Gemma 4 Good Hackathon submission — Health & Sciences track Gemma 4 running entirely on local hardware. No cloud. No API keys. No internet required.
There are 6 million+ Variants of Uncertain Significance in ClinVar. A clinical geneticist reviewing a rare disease case must manually cross-reference dozens of high-pathogenicity candidates against a patient's phenotype profile — a process that takes hours and requires specialist expertise that most of the world doesn't have access to.
In resource-limited settings — rural hospitals, low-income country clinics, disaster zones — there is no cloud genomics service, no specialist on call, no reliable internet. The patient waits. The diagnosis doesn't come.
A fully local clinical reasoning agent powered by Gemma 4. The doctor photographs a patient's clinical report on any Android phone running Gemma. Gemma 4 E2B (running on the phone via LiteRT) extracts phenotype terms from the image. Those phenotypes are sent over local WiFi to a desktop/laptop running Gemma 4 E4B via Ollama.
The E4B model uses Gemma 4's native function calling to navigate a 57,124-node gene-disease knowledge graph built from real ClinVar and HPO data — hopping from phenotype → disease → gene → variant nodes until it flags the most likely causal genetic variant.
Everything runs offline. Nothing leaves the building.
Baseline benchmark — gemma4:e4b, 10 episodes x 3 task types (seeds 1-10)
| Task | Success Rate | Mean Reward | Avg Steps | Clinician Candidates | Agent BFS Hops | Burden Ratio |
|---|---|---|---|---|---|---|
| Monogenic | 30% | 0.333 | ~7 | 6.2 | 3.0 | 2.07x |
| Oligogenic | 10% | 0.192 | 25 | 9.5 | 3.0 | 3.17x |
| Phenotype Mismatch | 40% | 0.441 | ~5 | 8.2 | 3.0 | 2.73x |
Burden ratio = high-pathogenicity candidates a clinician must manually review divided by the agent's graph hops to reach the causal variant. Higher = greater reduction in cognitive load.
Zero-shot, no fine-tuning. Full benchmark log: results/baseline_e4b.json
This phone→desktop split keeps extraction on the edge (E2B / LiteRT) and complex reasoning on local compute (E4B / Ollama) — works with any Android phone that can run Gemma.
| Feature | How Used |
|---|---|
| Native function calling | Graph navigation actions are real tool calls — no JSON parsing, no regex |
| Multimodal (vision) | Clinical report image → phenotype extraction on phone before episode starts |
| E2B edge model (LiteRT) | On-device inference on any Android phone via LiteRT — fully offline intake |
| E4B local deployment (Ollama) | Zero cloud dependency, works fully offline |
For each candidate gene, GenoPath computes HPO terms associated with that gene's diseases that the patient lacks. This enables elimination reasoning without any fine-tuning:
EXCLUSION SIGNALS:
BRCA1: patient LACKS → Breast carcinoma, Ovarian neoplasm
KCNQ1: patient LACKS → Prolonged QT interval, Atrial fibrillation
Gemma 4 E4B uses this signal zero-shot — structured data + a clinical system prompt is enough.
# 1. Pull Gemma 4 E4B via Ollama (one-time, ~9.6GB download)
ollama pull gemma4:e4b
# 2. Install Python deps
pip install -r requirements.txt
# 3. Set up env
cp .env.example .env
# 4. Build graph data — only needed if clinvar_pathogenic.tsv is missing
# (it is included in the repo, so skip this unless regenerating)
python scripts/filter_clinvar.py
# 5. Launch the app
python src/ui/app.py
# → Open http://localhost:7860
# 6. (Optional) Phone intake server — separate terminal
uvicorn src.server.intake_api:app --host 0.0.0.0 --port 8000Desktop / Laptop (reasoning device):
- GPU: 4060 Ti 16GB VRAM or equivalent
- Gemma 4 E4B at Q4_K_M ≈ 9GB VRAM — fits comfortably
- Do NOT attempt 27B — needs ~17GB VRAM, will OOM on 16GB
- Runs at ~20 tok/s
Phone (intake device) — any Android with Gemma support:
- Any Android phone running Gemma via AI Edge Gallery or a custom LiteRT APK
- E2B (2B 4-bit): ~1.5GB — fits any phone with >=4GB RAM, nothing to close
- E4B (4B 4-bit, "performance mode"): ~2.5GB — needs >=6GB RAM, close heavy apps first
- Phone will warm up during sustained inference — normal, not damaging
Task | Avg high-path candidates | Avg BFS hops to causal | Burden ratio
monogenic | 6.2 | 3.0 | 2.07x
oligogenic | 9.5 | 3.0 | 3.17x
phenotype_mismatch | 8.2 | 3.0 | 2.73x
A clinician manually reviews ~6 high-risk candidates for a monogenic case. GenoPath reaches the causal variant in 3 graph hops — a 2x reduction in cognitive load, zero-shot.
- HPO Ontology (
data/hp.obo) — Human Phenotype Ontology, 19,389 terms. hpo.jax.org - ClinVar Pathogenic Variants (
data/clinvar_pathogenic.tsv) — ~92,000 GRCh38 expert-reviewed pathogenic variants, filtered from ClinVar variant_summary.txt byscripts/filter_clinvar.py.
genopath/
├── src/
│ ├── graph/ # GenoPathGraph: build + query the knowledge graph
│ ├── episode/ # Environment, Pydantic models, reward logic
│ ├── agent/ # Ollama client, GENOPATH_TOOLS, phone intake
│ ├── server/ # FastAPI phone→desktop intake endpoint
│ └── ui/ # Gradio demo UI
├── scripts/
│ ├── filter_clinvar.py # Filter raw ClinVar → clinvar_pathogenic.tsv
│ ├── sanity_check.py # Integration smoke test (run after every change)
│ └── cognitive_load.py # Clinician burden vs agent hops analysis
├── training/ # Optional: Unsloth GRPO notebook (Colab A100/L4)
├── data/ # hp.obo + clinvar_pathogenic.tsv
└── results/ # Benchmark outputs
| Prize | Evidence |
|---|---|
| Main Track | Working Gradio demo + 3-min YouTube video |
| Health & Sciences | Real ClinVar/HPO data, rural clinic framing |
| Ollama ($10K) | ollama pull gemma4:e4b above, Ollama visible in demo video |
| LiteRT ($10K) | MediaPipe LLM Inference code in src/agent/intake.py, AI Edge Gallery on Android in video |
| Cactus ($10K) | Phone→desktop routing diagram above, routing logic in code |
| Unsloth ($10K, optional) | HF adapter published, training notebook, before/after benchmark |
Krishna Venkatesh — GitHub