Skip to content

Jingyi Li Scientific AI LLCLife sciences and materials sciencePennsylvania, United States

I build AI systems for science,and the evidence that they work.

An independent practice for teams whose product is a scientific result. I take the question, decide how the science should be represented, build the system that learns from it, and put in place the checks that tell you whether the output is real.

Founded and run by Jingyi Li, Ph.D. in Computational BiophysicsPenn State, 2026

Start a conversation What I build

How the work runs01

Frame the questionwhat a useful answer would look like
Choose the representationgeometry, symmetry, constraints, units
Build the systemdata pipeline, model, training, interface
Verify before anyone believes itphysics, simulation, baselines, failure maps

What fails verification returns to step two, not to a nicer chart

01 / THE PRACTICE

Three kinds of system, one way of working.

My training is in mechanics, simulation and computational biophysics, and my method is not tied to a single molecule, material or model family. What follows is what I build. The domain is yours to bring.

01 / Modeling

Scientific representation and generative modeling

Most scientific machine learning is decided before the first training run, in the choice of what the model is allowed to see. I encode coordinates, interactions, symmetries and physical constraints first, then pick an architecture that respects them.

  • Equivariant graph and geometric networks
  • Diffusion models and conditional generation
  • Surrogates for expensive simulation
  • Property prediction and inverse design
  • Benchmarking against domain baselines

02 / Infrastructure

Scientific data infrastructure and search

Research value is often locked inside archives nobody can query. I build the pipelines that turn experimental records, structures and simulation output into structured assets, and the search layer for cases where an ordinary database is not enough.

  • Ingestion, curation and quality auditing
  • Structure, graph and similarity search
  • Internal libraries and indexes
  • Simulation and experiment data at volume
  • Reproducible pipelines on computing clusters

03 / Automation

Language model agents and research automation

Large language models earn their place in research when they are wired into a workflow that checks them, not into a chat window. I build agents and pipelines that read at scale, extract at volume with an audit step, and return results a scientist can trace back to the source.

  • Literature and landscape intelligence
  • Extraction and classification with auditing
  • Evaluation harnesses for model output
  • Coding agents inside a research codebase
  • Internal tools and reviewable interfaces

02 / VERIFICATION

The part most teams skip.

A generative model always produces something. The real question is whether what it produced survives inspection by the field it came from. Every system I deliver comes with the means to answer that question, because a result nobody can check is not a result.

APhysical and geometric checks

Bond geometry, steric feasibility, conservation, energetics, units and boundary behavior. The output has to obey the science before anyone looks at the metric.

BSimulation as an independent witness

Molecular dynamics and finite element modeling as a second opinion on what the network claims. My simulation background means this is a working tool for me, not a subcontracted step.

CBaselines and failure maps

Comparison against what the field already does, and an honest account of where the model breaks. One aggregate number is how a project gets into trouble later.

03 / EVIDENCE

Systems I designed and shipped.

Two studies where I was first author, each starting from a representation problem and ending in something a scientist can use. They are examples of the method, not the limits of it.

ChironRNAFirst authorbioRxiv 2026Generative modeling

Repairing broken structures by generating the fix

A hierarchical diffusion system with E(3)-equivariant graph neural networks.

Damaged regions of an RNA structure are located by automated geometric diagnostics, then regenerated by a conditional diffusion model while the rest of the structure and its valid geometry are held in place. The same shape of problem appears anywhere a structure has to be corrected rather than built from nothing.

  • 80% clash reduction on more than 80% of the test set
  • All atom and coarse grained diffusion
  • PyTorch, equivariant networks, MolProbity
Read the preprint See it explained visually
ATLASFirst authorbioRxiv 2026Data infrastructure

Turning a public archive into a searchable asset

An automated pipeline from experimental structures to a queryable graph library.

Thousands of deposited structures were converted into atomic coordinates and graph representations, then made searchable by graph isomorphism so that a substructure can be found rather than remembered. The result exposed structural diversity that the previous categories had flattened.

  • 8,791 structures processed
  • 433,996 motifs across five categories
  • 1,487 distinct tetraloop conformations revealed
Read the preprint Browse the source

The domain changes. The method does not.

The same way of working has already moved across four fields that share no vocabulary, which is the reason I do not treat a new domain as a reason to start over.

Mechanics and simulationPlant cell wall mechanics, force field development, molecular dynamics at scale
Learned physicsNeural constitutive models for brain tissue inside finite element solvers, 95% accuracy at 80% lower cost
Structural biologyEquivariant generative models and graph libraries for RNA structure
Molecules and materialsMolecular design in flexible protein pockets, generative models for mechanical metamaterials

04 / WORKING TOGETHER

Three ways to start.

Most work begins small, because the first useful thing I can do for a team is usually to say precisely what the problem is.

Diagnostic review

A short engagement. I read your problem, your data and your current approach, then return a written assessment: what is feasible, what representation the problem actually needs, what it will take to verify, and what I would do first.

Build the system

I design and deliver the working system, including the data pipeline, the model, the evaluation, and documentation written so your team can keep running it after I step away.

Standing advisor

Recurring time for teams that have their own scientists and want a second opinion on method, a review of results before they are acted on, or help judging technical claims from outside the company.

Scope, timeline and terms are set per project. Tell me the problem in a paragraph and I will tell you whether I am the right person for it, and what I would do about it if I am.

05 / WHO THIS IS FOR

Teams whose result has to hold up.

The common thread is not an industry. It is that someone eventually has to defend the output to a reviewer, a regulator, an investor, or an experiment.

Biotechnology and pharmaceutical research

Structure, molecular design, target work and internal platforms that need to be more than a demonstration.

Materials, chemicals and engineering

Property prediction, inverse design, and machine learning that stands in for simulation without discarding the physics.

AI teams entering a scientific domain

Strong engineering, unfamiliar science. I supply the representation and the validation that domain reviewers will ask for.

Investors and diligence teams

An independent read on whether a scientific AI claim is supported by the evidence behind it.

Academic and institutional labs

Funded computational work that needs a system built properly once instead of rebuilt by each new student.

Teams building research automation

Literature intelligence, extraction pipelines and agents that have to be auditable before anyone trusts them.

06 / IN PUBLIC

How I think, written down.

I publish the reading and the tooling behind this work. It is the fastest way to judge whether my way of thinking fits yours, before either of us spends money on the question.

Literature atlasComputational structural biology

A field map built with an automated collection and classification pipeline, then audited rather than trusted.

Open the atlas →Literature atlasMachine learning in mechanics

The same treatment applied to a field with entirely different objects, which is the transfer argument made concrete.

Open the atlas →Reading deskPapers rebuilt as explanations

Interactive readings of research I care about, written so a non specialist can follow the argument.

Open the desk →

07 / CONTACT

Send me the hard version.

A paragraph about the problem is enough to start. If it is outside what I can do well, I will say so in the first reply and point you somewhere better.

jingyiliai4bio@gmail.com
LinkedIn Resume Personal site