What breaks when AI agents meet real work.

An independent research and engineering practice in London. We build open-source tools for the problems that appear once an agent is in real use: lost state, unchecked output, weak retrieval, runaway cost. The research is published in enough detail to reproduce.

LONDON

Open Source

Things we’re working on

Each of these started as a problem in a working system. Released projects are open source; the rest are in progress.

Memory, with judgement

Private product

AskZoe

A private AI assistant built around memory: what to keep, what to trust, and when to bring it up.

PERSONAL AIMEMORYRESEARCH

released

Felix

An agent harness that runs autonomous coding loops against real codebases without letting tests, types, or builds slide.

AGENT HARNESSQUALITY GATESCLI

released

Latch

A runtime for structured conversations: the model proposes progress, and a deterministic layer decides what gets recorded.

PROCESS RUNTIMECONVERSATIONAL STATETYPESCRIPT

released

RAGdoll

A RAG SDK for pipelines that have outgrown a notebook but do not need a managed retrieval platform.

RAGRETRIEVALKNOWLEDGE GRAPHS

released

Chomp

URL and HTML parsers that turn web pages into clean, LLM-friendly Markdown.

INGESTIONHTML TO MARKDOWNPYTHON

released

langchain-markitdown

LangChain document loaders built on the MarkItDown converter.

LANGCHAINDOCUMENT LOADERSPYPI

experimental

Echo

Fine-tuning Whisper for speakers with speech impediments, where the base model makes the most errors.

WHISPERFINE-TUNINGNOTEBOOKS

Writing

Notes and research

Longer write-ups, detailed enough to check or build on.

From the journal6 October 2026

Memory is a judgement problem: what we learned building it for an AI assistant

An account of how building Zoe changed our view of AI memory: from saving and retrieving facts to deciding what deserves to be remembered, trusted and brought into a conversation.

Read the story

From the journal16 September 2026

Strategic Workforce Planning Belongs to a Different Era

Plan from a defensible ceiling on what AI could do, not a headcount forecast

Read the story

October 2025OSF, open access

The Optimistic Speculation Strategy: A Three-Layer Approach to Reducing Conversational AI Latency

Conversational systems wait for end-of-speech before they start working. OSS starts earlier: when a user pauses mid-sentence, it launches a speculative call on the partial transcript and keeps the result invisible, so the model is already warm by the time the user actually stops. Nothing speculative is ever shown, so a wrong guess gets discarded rather than corrected on screen.

Read the paper

Conversational AISpeculative decodingLatency reductionSpeech recognitionPredictive inferenceTurn-taking