app, with every purchase a genuine on-chain settlement shown via clickable explorer proofs.
Demo: YouTube.
SiliconGPT — Learning & Benchmarking Semiconductor Process Logic (Hackathon)
Zero One Hack: Vienna's Supercompute AI Hackathon · AI:AT — AI Factory Austria · Industrial AI / Infineon track · Vienna · 36h · 29 — 31 May 2026.
Built SiliconGPT, a 1.37M-parameter transformer trained from scratch (one fab process step = one token) that learns semiconductor process logic across four tasks: next-step prediction, sequence completion, anomaly detection (valid vs. rule-violating recipes), and out-of-distribution (OOD) generalization to unseen product families. No pretrained LLM and no API: it runs on a CPU, yet beats GPT-5, Gemini, DeepSeek and Qwen on the benchmark tasks while being ~1000× smaller.
Designed the model with a measurement-grounded multi-agent discovery loop (inspired by Google's AI Co-Scientist) plus a custom Experiment agent that trains and benchmarks every hypothesis on A100 GPUs. Ranking ideas on measured evidence rather than debate, it found that a smaller model generalizes better out-of-distribution.
Links: code + report, discovery-loop skill, live demo, video demo.
Tickmark — AI Reinforcement Layer for Professional Expertise (Buildathon)
Buildathon · Erasmus University Rotterdam · Netherlands · 24h · 18 — 19 May 2026.
As AI takes over the fundamentals in high-stakes fields like medicine and finance, professionals risk losing the judgment they originally trained for. Built voice agents that call people back after a task to ask contextual questions about their decisions and reinforce understanding, with an analytics layer tracking long-term knowledge retention and growth across teams.
Event: buildathon.eu.
AI-Native Business Management Platform for SMBs (Hackathon — 2nd Place)
Ignition: AI Hackathon · Entrepreneur Club EPFL × UNIL · 2nd place · Lausanne · 36h on-site · 10 — 11 May 2026.
Co-built an AI-native platform that takes a Swiss SMB still running on paper, Excel, scanned PDFs, and WhatsApp and turns it into a working operational system (clients, Swiss QR-bill invoicing, suppliers, appointments, inventory, expenses), all in French and Swiss-compliant out of the box. The killer onboarding move: upload your existing spreadsheets and the platform stands up your live back-office automatically.
Owned the backend and the AI surface: a voice agent that places outbound calls to clients with overdue invoices, negotiates payment dates in natural conversation, logs the commitment, and writes the result back to the books, with email and SMS follow-ups orchestrated end-to-end. Designed the Supabase backend with row-level security, the multi-vertical schema, and the document-AI ingestion pipeline for messy supplier paperwork.
AI Context Layer for Structural Engineering Projects (AI Pipeline)
Take-home prototype · May 2026.
Engineering firms hold years of project data in unstructured PDFs and spreadsheets, forcing planners to spend 3–6 hours per project manually digesting the folder before any design begins, with the most useful context trapped in personal notebooks.
Built an AI pipeline that turns an entire German-language project folder into a single citation-anchored markdown file per project, consumed directly by engineers through a web app and reusable as a grounded context layer for downstream AI agents.
Validated on 4 real engineering archives (60+ documents and 6 file formats per project); each run produces a bilingual (DE/EN) briefing with ~180 cited facts, automatic cross-document conflict detection, and severity-ranked open questions, collapsing hours of manual archive review into minutes of guided review.
Demo: YouTube.
Anytrace — AI Founder Detection for Early-Stage VCs (Hackathon)
TUM.ai × Yellow × Project A Hackathon · Munich · 30 April — 1 May 2026.
Co-built an AI deal-sourcing tool that detects founders before they show up in any startup database, by reading the public signals of a VC's trusted network and surfacing convergence patterns that point to early-stage talent.
Shipped an end-to-end product within the hackathon: ranked outreach feed, interactive network graph explorer, evidence-linked founder dossiers, and an automated daily digest.
Demo: YouTube.
Powerly — AI Renewable Energy Designer (Hackathon)
Tech Europe Big Berlin Hackathon · Reonic Track · Berlin · 24h · 25 — 26 April 2026.
Tackled a real installer pain point: residential renewable-energy system design today takes hours of manual tailoring per customer. Co-built an AI tool that produces a complete proposal in seconds (solar PV, battery, heat pump, and EV wallbox sized to the household), presented as three honest options (Budget / Balanced / Premium) with cost, self-sufficiency, payback, and a one-line reason per part, refinable live by the user.
Designed the engine as deterministic sizing rules + retrieval over 1,000+ real installations + an LLM for composition with safe rule-based fallback. Owned the roof intelligence and 3D experience: built a computer-vision pipeline over photogrammetry building models that detects roof planes and auto-places panels, and integrated Google's photorealistic 3D Maps so users could click directly on their own roof to add or remove panels and watch cost, autarky, and payback update live.
Links: GitHub, demo.
WhatsApp AI Lead Qualification & Scheduling Agent (Customer Deployed)
Production deployment for a fitness business · Mar — Apr 2026.
Shipped a production AI agent on WhatsApp: qualifies leads conversationally, proposes time slots, obtains coach approval, and sends Google Meet confirmations, deployed to a paying customer with zero manual intervention in the booking flow.
Built on Next.js, Vercel AI SDK, Neon Postgres, Upstash Redis, and Meta Cloud API; engineered stateful lead tracking, webhook deduplication, per-number rate limiting, human-handoff triggers, and Vercel Cron for automated reminders and follow-ups.
AI Factory Scheduling Agent (Hackathon)
Physical AI × Manufacturing Hackathon · Forgis · Zurich · 24h on-site · 27 — 28 Feb 2026.
Co-built an AI agent that schedules factory production orders on the Arke MES, takes operator approval over Telegram, and executes each phase on a robot arm with vision-based verification. Designed and implemented the scheduling algorithm as a mixed-integer linear program (Google OR-Tools) producing provably optimal sequences that minimise deadline tardiness across multi-phase orders.
Built the vision-verification module that checks the workspace state after each robot move before the system is allowed to advance; simulated the full schedule live on an SO-101 robot arm during the 24h on-site demo.
Real-Time Voice Meeting Agent for Discord (Hackathon)
Cursor Hackathon · Stuttgart · 24 January 2026.
Built an autonomous voice agent that joins Discord voice channels, engages in real-time conversation using STT/TTS, generates structured meeting summaries, and maintains persistent memory across sessions for longitudinal context.
YouTube Learning Companion — DeepNoteAI (Open Source)
Personal project · Feb — Jun 2025.
AI web app converting long-form YouTube videos into structured study sessions with segment-wise explanations and comprehension questions; multi-expert LLM pipeline with an async backend and FAISS-based vector retrieval. GitHub.
Stack
Frameworks: PyTorch, TensorFlow, HuggingFace Diffusers & Transformers.
Infrastructure: SLURM (multi-GPU H200/A6000), FP16 mixed-precision, Weights & Biases, distributed training.
Generative Modeling: Latent diffusion (SVD, DDPM, flow matching), ControlNet, VAE / VQVAE, autoregressive token models.
Computer Vision: Semantic segmentation (Cityscapes / KITTI-360), DRN-D-105, FID / FVD / mIoU evaluation.
Web & Apps: React, Next.js, Vercel AI SDK, Electron, Tailwind CSS.
Languages: Python, TypeScript / JavaScript, C / C++ (AUTOSAR / embedded).