Unpill

An AI deprescription agent that safely removes medications for patients

Comment

Inspiration

Every year in the US alone, adverse drug events drive nearly 100,000 emergency hospitalizations for older adults. The root cause is polypharmacy. Today, over 40% of seniors take five or more prescription medications, creating a minefield of toxic interactions and compounded side effects.

Doctors know they need to deprescribe, but manually pulling lab results, performing complex renal math, and cross-referencing thousands of pages of geriatric guidelines in a 15-minute consultation is virtually impossible.

I built Unpill to bridge that gap: a safe, mathematically grounded clinical AI that doctors can actually trust for deprescribing medications with a safe plan.

What it does

Project Unpill is a stateless, SMART on FHIR-integrated clinical agent that performs automated, mathematically perfect deprescribing audits.

When a clinician selects a patient in our UI, Unpill instantly:

  • Pulls Live Data: Reaches into the hospital's EHR to pull the patient's vitals, active medications, and recent labs (like serum creatinine).

  • Calculates Renal Function: Bypasses the AI entirely to perform deterministic Cockcroft-Gault calculations for precise Creatinine Clearance (CrCl).

  • Cross-References Guidelines: Simultaneously evaluates the patient's state against the 2023 AGS Beers Criteria, the NLM RxNav Toxicology database, and the Nebraska Deprescribing Protocols.

  • Generates an Action Plan: Outputs a structured clinical report detailing exact dosage adjustments (e.g., reducing Apixaban for renal impairment) and step-by-step tapering schedules (e.g., safely weaning off Zolpidem to reduce fall risk).

How we built it

We built a true zero-trust, multi-agent architecture rather than a simple chatbot.

  • Frontend: We utilized the Prompt Opinion platform as our marketplace and UI layer.

  • Backend: A completely stateless orchestrator deployed on Google Cloud Run, built using the po-adk-python framework.

  • Zero-Trust Tools: We wrote strict Python tools that force the AI to use hardcoded logic for medical math and guideline retrieval, entirely preventing LLM hallucination.

Accomplishments that we're proud of

  • Flawless Clinical Math: Successfully isolating medical calculations from the LLM to guarantee 100% accurate Creatinine Clearance every single time.

  • Safety First: Building a system that actively refuses to guess. If a patient's FHIR record is missing a critical lab result (like Creatinine), our agent safely aborts the audit rather than hallucinating an assumption.

  • The Interoperability Handshake: Creating a seamless, stateless bridge between a beautiful UI frontend, a Google Cloud Run backend, and a live SMART on FHIR database.

What's next for Unpill

The plumbing is built, and the architecture is proven. Next, we want to scale.

  • Direct EHR Integration: Moving from FHIR sandboxes to live Epic and Cerner App Orchards.

  • Open-Source Fine-Tuning: Leveraging our LiteLLM router to point Unpill toward a custom, fine-tuned medical model (like a specialized Med-Llama) hosted on our own infrastructure.

  • Expanding the Database: Incorporating more global deprescribing protocols (like STOPP/START) to cover a wider array of polypharmacy scenarios beyond the Beers Criteria.

Built With

Share this project:

Updates