The practice / 47 companies

Startups we work with

These are not the companies we merely cover. This is the roster the practice actually works on — read daily off their own changelogs, careers pages and status pages, documented, and written up. Each carries one line on why we track it. The list changes as we add and drop names.

tracked

47

hiring now

14

open roles

120

Counted from the roster below · as of 19 Sept 2026, 11:45 UTC, refreshed every 60s

Agenta

The agent harness as a product — an open-source workspace that versions an agent's instructions, skills and tools like code and evaluates a change before it ships, which is the layer the whole "build your own system" argument runs through.

since September 2026

AI Score

The governance and observability layer over enterprise inference — the system of record for what an organisation's AI and agents actually did.

since September 2026

Bluedot

A high-volume inference consumer paying retail — every meeting they capture is transcription plus summarisation billed per token, and their margin is the gap between a $14–$39 seat and what that seat costs them to serve.

since September 2026

Cactus Compute

The bottom of the inference market — a 14MB model and a dependency-free C++ runtime aimed at the four in five edge devices that cost under $200, where the unit economics are flash reads and milliwatt-hours rather than tokens.

since September 2026

Callosum

The routing layer across heterogeneous silicon — the software that decides which model runs on which chip, sitting between the chip makers and the teams buying inference.

13 open rolessince September 2026

Cerbos

The runtime permission layer for AI agents — the company answering "what is this agent actually allowed to do", which is the control almost every agent stack is currently missing.

since September 2026

Cognee

The memory layer of the agent stack — where a knowledge graph, not a bigger context window, is the bet on what an agent should remember, and where the open-source project and the enterprise deployment have to stay the same product.

since September 2026

Daytona

The sandbox layer underneath every coding agent — the place where "give the model its own computer" stops being a metaphor and becomes a per-second price.

since September 2026

Deeploy

The people betting that AI governance becomes a monitored, evidenced, continuously-audited thing rather than a document — and the clearest European read on what the EU AI Act actually asks a deployer to do.

since September 2026

Doubleword

Open-weight inference sold by latency tier, plus an open-source gateway teams self-host — the layer where a buy-versus-rent decision turns into a per-token bill.

since September 2026

E2B

The sandbox layer — whoever owns "where the agent's code actually runs" sits under every coding agent and every computer-use product, and E2B is the one selling it as a primitive rather than as part of a platform.

10 open rolessince September 2026

EXO Labs

The local-inference edge — the software layer that turns hardware a team already owns into the cluster, where the buy-versus-rent decision gets settled in Thunderbolt cables rather than API bills.

1 open rolesince September 2026

FlexAI

The clearest published attempt to make inference pricing checkable rather than trusted — they write down the rule that sets each rate and ship calculators that show where per-token billing stops being the cheap option.

since September 2026

Flower Labs

Federated learning *is* the direction's core question made into a product — what it costs to send your data somewhere else — and they are the open-source default for it, sitting on four separate documentation roots with no single door in.

since September 2026

Giskard

The European red-team lab — attacking agents for a living, publishing the benchmarks and the incident datasets in the open, and selling the same machinery back to banks and insurers as a compliance instrument.

since September 2026

Gradient Labs

The applied end of the agent stack — what it takes to run an autonomous agent inside a regulated bank, where the eval harness and the audit trail matter more than the model.

6 open rolessince September 2026

Hopsworks

The feature store that predates the current wave and is now re-aiming itself at agents — a useful control on whether "AI infrastructure" is a new stack or the old data stack under new lighting.

since September 2026

Implement AI

The applications end of the stack — they sell the outcome of inference to non-technical UK businesses, so their pricing is where token cost becomes a task price a buyer actually sees.

since September 2026

Inception Labs

A different inference architecture — diffusion rather than autoregression — which changes the latency and cost-per-task arithmetic for any product that calls a model many times inside a single user interaction.

since September 2026

Inngest

The durable-execution layer that agent builders reach for when "retry the whole thing" stops being an acceptable answer — and one of the few in that layer publishing its own reliability post-mortems.

1 open rolesince September 2026

LangWatch

The test-harness layer of the agent stack — the people trying to make "does this agent actually work" a number you can put in CI, priced per event in euros and open enough to run on your own metal.

since September 2026

Lightpanda

They are rebuilding the browser on the premise that nobody is looking at it — which is the cheapest, most measurable claim on the roster, and nobody has independently benchmarked it against headless Chrome.

since September 2026

LM Studio

The default way a normal person runs a model on their own hardware — and, as of this year, a local-first company that has started selling hosted inference, which is the direction's central trade-off happening inside one product.

since September 2026

Maitai

The "your production traffic is your training set" bet, priced at $50 a month — the clearest small-team attempt to make fine-tuned task models cheaper to run than the frontier model you are calling today.

since September 2026

Mastra

The most-downloaded TypeScript agent framework there is — 1,240,131 weekly installs of @mastra/core — which makes it the reference point for anything we publish about what building an agent in TypeScript actually costs.

