Aditya Pai.
Résumé
← All work

04 / AI Research / Scientific Infrastructure

Autonomous Scientific Discovery

Infrastructure for an AI-assisted laboratory

Project reference ↗
Role
Research Assistant · NC State
Period
2025 — Present
Built with
Python · AutoDock Vina · Machine Learning · AWS Lambda · SQL
An experimental loopCONCEPT STUDY
DATAPREDICTREVIEWEXPERIMENT
Data → prediction → review → experiment
Conceptual engineering diagram · Illustrative, not a product screenshot or measured plot.
labeled assay rows
3M+
in the loop
Human

The problem

Scientific discovery depends on moving between data, predictions, experiments, and human judgment. Automating one step is useful; connecting the steps into a repeatable workflow is the larger infrastructure problem.

At NC State, I built AI-assisted chemical discovery workflows combining AutoDock Vina docking with a binding-affinity model trained on more than three million labeled biochemical assay rows.

An experimental loop

The workflow brings together computational screening, model-based evaluation, and human-in-the-loop cycles. I also co-developed remote experiment orchestration and AI-driven co-scientist workflows with CMU’s self-driving cloud laboratory.

The diagram illustrates the relationship between stages. It does not claim a fully unattended laboratory or a validated scientific discovery outcome.

Modernizing the infrastructure

ILLUSTRATIVE ARCHITECTURESystem, at a glance.
  1. 01
    Scientific dataIngest and verify biochemical assay records
  2. 02
    Computational screeningAutoDock Vina and ML binding-affinity predictions
  3. 03
    Human reviewSelect and refine the next experiment
  4. 04
    Experiment & feedbackOrchestrate experiments and inform the next cycle

An illustrative human-in-the-loop workflow, not a claim of fully unattended discovery.

I migrated monolithic C and PHP services toward Python microservices deployed on AWS Lambda. This work separated legacy service responsibilities into independently managed components.

In parallel, I built an automated ingestion pipeline using Python, Selenium, and Pywin32 to update and verify more than three million SQL records.

Outcomes

  • Chemical discovery workflows combining docking and ML predictions.
  • A binding-affinity model trained on 3M+ labeled assay rows.
  • Legacy services migrated toward Python microservices on AWS.
  • Automated scientific-data ingestion and verification.
  • Contributions to remote experimentation and co-scientist orchestration.

Engineering takeaway

Automation is most useful when the data and handoffs are dependable. In a scientific workflow, human review remains a meaningful part of the system rather than an inconvenience to remove.