Autonomous agents for scientific discovery

Research agents that run the experiment, not just describe it.

Empirion deploys autonomous agents that survey the literature, form hypotheses, run analyses and simulations on your datasets, and return reproducible, cited reports — built by physicists, trained with reinforcement learning.

// Built by physicists. Trained with RL.

fig.1 — hypothesis → experiment → result
Hypothesis H₀ vs H₁ Experiment sandboxed run Result cited + repro
fig.2 — fitted regression · R²=0.94
dose (mg) response

Trusted by research & data teams at

  • Helix Bio
  • Lattice Materials
  • Northwind Labs
  • Argo Genomics
  • Meridian Data

Research is bottlenecked by people, not ideas

The questions are everywhere. The expert hours to chase them down are not. That gap is where good hypotheses go to die.

01

Literature review takes weeks

A thorough survey across thousands of papers, preprints, and internal reports can eat a researcher's entire month — before a single experiment begins.

02

Analyses aren't reproducible

Results live in unversioned notebooks with undocumented seeds and lost environments. Six months later, no one can rerun the figure that mattered.

03

Expert time is scarce

Your best scientists spend their days wrangling data and rerunning pipelines instead of asking the next sharp question.

The platform

A full research loop, run by agents

From the first literature query to a signed, reproducible report — every step instrumented, every claim sourced.

Literature review with citations

Surveys open repositories and your private corpus, then synthesizes findings with inline reference markers that resolve to real sources.

Hypothesis generation

Proposes testable hypotheses ranked by expected information gain, with the reasoning and evidence behind each one laid out for review.

Sandboxed analysis & simulation

Writes and executes real code in isolated, network-restricted environments — statistical analyses, numerical simulations, model fits.

RL-trained reliable reasoning

Fine-tuned with reinforcement learning on verifiable tasks, so multi-step plans actually complete and hold up under scrutiny.

Reproducible runs

Captures code, data versions, package pins, and seeds. Every report carries a content hash so anyone can reproduce it bit for bit.

Connects to your data lake & HPC

Reads from your warehouses and object storage; dispatches heavy compute to your Slurm or Kubernetes HPC and GPU clusters.

Exportable papers & notebooks

One click to a clean manuscript, a runnable notebook, or a LaTeX draft — figures, methods, and bibliography included.

Human-in-the-loop review

Gate any stage on expert approval. Reviewers can comment, rerun, or fork a step — Empirion augments judgment, it doesn't replace it.

How it works

Four steps from question to cited result

  1. 1

    Pose a research question

    Ask in plain language, point Empirion at the relevant datasets, and set any constraints or success criteria.

  2. 2

    Agent surveys literature & forms hypotheses

    It reads the relevant work, synthesizes what's known, and proposes ranked, testable hypotheses with their evidence.

  3. 3

    Runs analyses in sandboxed compute

    It writes code, executes it on your HPC or GPU cluster inside an isolated sandbox, and iterates on the results.

  4. 4

    Returns a reproducible, cited report

    You get a written-up result with references, figures, the full run log, and a hash to reproduce it exactly.

empirion run · session 0xA7F3

prompt> Does compound KX-204 dose correlate with cell viability in the screen_2026q1 dataset? Survey prior work and test it.

lit Retrieved 1,284 papers · synthesized 18 relevant findings

hyp H₁: viability decreases monotonically with dose (sigmoidal)

run fit_dose_response.py · scipy 1.13 · seed=42 · cluster=gpu-a100

Result: significant negative dose–response, IC₅₀ = 4.2 µM [1][2] · consistent with prior screens [3].

repro hash sha256:9f3c…b71e · reproducible ✓

Research notes

Trained to do science, not to talk about it

Empirion's agents are post-trained with reinforcement learning on verifiable research tasks. We design rewards around outcomes a scientist would actually check.

// method notes

R = w₁·correctness + w₂·reproducibility + w₃·citation_validity

Illustrative. The agent earns reward when an analysis runs end to end, its result matches held-out ground truth, and every citation resolves to a real, relevant source. Hallucinated references and non-reproducible runs are penalized.

