# Lab notes and open-source work

Canonical: [Lab notes and open-source work](https://humanfacetech.com/lab)

## NicheBench: evaluating LLMs on framework-specific tasks

[Read the write-up](https://humanfacetech.com/lab/nichebench)

How our Python evaluation harness uses Drupal tasks, judge rubrics, and optional runtime checks to examine model behaviour in a framework context.


## NicheBench

[Open repository](https://github.com/HumanFace-Tech/nichebench)

Our Python CLI for evaluating LLMs on framework-specific tasks, starting with Drupal. The public repository contains the harness and documentation; task data needs separate access.

- Quiz, code-generation, and bug-fixing tasks with a model under test and a judge.
- Opt-in isolated Drupal/DDEV runtime checks, collected artifacts, and a final diff.
- Task-specific scoring rubrics, configurable weights, and critical-check gates.

## Whisper Recorder UI

[Open repository](https://github.com/HumanFace-Tech/whisper-recorder-ui)

Our browser recording PWA sends audio to a configured speech-recognition endpoint. You choose a reachable compatible server and model, such as Whisper or Qwen3-ASR; models run on the server.

- Record audio in the browser and send it to a local or OpenAI-compatible ASR endpoint.
- Add hints for technical vocabulary. You could configure LangChain or Langfuse hints when those terms are misheard.
- Optionally process the transcript with an LLM, keeping raw and processed text separate. Installable on supported browsers; transcription requires a reachable server.

## Start a conversation

For a faster route, [Book a 30-minute session with us](https://calendar.google.com/calendar/u/0/appointments/schedules/AcZssZ3MvmJCrtltllV0bv1h7BMvEUHyLWWS76p5viktzbOT5Jhs-dvfPRvTpFJ3Nt6X8CqHCM1f1q_M). Otherwise use the [contact form](https://humanfacetech.com/contact).