Inspiration
Resist AI began with a desire to make computational research into antimicrobial resistance more accessible and easier to trust.
As we explored AI-assisted molecular research, we kept returning to a fundamental question: can a researcher clearly distinguish what was calculated, what an AI model interpreted, and what still needs experimental validation?
A convincing explanation can obscure those boundaries. We wanted to make them visible.
That became the focus of Resist AI: a research copilot that brings molecular calculations, AI commentary, and their provenance into one inspectable workflow, helping researchers understand both the evidence they have and the questions that remain.
What it does
Resist AI turns a molecular structure into an inspectable research record.
A user enters a SMILES string, a text representation of a molecule. The application calculates molecular descriptors with RDKit, including properties such as molecular weight and hydrogen-bond donor and acceptor counts.
The workflow lets users:
- Calculate molecular properties and inspect the underlying structure.
- Optionally request AI commentary through an integration configured for NVIDIA Nemotron on Nebius Token Factory.
- View calculations and generated interpretation separately, with model commentary clearly labeled as unreviewed.
- Inspect provenance, including the canonical structure, software version, input fingerprint, and available model-response metadata.
- Download a JSON record for further inspection and comparison.
Invalid inputs and unavailable model requests produce explicit failure states. Molecular descriptors remain identified as calculations; they do not establish biological activity, clinical safety, or treatment suitability.
How we built it
We extended the existing ResistAI / Team BioHack research prototype, concentrating the hackathon work on evidence review, the Nebius integration, and a focused application for demonstrating the workflow.
The application uses Python and FastAPI, with a lightweight HTML and JavaScript interface. RDKit performs molecular calculations on the server and supplies the evidence passed into the optional model-review step.
We built a configurable client for Nebius Token Factory’s OpenAI-compatible API, targeting NVIDIA Nemotron. The integration requests structured commentary tied to the supplied evidence, checks response format and evidence references, and records requested and returned model identifiers, timing, hashes, and available usage metadata.
We added authentication, bounded requests, and SQLite-backed limits on model attempts that persist across restarts. Docker packages the focused application, while automated checks cover input handling, evidence records, provider failures, and access controls.
Challenges we ran into
The hardest challenge was preserving the boundary between computation and interpretation throughout the application.
A valid molecular structure, a calculated descriptor, and a well-formed model response each establish something different. The interface and exported records needed to make those differences understandable.
We also had to narrow a broad research prototype into a workflow we could inspect and test thoroughly. That meant separating inherited experimental features from the focused demonstration and reviewing how results were labeled.
Operational reliability mattered too. Missing credentials, malformed responses, timeouts, and exhausted request limits all needed clear outcomes. We designed those conditions to remain visible in the research record.
Accomplishments that we're proud of
We’re proud of building a concrete workflow that connects molecular input, real calculations, optional AI interpretation, and downloadable provenance.
Our key accomplishments include:
- A focused research interface that makes the evidence behind an interpretation visible.
- An implemented Nebius/Nemotron integration with structured response handling and inspectable request metadata.
- Explicit uncertainty and failure handling throughout the review workflow.
- Authentication and durable usage controls for managing access to model requests.
- 365 passing automated tests, alongside lint checks and a container smoke test covering authenticated access and real descriptor calculations.
These are software-engineering milestones. Scientific evaluation is the next layer of work.
What we learned
We learned that trust in an AI research tool depends on its data model, interface, and failure behavior as much as its choice of model.
A response can pass every formatting check and still require scientific review. Provenance helps explain where a result came from, while experimental evidence determines what conclusions it can support.
We also learned the value of a focused scope. Building one inspectable workflow helped us identify assumptions, test boundaries, and explain the system more clearly.
Making uncertainty understandable is part of making research software useful.
What's next for Resist AI
Our immediate priorities are to complete live Nebius validation and hosted demonstration checks, then evaluate the quality and usefulness of the generated commentary.
From there, we plan to:
- Benchmark explanations against curated molecular reference cases.
- Work with qualified researchers to assess correctness, clarity, and usefulness.
- Add independently curated evidence and stronger links between inputs, calculations, and research claims.
- Extend the workflow to additional molecular and biological tools as their outputs can be traced and interpreted reliably.
Our long-term ambition is to support more reliable research into antimicrobial resistance, with transparent computational evidence and experimental validation guiding the platform’s development.
