Ex-H Company researchers raise $30m seed for Remanence to train custom models on agent data
Paris-based Remanence has raised a $30m seed round co-led by EQT Ventures and FirstMark to post-train open-weight models on the work its customers' AI agents already do. Datadog's Olivier Pomel and Synthesia's founders are among the angels.


Remanence, a Paris startup founded this autumn by three former researchers at H Company, has raised a $30m seed round, about €25.5m, co-led by EQT Ventures and FirstMark. The company wants to give businesses that run AI agents a way to turn those agents' own records into specialised models they own, rather than paying per call to a large US lab.
That is a big cheque for a company that was only registered as a French SAS in September. Yellow, Global Founders Capital, Zero Prime, Drysdale and Emblem joined the round. The list of angels reads like a roll call of French and European software founders: Datadog chief executive Olivier Pomel, Synthesia co-founders Victor Riparbelli and Steffen Tjerrild, Pigment co-founder Romain Niccoli, Mirakl's Philippe Corrot and Mehdi Ghissassi of AI71.
FirstMark knows the founders' previous employer well. The fund took part in the $220m round H Company announced when it launched in May 2024.
From Holo to company procedures
Pierre-Louis Cedoz, the chief executive, and co-founders Hamza Benchekroun and Axel Moyal led post-training work on H Company's Holo family of open models. That work focused on teaching agents to operate software: find the right element on a screen, pick an action, check what happened, then carry on. Cedoz and Benchekroun are among the authors of the technical paper on Holo3. The company's announcement says the three have also worked at Google DeepMind, InstaDeep and Meta. The Paris registration filing names Benchekroun as president and Moyal as managing director.
The new company applies the same idea to the routines of a single business. When a customer-support agent looks up an order, checks the return rules and offers a refund, it leaves a log of each step. Remanence reads those traces and builds sandboxed copies of the task, complete with verifiers that decide whether a run was correct. An open-weight model then practises in those environments, and long-horizon reinforcement learning rewards the runs that succeed and adjusts the model's weights.
The hard part is the verifier. A log can show that a refund went out, but judging whether it should have means knowing the policy, the amount the customer was owed and any exceptions. Remanence has to encode that logic for every procedure it trains on.
“We started Remanence to close that gap and help businesses turn their own agent data into specialised models they control,” said Pierre-Louis Cedoz, co-founder and CEO.
Customers get their own version of the model. They can pick where it is trained and where it runs, either fully managed by Remanence or inside their own cloud, so their data stays on their own infrastructure. Whether a customer can take the model away and keep training it depends partly on the licence of the open model it started from.
Where the $30m goes
The money will pay for a dedicated GPU cluster, more training environments and evaluation systems. Most hiring will be in Paris, and the company plans its first footholds in London and San Francisco. It has not disclosed headcount.
The startup says it is working with design partners in customer support, legal and enterprise software across Europe and the US. It has not named any of them, published production results or set out pricing. That matters, because the case for a custom model is about cost per task completed correctly, not cost per token. A smaller specialised model is cheaper to run. Whether it saves money overall depends on volume, the cost of training and retraining as procedures change, and how many cases still go to a general-purpose model or a person.
FirstMark expects a hybrid set-up, in which companies keep access to frontier models from the big labs and run their own models on selected workflows. The pitch fits a wider European interest in owning the AI stack, the same argument Mistral AI makes with its open models and its own compute.
A crowded field for post-training
Rivals already exist. Adaptive ML, whose Adaptive Engine handles evaluation, post-training and learning from production feedback, is the closest rival; it has said AT&T picked it to build models tuned to the carrier's tools and data. Veris AI covers part of the same chain from the other end, simulating a company's systems and turning production traces into test scenarios. Engineering teams can also assemble some of this in-house with open-source tools such as OpenPipe's ART training framework.
The founders' Holo experience gives Remanence technical credibility. The commercial test is different. Investors will watch how long it takes to go from a customer's traces to a first usable model, and how much of that work carries over to the next customer. If every deployment needs specialists to rebuild procedures and their exceptions by hand, Remanence looks more like a services firm. If its tooling lets a second, similar procedure go live faster, it has a platform.
The first documented deployment will be the real proof point: success rate, cost per successful task and the share of cases handed back to humans, measured on the same procedure before and after.
What did you think of this story?
Written by
John Pulisic
50 stories on European Startup Insider
Have a story like this? Pitch it to us.
We publish founder profiles, company news and practical guides from across Europe.
More stories like this
Remanence raises $30M seed to help companies own their AI models
Upload a landscape 16:9 photo of at least 1280 × 720, such as an official press photo of the founders. I couldn't find a confirmed press kit.







