Confluence Labs
foundation models optimized for learning efficiency
Winter 2026ActiveB2BArtificial IntelligenceAISan Francisco, CA, USA
Company
https://confluence.shWhile modern AI excels in any area you can collect a lot of data for, it struggles in areas where data is sparse or costly to attain. Designing new molecules, discovering new physics, and engineering new materials, and even developing more effective systems of governance are all domains where collecting data is extremely costly.
We dream of a world where AI accelerates research in all of these domains and creates a more abundant future for humanity, but the current technology is not there. That’s why we started Confluence Labs.
We are building AI that can design highly effective experiments in data-sparse domains and learn maximally from the data it already has.
Verdict
High Signal
Market Opportunity
The target applications — drug design, hardware engineering, materials science, physics research — are multi-billion dollar markets where data efficiency is a genuine bottleneck. B2B enterprise customers in pharma, biotech, and semiconductors routinely spend heavily on R&D acceleration tools. The ICP (researchers gated by cost of physical experiments) is clear and well-articulated.
Medium Signal
Founder Signal
Niranjan Baskaran: Vassar CS + Dartmouth ECE (on leave for YC), RSI 2021 (one of 30 students selected worldwide, research under Prof. Charles Doran of Harvard CMSA), ISEF 2021+2022, Atlas Fellow. The team's reported 97.9% ARC-AGI-2 result is an extraordinary technical signal that is difficult to dismiss regardless of professional experience. Brent Burdick is a self-described college dropout and self-taught engineer with limited verifiable public profile. Neither founder has professional experience in the target domains (drug design, materials science, pharma R&D), but the ARC-AGI-2 result suggests genuine research capability.
Medium Signal
Competition
They're competing against frontier labs (Anthropic, OpenAI, Google DeepMind) on reasoning benchmarks, which is a David vs. Goliath situation. In the scientific AI application space, competitors include Recursion Pharmaceuticals, Insilico Medicine, Exscientia, and Benchling for specific verticals, plus general AI labs pivoting to scientific applications. Their differentiation — LLM-driven program synthesis for data-sparse domains — is technically interesting but not yet a proven moat since it's built on top of existing LLMs (GPT, Claude) rather than proprietary model weights.
Medium Signal
Product
The company claims SOTA on ARC-AGI-2 at 97.9% accuracy for $11.77/task, beating Claude Opus 4.6 Thinking and GPT 5.2 Thinking High on the public eval. They've open-sourced their solver, which is reproducible — this is a real technical result, not vaporware. However, there are no paying customers, no revenue, no API docs, no pricing page, and the actual product (hypothesis generation + data-efficient modeling for drug/hardware/physics) is described as aspirational and future-looking.
OverallC Tier
Confluence Labs has a legitimately impressive technical result — 97.9% on ARC-AGI-2 is a real benchmark, not marketing, and is difficult to dismiss. The gap between a benchmark result and a commercial product is significant, particularly in drug design and materials science where enterprise sales cycles are long and domain trust matters enormously. The team has no professional experience in the target verticals and no customers or revenue. Backed by YC and Paul Graham, which provides real runway and credibility. This is early-stage research with exceptional technical signal and an unproven commercial path — C tier reflects the real upside tempered by the distance to monetization.
Active Founders
Confluence Labs
TierC Tier
BatchWinter 2026
Team Size2
StatusActive
LocationSan Francisco, CA, USA
Last Updated4 days ago