2 open rolessince September 2026

MemryX

A US edge-AI accelerator you can actually buy off the shelf for the price of a mid-range GPU fan — the cheapest way to find out whether dedicated inference silicon belongs in a normal machine.

since September 2026

Metorial

They are building the plumbing that decides which tools an agent may touch and who gets to say so — the governance layer of the MCP stack — at a size where the founders still answer their own mail.

since September 2026

Mixedbread

The retrieval layer an agent actually runs on — they train the embedder, the reranker and the search agent together, which makes them the clearest public test of whether a co-designed stack beats an assembled one.

since September 2026

Nabu Casa

The only company shipping a privacy-first voice assistant people actually buy, which forces them to answer in public, feature by feature, what a house can run locally and what still has to leave it.

1 open rolesince September 2026

Northflank

The place an AI agent's code actually runs — microVM isolation, GPUs and your own VPC — which is where the agent-sandbox question stops being a model question and becomes an infrastructure one.

6 open rolessince September 2026

Ollama

The clearest live test of whether local inference can hold its ground — a 9-million-developer local runtime that has started selling cloud tokens, publishing the per-token price of both sides of that choice.

8 open rolessince September 2026

Prime Intellect

They are building the part of the stack most people cannot see — where a model is trained, given an environment to act in, and graded — and they publish the engineering behind it in the open, which makes them a primary source for us whether or not the relationship goes anywhere.

30 open rolessince September 2026

Pruna AI

Compression sold as a product rather than a feature — the clearest live test of whether "we make your model cheaper to run" is a business you can charge for, or something the serving layer eventually swallows.

since September 2026

Pydantic

The eval-and-observability layer of the agent stack, run by the people whose validation library is already inside almost every Python AI project — the clearest live test of whether "measure what your agent did" becomes a product line or a feature.

since September 2026

Restate

The durable-execution argument at its most technical — a Flink-lineage team insisting that an agent loop needs a log underneath it, not a checkpoint file, and publishing the reasoning rather than the marketing.

8 open rolessince September 2026

Sakana AI

They sell orchestration over open weights as a route to frontier scores at lower cost — the routing layer above the model, priced per token like a model, which is where the buy-versus-route decision actually gets settled.

since September 2026

Salus

Runtime authorization for AI agents — the layer that decides whether an agent's *next action* is allowed to touch a real system, which is the control problem every team hits the week after their agent starts writing.

since September 2026

Sauna (Wordware)

The clearest published pivot in the agent market — a $30M seed spent building developer infrastructure, then turned around in public to build the end product instead, with the whole argument written down and dated.

13 open rolessince September 2026

Trigger.dev

The durable-execution layer that agent builders reach for first because it is just TypeScript — the clearest test of whether "make the agent loop survive a restart" becomes a product category or a framework feature.

5 open rolessince September 2026

UbiOps

UbiOps runs and scales AI models on infrastructure the customer already owns — the orchestration layer a European organisation reaches once it has decided its models are not leaving its own hardware.

since September 2026

Valohai

A ten-year-old, ~29-person ML platform that runs on your own hardware and has never leaned on funding news — the closest thing on the roster to a real £500/mo buyer rather than a logo.

since September 2026

Vast.ai

Vast is the GPU spot-price layer MenFem's /data product reads — the hourly floor for a rented GPU-hour, which is the denominator under the task-price series we publish.

16 open rolessince September 2026

Velatir

The governance layer for AI a company did not choose — an EU-sovereign answer to shadow AI, sold to the people who have to defend the estate rather than to the people building on it.

since September 2026

Vespa.ai

The retrieval engine the RAG boom rediscovered — sixteen years of serving infrastructure now being asked whether it is also the right substrate for agents, which is a question worth watching someone answer in public.

since September 2026

XMOS

The British answer to "what runs the AI that has to live inside the device" — software-defined silicon with hard timing guarantees, sold into voice today and aimed at robotics next.

since September 2026

Zep

The memory layer argued as a database problem rather than a prompt trick — a temporal graph with validity windows, sold into enterprise VPCs, with the open-source engine under it in plain sight.

since September 2026

ZML

The clearest live test of whether inference is really portable — one server, one model, five accelerators, and published numbers for each, which is the only way to price a task instead of a token.

since September 2026

The roster is a subset of the research index. Every other company we cover — the dossiers, the coverage state, our conviction and trading relevance — is on the screener.

The full research index What the roster is hiring for