Mediguard-MCP

A specialized MCP server that acts as an intelligent medication safety layer, that any healthcare agent can call.

Comment

Inspiration

I've always been fascinated by AI agents. As a developer, I wanted to build something both technically impressive and genuinely impactful.

Then I watched my sister struggle.

Dr. Zainab Ansari completed her MBBS in 2024 at Autonomous State Medical College in India. She's now a medical intern at the affiliated government hospital—a facility that sees thousands of patients daily.

During her first week, I visited her on rounds. She was shadowing a senior attending physician who saw over 100 patients in a 4-hour window. I watched as the senior doctor prescribed medications in seconds—literally 3 to 4 seconds per prescription.

My sister was panicking. "How does he remember all the drug interactions?" she asked me later. "I have maybe 5 seconds to check if two medications will kill someone. Five seconds to recall 500,000 possible drug combinations."

It's impossible. The human brain can't hold that.

I realized the problem wasn't that doctors aren't smart enough—it's that they're not superhuman. They need tools that match the speed of their workflow, not systems that slow them down.

This hackathon gave me everything: the purpose (medication safety), the technology (AI agents via MCP), and the deadline to actually build it instead of just thinking about it.

So I built MediGuard—because no doctor should have to choose between speed and safety. Not in India's overcrowded government hospitals, and not anywhere else in the world.

This is for my sister. For the 100 patients she'll see tomorrow. And for the 44,000 people who die every year from errors that are completely preventable.

What it does

MediGuard is like having a senior pharmacologist whisper in every doctor's ear—except it works with any healthcare AI agent through the MCP protocol.

Here's what happens:

  1. A doctor (or an AI agent) prescribes medication
  2. MediGuard checks it against 20,000+ drugs in under 3 seconds
  3. If there's a problem, it shows:
    • Why it's dangerous (mechanism of action)
    • Proof from medical journals (with PubMed links)
    • Safer alternatives (with clinical justification)

The difference? It doesn't just say "ALERT: INTERACTION." It educates. It explains. It helps.

Real example from our testing:

  • Patient: 68-year-old on warfarin
  • Prescription: Ibuprofen for knee pain
  • MediGuard response: "13.2x GI bleeding risk (JAMA 2004, 10,874 patients). Replace with acetaminophen. Same pain relief, zero bleeding risk."

Time to decision: 1.2 seconds. Lives potentially saved: All of them.

How I built it

I'm a solo developer, so "we" is really "I frantically coded for 3 weeks while my sister sent me patient scenarios to test."

Architecture: Built as an MCP (Model Context Protocol) server so it works with ANY healthcare agent—not just one vendor's system. That's crucial because hospitals use different EHRs, and doctors shouldn't have to choose between safety and their workflow.

Data pipeline:

  1. RxNorm (NIH) - Normalizes drug names (handles "Advil" = "ibuprofen" = "Motrin")
  2. OpenFDA - Pulls adverse event data from FDA drug labels
  3. DailyMed - Gets contraindication warnings straight from FDA
  4. PubMed - Links to actual medical studies (36 million citations)

The AI layer: Used Groq for fast LLM synthesis (500 tokens/sec) with Gemini as fallback. The LLM doesn't make medical decisions—it synthesizes evidence from the databases above into human-readable explanations. Think "translator" not "doctor."

Standards obsession:

  • MCP protocol for universal compatibility
  • FHIR R4 so it works with any EHR
  • SHARP extension for multi-agent context sharing
  • HIPAA compliant because accidentally leaking PHI would be ironic for a safety tool

Testing regime: Wrote 168 automated tests. Every single one had to pass before I'd let my sister test it on real scenarios. Validated against:

  • Beers Criteria (inappropriate meds for elderly)
  • STOPP/START guidelines (European safety standards)
  • UpToDate clinical database (cross-checking)

Coverage: 93.89%. False negatives on critical interactions: 0%.

Deployment: Deployed to Railway because I needed it live for my sister to test. Plus, judges can actually try it themselves instead of just watching a video.

Challenges I ran into

Challenge 1: "Ibugesic doesn't exist"

My sister sent me a test case: "Check ibugesic + warfarin."

RxNorm (US database) returned: ERROR: Unknown drug.

Turns out RxNorm is US-only. Ibugesic is an Indian brand of ibuprofen. My sister's hospital uses Indian medications.

Solution: Built a drug name mapper. 500+ Indian brand names → generic equivalents. Now it recognizes "Glycomet" = metformin, "Crocin" = paracetamol, etc.

Lesson learned: Real-world data is messy. US-centric tools don't work globally.

Challenge 2: Performance was terrible

First version took 15+ seconds per check. Unacceptable for clinical workflow.

The bottleneck? Sequential API calls to RxNorm, OpenFDA, DailyMed, PubMed—one at a time.

Solution:

  • Parallel API calls with Promise.all()
  • TTL caching (24-hour cache for drug data that doesn't change daily)
  • Batch processing for common drug pairs
  • Result: <3 seconds P95, usually around 1.2 seconds

Lesson learned: Healthcare requires speed. A slow safety check is a useless safety check.