  • correctness — match against held-out outcomes and known benchmarks.
  • reproducibility — identical results on rerun from the captured environment + seed.
  • citation_validity — references resolve and support the claim they're attached to.

Use cases

Where teams point Empirion

Drug discovery & bio data

Mine screening data, fit dose–response curves, triage targets, and ground every finding in the assay results and prior literature.

Materials & physics simulation

Run parameter sweeps and numerical simulations on the cluster, then compare against published measurements and theory.

Data-science teams

Turn open questions into reproducible analyses — feature studies, causal checks, model evaluations — without burning analyst weeks.

Academic labs

Accelerate literature reviews and pilot analyses, with exportable manuscripts and notebooks ready for the group meeting.

0

papers synthesized per run

0

reproducible runs, by design

0

faster lit-review-to-result

0

analysis tools in the sandbox

// figures shown are illustrative

Pricing

Start small. Scale to the cluster.

Every plan ships reproducible runs and cited reports. Annual billing shown.

Lab

$0/ free

For individual researchers and small labs getting started.

  • Up to 50 runs / month
  • Literature review & citations
  • Sandboxed analysis (shared compute)
  • Reproducible run hashes
  • Notebook export
Contact us

Enterprise

Contact

On-prem HPC, custom RL training, and full data control.

  • Deploy in your VPC or on-prem HPC
  • Runs on your Slurm / K8s GPU cluster
  • Custom RL fine-tuning on your tasks
  • SSO, audit logs & access controls
  • Priority support & SLA
Contact us

Voices

What researchers say

“It compressed a literature review that would have taken my postdoc three weeks into an afternoon — and every claim was sourced. The reproducibility hash sold the rest of the lab.”
Dr. Anya Reyes
Principal Investigator, Computational Biology
“We pointed it at our data lake and it ran real analyses on our cluster, not toy examples. The fact that it's RL-trained on correctness shows — it finishes the multi-step work.”
Jordan Liu
Head of Data Science, Meridian Data
“Running on-prem against our HPC was the dealbreaker for everyone else. Empirion deployed into our environment and our data never left it.”
Dr. Sofia Marchetti
Director of Research Computing

About

Built by physicists

We think agents should be trained to do science, and that the compute under them should be real.

Kiran Kumar Das

Founder & CEO

Physicist and high-performance-computing engineer with roots in CERN-scale data and quantum computing. He builds both the science-trained agents and the sandboxed HPC/GPU compute they run on. His conviction: agents should be trained to do science, not just describe it.

FAQ

Questions, answered

What does it mean that Empirion is reproducible?

Every run records the agent's plan, the exact code it executed, the data versions and queries it touched, package versions, and the random seeds used. Each report ships with a content hash so a colleague can re-run the analysis and get the identical figures and numbers.

Is my data secure, and can Empirion run on-prem?

Yes. Analyses run in isolated, network-restricted sandboxes, and on Enterprise your data never leaves your environment. Empirion deploys into your VPC or directly onto your HPC/GPU cluster, with SSO, audit logs, and per-dataset access controls.

What does “trained with RL” actually mean here?

Empirion's agents are fine-tuned with reinforcement learning on verifiable research tasks. The reward rewards correct, reproducible outcomes — analyses that run end to end, results that match held-out ground truth, and citations that resolve to real sources — so the model learns reliable multi-step reasoning, not just fluent prose.

Does Empirion make up citations?

Citations are grounded in the documents Empirion retrieved during its literature survey. Every reference marker links to a source in the run's bibliography, and claims without supporting evidence are flagged for human review rather than asserted.

What data sources and compute can Empirion connect to?

Empirion connects to your data lake, warehouses, and object storage, and dispatches compute to your Slurm or Kubernetes HPC and GPU clusters. Literature connectors cover open repositories and your private document stores.

Can a human stay in the loop?

Always. You can require approval at any stage — hypothesis selection, analysis plan, or final report — and reviewers can comment on, rerun, or fork any step. Empirion is designed to augment expert researchers, not replace their judgment.

Put a research agent on your hardest question

Request access and we'll connect Empirion to your data and compute. Reproducible, cited results — start to finish.

// operations@empirion.tech · deploys in your VPC or on-prem HPC