The Chemistry Copilot That Hires Zero Chemists

b12 Labs Emerges from YC with a Chemistry Copilot

b12 Labs left stealth this month as a Y Combinator Summer 2025 startup with an AI copilot that plans chemical syntheses, promising to pull early-stage pharma research from years to months. The startup joined Y Combinator's Summer 2025 cohort, founded in 2025 by Andres Bran and Zlatko Jončev, who listed three team members on YC's directory. Bran and Jončev met at EPFL's Lab of Artificial Chemical Intelligence. Bran built the first AI agent that autonomously thinks and makes molecules in a robotic lab; he published that result in Nature Machine Intelligence, where it drew over 1,000 citations and won best paper at the NeurIPS AI for Science conference. His broader work on agents and large language models for chemistry carries more than 1,100 citations. b12's blog says the startup exited stealth within YC to ship AI agents for chemistry workflows.

The tool targets a bottleneck in early drug discovery. b12's YC listing argues that chemical synthesis slows the process: a new drug needs more than ten years and over $2 billion to reach patients, and only two of every ten thousand compounds make it. Chemists' software fragments, dates, and resists integration. b12 calls its product a GPS for chemistry — it helps pharma plan recipes for new drugs and pulls early-stage discovery from years to months.

The platform stacks AI agents that pair planning with robotic execution in minutes, not months. b12's blog describes Automation Copilot, which translates natural language into machine-executable workflows for real instrument stacks, and Plate Designer, which builds optimized reaction layouts from a conversational goal. Synthesis Expert acts as an agentic retrosynthesis tool. Three specialist agents run inside: Selenium plans synthesis by judging strategies and building trees; Palladium designs and tunes conditions for each reaction; Cobalt compiles and runs instructions for robotic instruments. The system turns natural language into executable protocols without code.

Paid pilots with major pharmaceutical companies show early traction. b12's YC profile says the platform planned experiments that hit full conversion on the first try, eliminating the iteration cycle. The metrics page lists the gains:

The technical lineage runs through ChemCrow, an LLM chemistry agent described in Nature Machine Intelligence. The paper says the agent planned and ran syntheses of an insect repellent and three organocatalysts, and expert chemists rated it above GPT-4 on chemical factuality and reasoning. b12 built safety limits into the loop: if a proposed reaction turns dangerous, execution stops.

The company aims long term for autonomous labs with minimal human hands, but today pitches integration and operability. It works with early pharma and biotech partners to test the tools on real workflows, turning chemists into operators who direct machines by conversation.

The Founding Engineer Hire Signals a Hybrid Skill Bar

b12's careers page lists five engineering roles, none asking for chemistry or robotics background. The hiring push sets a bar for AI infrastructure builders. Though the copilot maps chemistry routes for pharma, the wanted hires write Rust and run vLLM.

The five postings break into narrow software functions. Senior Research Engineer — Memory Systems requires vector stores, graph pruning, and distributed indexes like FAISS or Tantivy, with Rust or strong Python. Staff Engineer — Council Router owns LLM routing and task classification, preferring Go or TypeScript. Senior Product Designer — BOSS Studios designs dense data surfaces. Senior Full-Stack Engineer — Studios ships front-end code on Next.js 15 and React 19. Site Reliability Engineer — Owned Compute runs a four-VM cluster serving whisper, Gemma, and Qwen models through llama-server. Every role targets model plumbing and web delivery. No posting mentions synthetic chemistry, lab protocols, or robotic arms.

This skill bar contrasts with the image of b12 as a robotics-first chemistry copilot. The job posts show no touch of lab robots. A neighboring physical-AI startup paints a different picture. Alcor Labs, a $5M seed firm, advertised its first engineering hire on JackandJill at $160k–$220k plus up to 2% equity, demanding applied AI for visual perception or agentic reasoning loops in industrial settings. That founder role owns edge-to-cloud pipelines from wearable hardware to cloud VLM inference, turning industrial labor into physical AI. Alcor wants hybrid code-plus-hardware plus six-day field deployment in San Francisco.

b12's product context explains the gap. The YC directory describes the startup as an AI copilot that plans how to create and test new molecules for chemists — a planning layer upstream of any bench work. A system that outputs synthesis routes need not command a pipette on day one. The founders, not the hires, hold the chemistry expertise, evidenced by Bran's published robotic-lab synthesis work. Early engineering money builds a fast, reliable planner.

The wider talent market tells a different story for physical AI. greatrobots.ai counted active roles in early 2026 across defense-robotics firms: Anduril 2,157, Shield AI 420, Saronic 270. Zero G Talent's board data shows Figure AI's rapid role additions, with pay bands in the table above. Zero G Talent's figures put Zipline's software engineering manager role at $190k–$275k, and Zipline's postings similarly cluster in software management. Robotics-AI crossover demand rings loud at delivery and defense firms, not at b12's chemistry copilot.

b12's early hiring signals an AI-software bar with chemistry left to the founders, while the defense hiring wave shows where hybrid robotics talent is flowing. Coders who write inference routers fit b12 today. Roboticists who want to touch molecules should watch Figure AI and Alcor Labs.

How the Copilot Trains Robots on Synthesis Recipes

The copilot's purpose is to make existing lab hardware act on a chemist's words. The agents described in the opening section fuse planning with execution, but the training of robots happens at the integration layer, where natural language becomes machine instructions.

b12's launch blog says the system "uses AI agents to convert natural language instructions into executable protocols" and orchestrates execution end to end. Physical automation stays vendor agnostic. b12's YC profile says it integrates with Chemspeed, Unchained, Opentrons, and any lab doing experiments manually. The site update on b12-labs.com adds that you can "Translate your optimized conditions into execution-ready protocols for Chemspeed, Opentrons, or any automation platform. No coding required." That interface trains the robot: the AI output is the recipe, and the robot consumes it directly.

