🧬 ZeroHelix

The Offline Pharmacogenomic Firewall.

💡 Inspiration

The inspiration for ZeroHelix came from a terrifying statistic: 100,000 people die annually from Adverse Drug Reactions (ADRs)—deaths that are largely preventable.

We realized there is a "Deadly Gap" in modern medicine. 99.5% of humans carry genetic variants that affect how they process drugs, yet doctors still prescribe "one-size-fits-all" doses. While pharmacogenomics (DNA-based medicine) exists, it has a massive barrier: Privacy.

People are terrified to upload their DNA to the cloud due to data breaches (like the recent 23andMe hacks) and fear of insurance discrimination. We asked ourselves: What if we could bring the Intelligence to the Data, instead of sending the Data to the Cloud?

🤖 What it does

ZeroHelix is a "Zero-Trust Biological Operating System" that lives entirely on your phone. It turns your raw DNA data into a life-saving firewall without ever connecting to the internet.

1. Secure Ingestion: Users import their raw DNA file (from Ancestry/23andMe). ZeroHelix parses it into a secure, encrypted local vault and instantly wipes the original file.
2. Visual Recognition: When a user picks up a medication, they simply scan the box with their camera. Our OCR engine instantly identifies the drug (e.g., "Warfarin").
3. The "Bio-Beacon": The app silently cross-references the drug against the user's specific genetic markers stored in the vault.
4. Offline Reasoning: Using an on-device Small Language Model (DeepSeek R1), it analyzes the interaction using medical logic.
5. Life-Saving Alerts: It warns the user immediately if the drug is toxic for their specific biology (e.g., "STOP! You have a CYP2C9 variant. This dose could be fatal.").

⚙️ How we built it

We built ZeroHelix with a privacy-first architecture using a "Secure Enclave" approach:

• The Brain (AI): We utilized DeepSeek-R1-Distill-Qwen-1.5B, quantized to 4-bit GGUF. We chose this model over standard chat models (like Llama 3) because medical safety requires Chain-of-Thought (CoT) reasoning to handle complex "Polypharmacy" (drug-drug-gene interactions), not just text summarization.
• The Engine: We used the RunAnywhere SDK to load the model directly into the device's NPU/RAM. This ensures that decrypted DNA vectors exist in volatile memory only for the ~100ms duration of the inference and are instantly overwritten.
• The Frontend: Built with Flutter for a smooth, cross-platform mobile experience.
• The Eyes: Integrated ML Kit for real-time text recognition (OCR) to identify drug packaging instantly.

🚧 Challenges we ran into

• The "Heat" Problem: Running a 1.5B parameter model on a mobile device initially caused overheating and lag. We solved this by aggressively quantizing the model to 4-bit and optimizing the inference loop to only activate the "Reasoning Core" when a drug is explicitly detected.
• Llama vs. DeepSeek: Early tests with standard chat models failed to catch complex interactions (e.g., if Drug A inhibits an enzyme needed for Drug B). Switching to DeepSeek R1 gave us the "reasoning" capability to deduce these causal links accurately.
• Parsing Massive Files: Raw .vcf DNA files are huge. Building a parser that could extract only the relevant pharmacogenes (like CYP2D6) without crashing the phone's memory was a significant engineering hurdle.

🏆 Accomplishments that we're proud of

• True Offline Capability: We successfully simulated the "Unconscious Trekker" scenario—saving a patient in a zero-signal zone (like a remote village) where cloud-based AI would have failed.
• Privacy Architecture: We built a system where the user's DNA never leaves the device. We don't even have a server to store it.
• Speed: Achieving inference times of under 2 seconds on a standard Android device, making the "scan and check" experience feel instantaneous.

🧠 What we learned

• Small Models are Mighty: You don't need a massive server farm to save lives. A well-tuned, quantized SLM (Small Language Model) can outperform cloud giants when applied to a specific domain.
• The Power of CoT: Chain-of-Thought reasoning is essential for medical AI. The ability of the model to "show its work" (e.g., Why is this drug dangerous?) builds trust with the user.
• Privacy is a Feature, not a Constraint: Designing for "Zero Trust" actually simplified our architecture by removing the need for complex cloud security and compliance layers.

🚀 What's next for ZeroHelix

• Wearable Integration: Pushing alerts to smartwatches so a medic can see a patient's "Bio-Beacon" status (e.g., "Allergic to Penicillin") even if the phone is locked.
• Food-Gene Interactions: Expanding the database to scan food barcodes (e.g., warning a user with G6PD deficiency about Fava beans).
• Doctor Mode: A specialized interface for emergency responders to quickly scan a patient's QR code and get a "Safe List" of medications during trauma care.

👥 The Team: GenAI Taskforce

• Pranay Gujar
• Vishwakarma Institute of Technology, Pune

Built With

• android
• c++
• dart
• deepseek-r1-distill-qwen-1.5b
• executorch
• flutter
• gguf
• google-ml-kit-(ocr)
• nosql-(local-vector-store)
• python