Home Projects

Projects

Evaluation tooling built from the requirements of my own review work, and software projects built to solve a specific problem end to end.

Evaluation toolingSep 2026

TTS Review Tool

  • Live demo
  • Open source
  • In daily use

A keyboard-driven, single-page interface for text-to-speech A/B evaluation, built from the requirements of my own TTS review work.

Problem
Paired audio reviews are slow with a mouse and generic tools, and terminology drifts over a long session.
Solution
Play, score on a −2 to +2 CMOS scale, tag reasons, comment and advance without leaving the keyboard. Comment snippets keep wording consistent; progress autosaves.
Stack
Vanilla HTML, CSS and JavaScript. No build step, no dependencies, storage behind a swappable key-value layer.
Result
Used in my live review sessions. Scores, reasons and comments export as TSV or CSV straight into a spreadsheet.

The public demo is sanitized for NDA: placeholder text, no audio, no client data.

Open live demo (opens in a new tab) Source on GitHub (opens in a new tab)

CMOS, A vs. Bkeys 1–5

Keymapfrom the README

  • Play A → BSpace
  • Play only A / only BAB
  • Set CMOS score1–5
  • Toggle reason tagsQ…]
  • Previous / next item←→
  • Confirm comment & advanceEnter

Schematic of the review loop. Try the real interface in the live demo.

SoftwareApr – Jun 2026

Real Estate Price Prediction Tool

  • Open source
  • Full stack + infrastructure

Property valuation for the Karlsruhe/Bruchsal region with a 10-year forecast.

  1. Local market and macro data
  2. Ensemble model
  3. LLM adjustment from listing text
  4. Valuation and 10-year forecast
Problem
Much of what moves a property’s price is only described in free text, not in structured data.
Solution
An ensemble model over local market and macro data, plus an LLM step that turns listing descriptions into a bounded, structured price adjustment.
Stack
  • Python
  • Flask
  • SQLite
  • Groq LLM
  • Chart.js
  • AWS
  • Terraform
View on GitHub (opens in a new tab)

AI systemMar 2026

TikTok Clip Analyzer

  • Open source
  • LLM pipeline

Finds the clip-worthy segments in long YouTube videos and explains each pick.

  1. YouTube transcript
  2. Chunked segments
  3. Rubric scoring by Llama 3.3
  4. Timestamped, justified clips
Problem
Scanning a full video for short-form clips is manual and hard to justify.
Solution
Transcribes and chunks the video, scores each segment on a weighted rubric (hook strength, standalone value, emotional impact, shareability) and returns timestamped, score-justified suggestions.
Stack
  • Python
  • Streamlit
  • Groq API
  • Llama 3.3 70B
  • YouTube Transcript API
View on GitHub (opens in a new tab)

Evaluation case studiesConstructed examples, NDA-safe

Two outputs that look equal, until you read them.

Six recurring failure modes and two worked reviews, a German HR translation and a Python complexity claim, walked through line by line in the format I use for real client work.

Read the case studies

Need something like this built or evaluated?

Review interfaces, LLM-powered workflows, structured export pipelines and small internal apps. Open to freelance and contract work.

Discuss a project See services