Lab robots are expensive and powerful, but most of them sit underutilized because the barrier to operating them is too high. — b12 Labs launch blog, August 2025

The bottleneck b12 targets is early-stage chemical synthesis, where fragmented software and specialized programming keep expensive hardware idle. By dropping coding, the copilot turns a Chemspeed deck or Opentrons arm into a direct extension of the planner.

A May 2026 site update claims a new gain: byproduct assignment runs 75% faster than the manual norm. The pilot results from paid pharma work, cited in the first section, show full conversion on the first attempt, eliminating the iteration loop.

YC's S25 batch leans toward software agents over physical robots, a pattern visible in b12's own hiring. The defense hiring wave documented elsewhere sits outside this lab-automation wager.

Why YC Is Betting on Lab-Automation for Pharma

Y Combinator funded its first biotech, Ginkgo Bioworks, in 2014. The accelerator now backs more biotech startups each year than any other investor, it wrote on its biotech page. A live roster of 126 top biotech startups funded by July 2026 backs that claim. The portfolio size explains why the current S25 cohort's lab-automation tilt reads as strategy, not a one-off.

The accelerator's recent picks center on software agents that drive physical lab gear. b12 launched from YC to build AI agents that translate natural language into lab automation workflows, the company said. The startup, staffed by researchers from Roche, Chemspeed, Max Planck Institute and EPFL, designs, optimizes and runs chemical reactions with route planning and data analysis. Infera, another YC biotech, converts a plain-English experiment description into a validated, instrument-ready run across a lab's existing equipment. Enjamb builds a connection layer over legacy pharma systems like Benchling and Veeva so agents can run whole drug programs.

"We came out of stealth as part of YC to build AI agents that turn natural language into executable lab automation workflows." — b12 labs, August 2025

Pharma adopters answered within weeks. Enjamb reported that three weeks after launch, over 500 employees at Johnson & Johnson, AbbVie, Bristol Myers Squibb, Merck and Sanofi used its platform. Delineate, a separate YC startup writing AI agents for clinical trial design, works with two of the largest pharmaceutical companies. Tamarind serves tens of thousands of researchers in large pharma firms. Angstrom AI tells clients its simulations match wet-lab accuracy at over 100x speed, cutting preclinical checks.

YC's economic logic favors this layer. The accelerator says cost and cycle-time drops in life sciences now let small teams do work once locked in big labs, reminiscent of software's 1990s shift. Pharma's closed legacy stack — Benchling, Veeva, Medidata, SAS — lacks an AI connector, a gap these startups fill without ripping out instruments. b12 claims its copilot compresses early R&D from years to months.

Delineate sizes the clinical-trial slice at a $10 billion market where one saved day equals $1–5 million in revenue potential. Angstrom AI likewise achieves 100x faster wet-lab accuracy, cutting preclinical checks. Those figures pull founders to agent-first builds. YC reinforces the bet by skipping shared lab space; it arranges preferential deals with lab providers because founder needs vary too much, the accelerator said.

The defense hiring wave documented in earlier sections sits outside YC's biotech push, confirming a software-led lab-automation wager but stopping short of proving broad physical robotics appetite across the accelerator.

YC's network closes the gap for new entrants. The accelerator says the most valuable part for bio companies is the circle of other funded bio founders who trade help.

What This Is Not: Stepping Away from Genomic Bio-AI

b12 entered YC's Summer 2025 batch with a tight profile: an AI copilot for planning molecular synthesis. The YC directory says it helps pharma plan drug recipes, accelerating early discovery from years to months. That pitch sits close to the bio-AI boom, but it is not a genomic data play.

Genomic bio-AI platforms build value on massive biological datasets. b12's copilot outputs synthesis routes for small molecules; it does not train on genome sequences or protein structures.

The talent bar reflects the split. A genomic bio-AI shop staffs up with bioinformatics, computational biology, and data infrastructure engineers. b12's early hiring asks for AI infrastructure skills. The verified YC listing shows three people, and no public robotics deployment. While the startup is labeled robotics-adjacent, the evidence fails to confirm lab robots or accelerator-wide robot deployments, though the defense-tech recruitment surge is clear.

That surge is happening elsewhere. That surge is happening elsewhere. The same data shows:

Those three account for the bulk of defense-robotics hiring. Gecko Robotics, which fuses AI and robotics for critical infrastructure inspection, sits in the same tangible-hardware wave but outside pharma. The numbers underline where the headcount war actually rages: not in b12's three-person chemistry copilot, but in autonomous systems for defense.

The market narrative has shifted beyond pure software to embrace physical systems. That trend makes it easy to file b12 alongside every robotics label. But b12 plans chemical recipes; genomic bio-AI mines life's code; defense robotics builds machines that sense and act. Each pulls a different hybrid skill set.

For candidates, the fence is practical. A chemist who codes should track b12's copilot work. A data scientist fluent in gene sequencing belongs at genomic operations. The defense side, with Anduril's 2,157 open roles above, will absorb engineers leaving big tech for new missions, a separate flood from the pharma bench.

b12's copilot does not yet tie to lab arms; its recorded scope is molecular planning software. The genomics giants gather terabytes; b12 hands a chemist a route map. One fills databases, the other fills reaction vessels.

Working in robotics? Zero G Talent tracks the openings: see every open Zipline role, browse robotics jobs, openings at Nuro and Figure AI, and the people building the field.