AI Infrastructure for life-science agents

Today, we introduce Rafflesia. We are building the infrastructure for life-science research agents.

It began with a simple question: given the right tools, can an agent reproduce the conclusions of a research paper from its raw data?

The answer was no. Not because the agent was not capable enough — because the tools it needed did not exist.

Accelerating the feedback loop

Agents are extraordinarily capable, and they inherit a world that was never built for them. Perform a sequence homology search. Download a dataset. Run a molecular dynamics simulation. Visualize a result. Each of these should be one step inside a larger scientific question. Instead, every task becomes a meta-quest of its own — locating the tool, installing it, coaxing it into a machine-readable answer, and discovering hours later that the output was a PDF.

What matters is not any single one of those tasks. It is the loop they sit inside: the hypothesis-to-theory-to-experiment cycle that actually creates knowledge. Every meta-quest is time taken out of that loop.

The perfect toolbox will not be a stitching-together of the services humans already use. Wrapping a web UI in an API does not make it agent-native; it makes it a slower web UI. The toolbox will be new, and built for agents from the ground up:

• Headless interfaces. No human in the loop, by design rather than by workaround.
• Retrieval that answers without bloating the context window. An agent that must read a gigabyte to learn one number has already lost.
• Curated databases that live in the cloud and can be queried blazingly fast.

And wet lab can be part of these tools, not a separate world the software politely hands off to.

Our own iteration loop is deliberately simple:

create evals        real scientific tasks that ask agents to
                    reproduce results from the literature
      |
      v
watch failures      agent breakdowns show what is missing
                    from the toolbox
      |
      v
build infrastructure  missing primitives become durable systems,
                      with the evals as the proving ground
      |
      v
raise the bar       harder evals, and around again

The failures are the specification. We do not guess at which primitives matter; we watch where agents break and build exactly that.

Why structural biology

Structural biology studies the 3D structure, interactions, and dynamics of proteins and the complexes they form. That foundational understanding drives the design and discovery of better drugs.

We are starting here because the field has four characteristics that are rare in combination.

It produces experimental results at exponential scale. The storage and processing systems around that data were designed for local HPC centres tied to individual research labs. The infrastructure to index, curate, organize, and serve it needs to be rebuilt for the cloud and for AI agents — not ported to them.

It is one of the most verifiable environments in all of science. The feedback loops are everywhere you look: stoichiometry, hydrophobic and ionic forces, primary, secondary and tertiary structure, bond geometry, angstrom-level positions you can measure directly. These systems are also easy to simulate. The laws of physics provide the reinforcement-learning environment, and they do not need to be annotated by hand.

The problems span the full range of difficulty. Some are simple enough for an agent to handle today. Others sit at the edge of what anyone understands — large multi-protein complexes, for instance. That spread is what makes the field ideal for agentic hill-climbing, where each solved problem opens the next one rather than exhausting the supply.

The payoff is enormous. If we can make AI genuinely good at protein engineering and nanoscale design, the ceiling is hard to see. Imagine an AI that could design something like the nuclear pore complex, or a ribosome.

Our principles

Open. Our goal is to move biology out of the walled garden. We contribute to open source wherever we can.

Ship progressively. We believe moonshots are reached part by part, iteratively — not through inspirational, fuzzily-defined goals. Build the piece, ship the piece, then build the next one. We also believe that is how safe AI systems for biology get built: each piece is small enough to actually verify.

Software first. The data already exists, and excellent wet labs already generate more of it. Our job is the software layer around that decentralized network — standardized procedures, reproducible datasets, and tight loops that let better algorithms learn from real evidence.

Cloud native. Building for the cloud is building for AIs. It is also building for scale. Much of today's scientific computing was designed for HPC, and we think there is a huge opportunity to reinvent it for the cloud era.

What we have launched

• Rafflesia Homology — a serverless search engine built on object storage, as precise as Foldseek and MMseqs2, and much cheaper and more scalable.
• Headless Mol* and headless MDsrv — command-line tools that let agents create molecular visualizations without a human in the loop.
• TinkerBench — a benchmark for measuring how well agents perform on real structural biology tasks.

Each one came out of the loop above. TinkerBench exists because we needed to know where agents actually fail. The headless visualization tools exist because they failed at visualization. Rafflesia Homology exists because homology search was the primitive they reached for most and the one the existing infrastructure served worst.

We are assembling a world-class team of engineers and scientists to build groundbreaking infrastructure for life science. Join us.