A Berlin-based legal AI research lab founded by test-time training pioneer Moritz Hardt has closed a pre-seed funding round led by Point Nine, with backing from a roster of AI heavyweights including Discovery Loop co-founder and CEO Jeff Dean, Stanford professor Chris Ré, and UC Berkeley professor Ion Stoica, the company said. Grubel declined to disclose the amount.
The company's unusual pitch: rather than deploying general-purpose AI agents across legal work, let each legal matter train its own specialized model through what Grubel calls a "specialization loop" that curates data, adapts models, and continuously evaluates performance on the case at hand.
"Our vision is a future where every matter builds its own AI, one that knows its world and does the best possible work within it," the company wrote on its website.
Why General Models Miss the Mark
Grubel's founding thesis rests on a specific technical critique. In a research note published earlier, the startup argued that off-the-shelf agent systems stumble on high-stakes legal tasks because they produce what it terms "silent failures"—errors that go unnoticed until too late. The problem compounds when those systems try to navigate the messy, unstructured document troves typical of legal work, or when they overlook facts explicitly required by a legal rubric.
The solution Grubel proposes is automation built around the matter itself. Instead of generic workflows, the startup's platform purportedly handles data curation, model fine-tuning, and ongoing evaluation in a closed loop trained directly on the documents and requirements of a given case. Whether that approach scales beyond proof-of-concept remains to be seen.
Early Numbers, Caveats Included
To back its claims, Grubel published experimental results using DeepSeek models on a suite of legal tasks. The company said its task-specific setup achieved roughly 63 percent accuracy across 11 tasks drawn from Harvey's Legal Agent Benchmark, a collection of synthetic due diligence scenarios. On two highlighted diligence tasks, Grubel's system scored 66.24 percent at an estimated cost of around $12 per task in model inference—performance the company said bested what it characterized as "frontier model" baselines applied to the same rubric.
The benchmark itself comes from Harvey, the legal AI platform that raised a $550 million funding round at a $15.5 billion valuation in September 2026. Grubel's decision to test against Harvey's own benchmark while counting Harvey's co-founders as investors adds a layer of industry entanglement worth noting.
"A simple task-specific system beats frontier models," Grubel wrote, suggesting that naive coding-agent patterns borrowed from software engineering fail to capture the rubric structure and factual precision legal work demands.
Pedigree and Practicalities
Hardt brings considerable academic credibility. He directs the Max Planck Institute for Intelligent Systems and co-authored a widely cited 2020 paper on test-time training, a technique that allows models to adapt during inference rather than relying solely on pre-training. Co-founder Reinhard Heckel, a professor at the Technical University of Munich, contributed to projects including OpenThoughts and DataComp.
The technical team includes Vedant Nanda, who previously worked at Mistral and Aleph Alpha on models like Mistral Large and Lawma, and Navid Rekabsaz, a former lead AI scientist at Thomson Reuters. Grubel incorporated as a German GmbH in Berlin with a Delaware parent entity and filed an EU trademark (application no. 019361839, filed May 9, 2026) covering AI-as-a-service categories, according to company disclosures.
The Data Custody Pitch
Grubel's go-to-market strategy leans heavily on data control. The company promises local deployments where client files never leave a customer's own infrastructure and where insights gleaned from one matter don't bleed into another. "You own the data, you own the intelligence," the company wrote—a message aimed squarely at legal-tech vendors, law firms, and corporate legal departments wary of cloud-based AI that might compromise confidentiality.
The startup has not disclosed customers, revenue, or headcount. Job postings on its site indicate open roles in legal engineering, research engineering, and operations. For now, Grubel remains in the early-stage category where technical demonstrations matter more than market traction. Whether its matter-specific training loops prove viable at scale, or practical enough for lawyers under deadline pressure, will become clearer as the company moves from benchmarks to billable work.
