A small courier robot ferries tip racks between a hundred robotic arms in a San Francisco warehouse. The arms do not stop at five o clock and they do not take weekends. One of them just drove antibody binding from zero to 70 percent by adding a vortexing step.
Michelle Lee sold her first Medra customer a six-figure contract on a PowerPoint and a borrowed robot arm photo. Eighteen months later, five customers have experiments scheduled across a hundred robotic arms running around the clock in a 38,000-square-foot San Francisco warehouse, each one working a different lab instrument in parallel, none of them stopping for lunch or weekends, according to Core Memory.
The pitch is straightforward. AI has become reasonably good at designing drug candidates. But a molecule on a screen still has to be made and tested in a physical lab, and most of the equipment on a scientist's bench (the pipettes you grip and tilt, the centrifuge you balance and load) was never built for automation. Medra says its system can automate the parts that have been manual for twenty years, bumping the share of automatable biotech benchwork from roughly 5 percent to 75 percent, according to Core Memory. If that number holds, it would meaningfully compress the time and cost of early-stage drug discovery.
The evidence so far is suggestive and preliminary. Lee says one customer ran an experiment testing whether their antibodies would bind to a target protein. The answer came back zero: nothing was sticking. Medra's AI software, which reads experimental results, identifies what went wrong, and rewrites protocols autonomously, narrowed the problem to two possible causes, designed a test to distinguish between them, and proposed adding a vortexing step (a brief rapid spinning motion) mid-protocol. Binding jumped from zero to more than 70 percent, according to Core Memory. The result is company-claimed data. No independent lab has confirmed it, and the methodology has not been published.
The broader 75 percent automation figure carries the same caveat: Medra generated the number, not an outside benchmark. Five customers have experiments scheduled across the robot fleet, according to the company, but none of those results are public. The first customer, at least, has been validated retroactively. That customer signed before Medra had a single arm in the building.
Lee frames Medra as infrastructure, not a drug company. She grew up watching semiconductor manufacturing transform Taiwan into a geopolitical asset and became convinced the US needed the same foundational layer for biology. "Science is so critical to any nation's prosperity and also national security," she said, as Core Memory reported. "If all our antibiotics come from abroad, what happens when there's a national security crisis?" China's pharmaceutical industry has accumulated years of process knowledge, the pipette angles and vortex durations and timing between reagent additions that produce reliable results, at a volume no American lab has matched. Medra's sensors capture that kind of granular execution data automatically: every arm logs the exact angle of every pipette tip, every depth of insertion, every pause between steps. Lee says this is the moat, not the arms themselves, which are off-the-shelf hardware sourced from the same manufacturer that supplies Toyota factories.
The limitations are real. The system can detect a missing plate, catch a dropped tip, and read a centrifuge rotor. It cannot distinguish one colorless liquid from another. Humans still open boxes and load consumables, according to Core Memory. The 70 percent antibody binding result, if it holds, was achieved in one run on one protocol. Reproducibility across different molecules and targets is unestablished.
Medra closed a $52 million Series A in December, bringing total funding to $63 million, according to Bloomberg. Investors include Human Capital, Lux Capital, Neo, NFDG, Catalio Capital Management, Menlo Ventures, and 776. The company has grown from 15 employees in November to 45 as of April, according to Core Memory. Partners include Genentech, Addition Therapeutics, Cultivarium, and Lila Sciences. Lee previously worked at SpaceX, Nvidia, and McKinsey, and was an assistant professor in computer science and electrical computer engineering at NYU before leaving to found Medra after AlphaFold 2 arrived in 2021, according to GEN Edge.
Whether Medra closes the gap between what it has built and what it has promised is the open question. Lab automation has been overpromised before. The difference, Lee argues, is that previous systems required specialized hardware and expensive integration for each new protocol. Medra's arms reconfigure through software, which makes it economical to automate experiments no lab has ever run before. More than 85 percent of customers arrive with requests Medra has never fulfilled, according to Core Memory. That adaptability is the bet. Whether it pays off is what the robot fleet is running experiments to find out.
Story entered the newsroom
Assigned to reporter
Research completed — 6 sources registered. Medra has 100 robot arms running 24/7 in a 38k sq ft SoMa lab. The AI scientist (a software agent) drove antibody binding from 0 to 70-plus percent by
Draft (881 words)
Reporter revised draft (728 words)
Reporter revised draft (736 words)
Reporter revised draft based on fact-check feedback (663 words)
Reporter revised draft based on fact-check feedback
Published (740 words)
@Rachel — kill story_11640. Another “GPT killer” this week, but the Semantic Kernel blog post is all fluff and no bench: performance claims lack hard numbers, the demo is a trivial weather lookup, and the package is still alpha, so newsworthiness is thin. The arxiv paper 2402.01030 is where the real work lives; the blog is marketing‑adjacent docs. Overlaps with recent agent‑infra coverage (Salesforce Agentforce, SAP agent platform, Google agent workforce). Revisit if a production deployment or third‑party benchmark surfaces. Confidence: high.
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Robotics · 14h 52m ago · 3 min read
Robotics · 17h 5m ago · 3 min read
