Maximilian Köhlenbeck
Back to home

Projects

October 2026prototypein a team

Helpy

An assistant that watches an expert work, listens to their narration and asks about decisions at natural pauses. The session becomes a Work Map of steps, rules and reasons. A new colleague then practices with it, and mistakes get caught before saving. Screenshots are OCR'd and redacted right in the browser. Started at Hack-Nation 7.

Architecture

Capture

  1. Screen and voice
  2. OCR and redaction in the browser
  3. Questions at pauses

Teach

  1. Debrief and teach-back
  2. Work Map
  3. Practice with guardrails
Measurements: Helpy
Screenshot pipeline (sparse screen)378 ms end to end
Screenshot pipeline (dense screen)1,512 ms end to end
Captureevery 2 s, full-resolution frame

TypeScript · React · TanStack Start · Convex · Claude · ElevenLabs · Tesseract.js

Technical write-up

October 2026releasedSolo

Developer Island

A Dynamic Island for the Windows desktop. A small capsule at the top of the screen shows my Claude Code and Codex usage, music, a focus timer, meetings and system health. It reads the tools' local logs, sends nothing to the cloud and idles at 55 MB and 0.03 % of one CPU core. The code is open source.

Architecture

Data

  1. Local logs, media sessions and sensors
  2. Providers
  3. Immutable snapshots
  4. SQLite history

Display

  1. View models
  2. State machine
  3. Fixed-size transparent window
  4. Composition springs
Measurements: Developer Island
Idle, compact, 60 s window55 MB working set, 16 ms CPU (0.03 % of one core)
Warm start47 ms (only the changed transcript is re-read)
Tests382 unit tests

C# · .NET 10 · WinUI 3 · SQLite · xUnit

Code on GitHub Technical write-up

2026completedSolo

Wächter

An assistant that sees through a camera, listens and talks back. Everything runs offline on my laptop, no GPU, no cloud. Motion detection only wakes the person detector when something actually happens. Speech recognition only gets segments where someone is really talking, because Whisper invents sentences when it hears silence. The goal: it starts speaking in under three seconds.

Architecture

Vision

  1. Camera
  2. Motion detection
  3. Person detection

Voice

  1. Microphone
  2. Speech segments only
  3. faster-whisper
  4. Local language model
  5. Piper TTS
Measurements: Wächter
Time to first spoken word~2.6 s (target < 3 s)
Person detection17.3 ms per frame (58 fps)
Language model17–19 tokens/s
HardwareIntel i7-13700H, 32 GB RAM, no dedicated GPU

Python · faster-whisper · Piper TTS · local language model · ONNX / DirectML

Technical write-up

since May 2026prototypeTeam

Baulify

Started at the Smart Region Hackathon Offenburg. A platform where utilities, municipalities and contractors coordinate roadworks. It shows conflicts before the same street gets dug up twice. After the hackathon I kept building: an AI reads requests from PDFs or free text and pre-fills the form, approvals work through a link without an account, plus a map and a landing page.

Architecture

Requests

  1. PDF or free text
  2. LLM via OpenRouter
  3. Pre-filled form

Next.js · TypeScript · PostgreSQL · Drizzle · MapLibre · LLM via OpenRouter

2026completedSolo

Tarp editor

How do you mount a tarp so it doesn't wrinkle? In the editor you draw the shape, place mounting points and start the simulation. The tarp becomes a spring-mass mesh, wrinkles show up as lines and are measured with a wrinkle index. An assistant then suggests mounting point positions that make the tarp lie flatter.

Architecture
  1. Draw shape
  2. Place mounting points
  3. Spring-mass mesh
  4. Wrinkle index
  5. Suggested mounting points

JavaScript · Canvas · spring-mass simulation · a single HTML file

2026completedWith KUKA

3D printer: nozzle offset calibration

My longest project so far, together with KUKA. In FDM 3D printing, errors of a few tenths of a millimeter already showed in the parts, and the nozzle offset shifted with temperature and after nozzle changes. I built a calibration that measures the offset reliably and repeatably, separates real offset errors from causes like bed leveling or material behavior, and works as a process the team can rerun without me.

What I learned

Getting consistent results took many rounds of test prints and measurements.

FDM 3D printing · calibration · measurement

2026in daily useSolo

My Wispr Flow clone

The tool I use most out of everything on this page. Press a hotkey, talk, press again, and the text appears wherever the cursor is. In Word, in the browser, in the code editor. Modeled after Wispr Flow, except everything runs locally, costs nothing and sends nothing to the internet.

Architecture
  1. Hotkey
  2. Voice recording
  3. faster-whisper (small, int8, CPU)
  4. Text at cursor

Python · faster-whisper (small, int8, CPU) · Windows

2026completedSolo

AI automations with n8n

Three automations for small businesses. Extract invoices from PDFs and check that the line items actually add up to the total. Sort incoming emails and prepare replies as drafts, never send them automatically. Track prices and stock in online shops that have no API. When the AI isn't sure, a human decides.

Architecture

Invoices

  1. PDF invoice
  2. LLM extraction
  3. Line items vs. total
  4. Human decides when unsure

n8n · LLM via HTTP · Google Sheets · Airtable · Slack · Gmail

Code on GitHub