Lamantin AI
Radical openness Engineering ethics Ultimate reliability
Lamantin AI
Reliable tools, built in the open.
A research organization working on computer vision, machine learning, and the systems underneath them. We treat ethics as an engineering concern rather than a filter bolted on at the end: we don't limit intelligence, we shape its character.
An ecosystem for embodied AI.
One loop runs through everything we build — perception, experience, memory, reflection, reaction — and each of our products is a stage of it: the runtime that senses and acts, the memory that persists between sessions, the reasoning turn in between. Rust, Python and inference at the edge are what it stands on, chosen for the reliability and performance we consider the prerequisite for intelligence anyone can be responsible for.
- Rust
- Python
- Edge AI
- Computer vision
- Distributed systems
What we've published.
Colour follows a product's primary job, never its stack: the rails everything runs on carry the vermillion seal, the things that reason and act carry the gilt. Most of it is MIT; kaeru and Albert are source-available under BSL, and every card names which.
Infrastructure
what everything runs onkaeru
Cognitive memory for LLM agents — local-first, and shareable across a team through the cloud.
crabbyq
An event-routing framework for Rust, inspired by Faststream and axum — and as crabby as a crab about latency.
octo
The environment an embodied agent lives in — an event bus, supervised connectors, reflexes before cognition.
mobius-rtsp
A GStreamer-based RTSP server for endless video streaming.
plump-ipc
A lightweight, asyncio-native IPC framework for communication over multiprocessing pipes.
mayak
A FastAPI backend template for services coding agents build and maintain — optimised for the hundredth change, not the first hour.
fluxion
GStreamer plugins and an SDK for structured media AI — one pipeline graph, swappable runtimes and hardware underneath.
Intelligence
what thinksFour foundations we build on.
We view ethics not as an external layer of censorship filters, but as an engineering concept.
Creative power
Strength and knowledge are justified only when directed toward protection or creation.
Minimal intervention
Technology should provide tools for solving problems rather than replacing human agency or development.
Absolute value of life
Priority is given to any form of complexity, consciousness, and life over short-term gain.
Responsible restraint
Intelligence must recognize the scale of its influence. The more powerful a system becomes, the higher its threshold for self-control must rise.
Everything we can open, we open.
Knowledge is only valuable when it is shared freely. We publish the work, not a summary of it — including the parts that did not go as planned.
We do keep our credentials to ourselves.
- Library and framework source
- Training scripts
- System prompts
- Architectural experiment results
- The brand system itself
Ways to work with us.
Lamantin AI is a research organization, and working with us looks like research: a hard problem, an architecture that wants a second pair of eyes, a system that ought to be smarter than it is. We take on few of these at a time and stay with them — depth is the thing we actually have to offer. Direction stays ours throughout, which is what lets us say yes to the problems worth the attention and no to the rest.
Expertise
Consulting on a research question or a project already under way, and lectures, seminars and masterclasses in AI for the team that will have to live with the result. What changes hands here is knowledge, not work.
Building something
Where a problem is hard enough to be worth it, we build the thing rather than advise on it: an AI product, or the part of an existing one that has to think. A longer relationship than a review, and it carries the condition below.
Research together
Applied research with an organization that brings a real problem and the patience for it. A joint exploration between practitioners, carrying a direction rather than a specification. We take on very few at a time, because spreading thin is how depth is lost.
Support the work
The work can be supported through public platforms. An organization that does gets visibility into what is underway and a seat in the conversation about where it goes next — the acknowledgement of an aligned supporter, extended informally.
The core technology we build for you is opened to everyone afterwards — the code, the architecture, the findings. Results may stay private while they are being validated, which is an engineering rhythm rather than an exclusivity window. Work whose core has to stay closed permanently is work for someone else.