A core team that has already shipped production-grade agentic AI together — with research, product, engineering, and evidence-provenance owners, plus a senior advisor in the literature. The open Stage-1 gap is software engineering.
Owns Hyper's scientific and technical architecture. Author of
flynt,
an open-source Python AST-rewriting tool (~700 GitHub stars). Led a
5-person team at the Bosch Center for AI building their research
RL framework (async data collection, distributed execution, CI
benchmarking; adopted across multiple Bosch products and by external PhD
students). Prior: Amazon Alexa compute-efficiency ($500k/yr saved on NLU
integration testing), Capgemini CV, NXP compiler. CTO / cofounder of
Covenance.ai. Based in Germany.
Owns commercial strategy, fundraising, partnerships, and the path to Series A. CEO / cofounder of Covenance.ai. 14+ years in Amazon's tech org (Prime Now → EU Promotions Tech → Product Insurance; PM tech lead), delivering systems at billion-dollar scale. Based in Luxembourg.
Owns Hyper's experimental rigour — PoC study design, evaluation robustness, and the statistical validity that makes a Stage-1 result credible to reviewers. Previously a Senior Scientist at Yahoo; combines mathematical depth with production ML experience. (Google Scholar.)
Owns product direction and bootstraps VC relations. Ex-Uber AI; exited founder.
Owns Stage-1 engineering. Ex-Amazon AGI.
Research & evidence bench. Soumajit and Alessandro add CS and ML research depth; Felipe owns the provenance of evidence through Mareforma.
Advises on AutoML, neural architecture search, and AI-Scientist systems — the nearest prior art to Hyper's loop.
Not a first collaboration. Ilya and Valerio cofounded Covenance.ai, incorporated the Luxembourg entity, shipped an AI-native product, were selected for the EU Regulatory AI Sandbox (EUSAiR) certification, and converted an Eversheds-Sutherland Italy design partnership to paid. Founder-chemistry, EU-entity-formation, and regulator-engagement risks are already retired for this pairing.
Research direction and experimental rigour are already owned (see above), so the Stage-1 hiring need is engineering. Leads are identified; freelancers are bridging in the interim. The filter is taste for the problem — familiarity with the nearest prior art (FunSearch, EoH, AlphaEvolve, Darwin Gödel Machine) helps but isn't a prerequisite.
Full-time from Stage-1 kickoff. You'd build and harden the pieces the discovery loop runs on.
Leads are identified; bridging freelancers in the interim: Edvard Grei, Jochen Wersdörfer.
Frontier-scale training and MLOps aren't needed in Stage 1. We hire this lead before the Stage-2 gate, gated by Stage-1 evidence, to take the loop from low-compute targets up to small-LLM scale.
€3M, 7 months, under the SPRIND Next Frontier AI Challenge. The headline deliverable: a Hyper-discovered method that beats an established, peer-reviewed baseline on a low-compute ML benchmark — statistically significant (p < 0.05, 3+ seeds), reproducible from the canonical library without human edits — as a preprint with full provenance, targeted at 1+ by month 5, alongside 5+ toy or small-scale breadth wins. The loop has to show four properties — proposal breadth, implementation fidelity, evaluation alignment, and evidence reuse — and a month-5 stop condition turns an honest miss into a rigorous report on the binding subsystem. Full plan available on request.
Interested in one of the open roles — or know someone who fits? Write to ilya@covenance.ai with two or three sentences on which design decision in the engine you'd most enjoy or most disagree with. We respond to everyone.