The Gemma 4 Good Hackathon

The Gemma 4 Good Hackathon · Impact Track · Health & Sciences prize

Project Rosie

Built by Shashank Padala

A veterinary-oncology research prototype with code for ranking neoantigens and templating an mRNA synthesis specification. The public demo is synthetic; we did not run the tumor-to-construct pipeline or confirm its one-hour timing.

I also participated in this hackathon with Memory Moment. This coverage was prepared independently after results were announced.

The idea worth stealing

Let the model interpret; never let it write the order form.

Rosie first had Gemma 4 draft the mRNA synthesis specification that goes to a contract manufacturer, then replaced it with a Jinja template. A formulation scientist who spots one hallucinated catalog number dismisses the whole tool, and the manufacturing science does not vary by patient. So the deterministic pipeline owns everything auditable, and Gemma works around it: flagging a VCF with no matched normal before the pipeline runs, reading the result charts to write the clinical report, and narrating what changes when the vet drags a threshold slider.

What we checked

Date checked: .

Public repository at the submission revision, Kaggle writeup, and linked artifacts; code was read, not executed. Deployment checks and limits are described below.

Labels apply to each finding and the evidence described, not to the project as a whole.

Not independently verified means the available evidence was insufficient to confirm a claim. It does not mean the claim is false.

  1. Gemma wrote the order form, then lost the job

    Verified

    Commit 7ba0dc2 added a Gemma-written synthesis specification; commit 22bd44d on May 9 replaced it with a Jinja template whose 26 variables are all case data. The catalog numbers are constants. Source

  2. The advisor never blocks

    Verified

    The pre-flight route asks for strict JSON, parses it defensively, and on any failure returns nothing rather than an error; the comment says never break the Submit flow. Source

  3. The public demo case is synthetic

    Verified

    build_demo.py writes seven plausible canine neoantigens across PIK3CA, TP53, BRCA2, KIT, and PTEN, then runs the real chart, mRNA-design, and template code over them. The writeup discloses this. Source

Under the hood

A hundred lines of scoring
Candidates weaker than 500 nM binding or below 1 percent VAF are dropped; the rest are ranked by 0.5 log-scaled IC50, 0.3 BigMHC immunogenicity, and 0.2 VAF. Anyone can read it.
The sliders never call the model
Re-ranking by IC50 and VAF threshold happens in the browser; Gemma narrates only when the vet presses a button, with the case loaded into its system prompt.

Nuance

Gemma here is gemma-4-26b-a4b-it on Vertex AI, so on-premise inference is a possibility the writeup raises, not the deployment. The live app answered, and the demo page shows the synthetic case; we did not submit a VCF or run the Cloud Run pipeline, and there are no automated tests. Clinical validation is one reviewer’s opinion, which the writeup says.

Sources

Last updated: . Editorial updates do not imply a new technical check.

Builder credits reviewed: , using the official announcement, submission, and any linked credit sources.

Builder? Add context, request a correction, or ask for a re-check →