Inephany Steers EIC Accelerator Money to Self-Tuning AI Training

Before a large AI model is trained for real, its developers usually train it many times over with different settings, such as the learning rate, and keep the run that works best. Inephany, a London company incorporated in July 2024, says that search wastes most of the compute it consumes. Its answer, called Paramorph, is a reinforcement learning system that adjusts those settings for each layer of the network while one training run is in progress, instead of fixing them in advance.

The EIC Accelerator selected the project at its 6 May 2026 cut-off as a grant-only award. CORDIS records an EU contribution of €2,441,694.94 for Paramorph from 1 August 2026 to 31 July 2028, with Inephany Ltd as coordinator. Eight weeks after the project started, Companies House records that Inephany Ltd changed its name to Recurvia Ltd on 26 September 2026, and the old inephany.com address now forwards to recurvia.ai.

Tuning a model while it trains

The settings that govern how a neural network learns, known as hyperparameters, are normally chosen before training starts and held fixed, or changed on a schedule written by hand. Finding good values means launching many runs. The CORDIS project summary describes this as a process that "consumes substantial GPU resources" and "may yield suboptimal models". Paramorph instead uses what the project calls a proprietary multi-agent reinforcement learning framework: software agents watch the training as it runs and change key hyperparameters at the level of individual layers.

The project states that no deployed solution provides real-time, layer-level, learned hyperparameter control within one training run. The idea itself is not new in research. A 2023 paper presented at a NeurIPS workshop, Generalisable Agents for Neural Network Optimisation, used one reinforcement learning agent per layer to control layer-wise learning rates and reported schedules competitive with hand-made ones. That paper is by other authors, and the project's claim is about deployed products rather than research.

The company's own site now describes Paramorph as a foundation model trained with multi-agent reinforcement learning to optimise the training of other models, working within limits set by human operators. Recurvia also lists a second product, Metrana, a metric logging and observability tool it says it built to develop Paramorph and offers to "select lab partners".

Three speed-up figures from one company

The numbers in circulation all come from the company or its application. At the pre-seed round in April 2025, Amadeus Capital Partners announced that the platform "could be at least ten times more cost-effective than current approaches", a claim Tech.eu repeated. The CORDIS summary gives two narrower figures: up to 3.5 times faster convergence than widely used optimisers such as Adam "under standard settings", validated on the transformer models GPT-2 and OLMo, and efficiency gains of up to 5 times in large-scale training, which it says could cut carbon emissions by several thousand tonnes per training cycle.

None of the three has been published in a form others can check. A search of arXiv and the web on 4 November 2026 found no paper or benchmark by the company, and its website gives no speed-up figure at all. CORDIS says the technology is supported by an EU patent application; a Google Patents search by applicant for both Inephany and Recurvia listed no published application on the same date. Patent applications are normally published 18 months after filing, so the absence of a publication is not evidence that no application exists.

Two years, one round, a new name

Inephany announced a pre-seed round of $2.2 million on 16 April 2025, led by Amadeus Capital Partners with Sure Valley Ventures and Steve Young, per Amadeus and the stock exchange notice of Sure Valley's fund, which does not disclose its own share. UKTN reported the round in sterling and gave two figures for it, £1.8 million in its headline and £1.7 million in its summary. The company said the money would grow its engineering team and bring in its first enterprise customers, and that its first products would follow later in 2025.

The filings at Companies House track the company's changes since then. Amelia Armour of Amadeus and Steve Young joined the board on 7 March 2025, the date of the round's share allotment. Co-founders Hami Bahraynian and Maurice von Sturm ceased to be persons with significant control on the same date and resigned as directors on 19 September 2025; the filings do not give a reason, and whether they still work for the company is not public. Two further directors were appointed on 20 January 2026. On 21 September 2026 the shareholders resolved to change the name, create a new share class and allot shares, and two allotments are recorded for that day. No funding round from 2026 has been announced.

The first accounts

The company's first accounts, for the 13 months from 5 July 2024 to 31 July 2025, were filed on 10 March 2026 under the micro-entity regime, which does not require a profit and loss account. They show net assets of £1,378,219, almost all of it current assets of £1,376,901, against £20,818 owed within a year, and an average of five people employed, directors included, per the filed balance sheet. Accounts for the year to 31 July 2026, which would cover the start of the EIC project, are due by 30 April 2027.

What the grant is meant to buy

The project's stated aim is to take Paramorph from TRL 5 to TRL 8 through large-scale experiments and structured pilots with companies, which the summary describes as de-risking the technology and unlocking private investment. It names its target users as AI research organisations, foundation model developers, quantitative finance firms, biotechnology companies and other businesses with heavy AI workloads, and lists a proprietary dataset covering more than 2,800 model configurations among its assets, per CORDIS.

The award is grant only, which for a British company is the only form available. The UK's association to Horizon Europe excludes the EIC Fund, the part of the EIC Accelerator that takes equity, according to the UK government's explainer of the agreement and the University of Cambridge's research funding guidance.

1. Working training optimiserTRL 5Validated on GPT-2 and OLMo per CORDIS; no public benchmark.
2. 3.5x faster convergence than AdamClaimStated in the project summary; no paper or third-party test found.
3. Large-scale experiments to TRL 8Under wayProject runs to 31 July 2028; no report filed yet.
4. Enterprise pilotsNot namedRecurvia refers only to "select lab partners".

The European compute question

Paramorph's case rests on the cost of training. The Commission's AI Continent Action Plan, presented on 9 April 2025, plans up to five AI gigafactories for training foundation models and an InvestAI facility meant to mobilise €20 billion, per InfoQ and Clifford Chance. Software that gets more training out of the same chips addresses the same constraint from the other side, and the project summary frames the grant as strengthening Europe's position in what it calls sustainable, competitive AI infrastructure. Whether the method delivers that at the scale of a foundation model is what the two years of the grant are meant to show.

The reporting is based on publicly available sources, listed below.