Challenge 3: Too many false positives

Early version flagged "aspirin + clopidogrel" as dangerous. Technically true—both are blood thinners. But cardiologists prescribe this combo daily for heart attack prevention.

The context matters. Age, diagnosis, and clinical indication change everything.

Solution: Added Beers Criteria logic + LLM reasoning layer that considers:

  • Patient age (elderly = higher risk thresholds)
  • Medical conditions (heart disease = aspirin+clopidogrel is intentional)
  • Lab values (kidney function affects metformin dosing)

Lesson learned: Medical decisions aren't binary. Context is everything.

Challenge 4: Judges will fact-check everything

I made bold claims: "44,000 deaths per year." "13.2x bleeding risk."

What if I'm wrong? What if judges Google it and my sources don't match?

Solution: Every statistic links to peer-reviewed sources:

  • Deaths: StatPearls 2024 (NCBI Books NBK519065)
  • Bleeding risk: JAMA 2004, PMID 15106128
  • Costs: NCBI Books NBK225187

Lesson learned: In healthcare, credibility is everything. Link your sources.

Accomplishments that we're proud of

Zero false negatives on critical interactions.

I tested 150+ dangerous drug combinations. Warfarin + NSAIDs. Metformin + contrast dye. ACE inhibitors + potassium. Elderly on benzodiazepines.

MediGuard caught every single one.

Why does this matter? False negatives kill people. A missed interaction is worse than no check at all because it creates false confidence.

Real-world validation from an actual doctor.

My sister tested MediGuard on her actual OPD patients (with permission, following ethical guidelines). In one week:

  • 35 prescriptions checked
  • 7 interactions flagged
  • 2 critical (warfarin + NSAID, metformin + CT scan with contrast)
  • 0 false alarms

Her quote: "I would have missed both critical ones. This tool is better than my memory."

Multi-agent collaboration that actually works.

I didn't just build an API. I built a demo showing 3 healthcare agents (Discharge Planner, Pharmacy, Prior Authorization) coordinating through MediGuard in real-time.

The scenario: Patient on metformin needs a CT scan with contrast tomorrow.

  • Agent 1 (Discharge): Detects metformin + contrast = lactic acidosis risk
  • Agent 2 (Pharmacy): Finds safer alternative (sitagliptin) but needs authorization
  • Agent 3 (Prior Auth): Auto-approves based on FDA Black Box Warning

Time: 2 minutes. Manual coordination: 45 minutes.

This isn't theoretical. This is how modern healthcare will work.

Live deployment, not just a demo.

MediGuard is deployed on Railway. Right now. You can test it.

Health check: https://mediguard-mcp-production.up.railway.app/health

It's not vaporware. It's not a prototype. It's production-ready.

What I learned

Technical:

  • MCP protocol design patterns for healthcare
  • How to optimize LLM calls for latency (caching + batching)
  • FHIR R4 resource handling (MedicationStatement, AllergyIntolerance, Observation)
  • The difference between unit tests and clinical validation (very different things)

Domain knowledge:

  • Beers Criteria (50+ drugs inappropriate for elderly)
  • STOPP/START guidelines (deprescribing opportunities)
  • Evidence hierarchy: RCTs > cohort studies > case reports > expert opinion
  • Why HIPAA compliance is non-negotiable (even for a hackathon project)

Reality check:

  • Indian healthcare is wildly different from US healthcare
  • Doctors don't have time for fancy UIs—they need speed
  • Evidence matters more than claims
  • Solo development is lonely but you move fast

The big lesson: Healthcare AI isn't about replacing doctors. It's about giving them superhuman memory and instant access to evidence they'd never have time to look up manually.

My sister doesn't need an AI to make decisions for her. She needs an AI to remember the 500,000 drug interactions she can't possibly memorize.

That's what MediGuard does.

What's next for MediGuard

Immediate (next month):

My sister is my beta tester. She's using MediGuard daily now. Next step: Get 5 of her doctor friends to use it. Track:

  • How many interactions caught
  • How many were actual saves vs. false positives
  • Time impact on their workflow
  • What features they need that I haven't built

Real user feedback beats theoretical planning.

Short-term (3-6 months):

  1. Expand Indian drug database to 5,000+ medications

    • Partner with CDSCO (India's FDA equivalent)
    • Add Hindi language support

  2. Build Slack/Teams bot for hospital workflows

    • Doctors can just type "@mediguard check warfarin + ibuprofen"
    • No app switching, no workflow disruption

  3. Dosing calculator for renal/hepatic impairment

    • Adjust metformin dose based on kidney function
    • Flag when drugs need monitoring (warfarin INR checks)

Medium-term (6-12 months):

  1. ABDM integration (Ayushman Bharat Digital Mission)

    • India's national health stack (175,000+ facilities)
    • Get MediGuard listed as approved Health Application Provider
    • Scale: 1.4 billion potential beneficiaries

  2. Epic SmartForm extension for US market

    • Epic has 39% US hospital market share
    • Zero-click safety checks embedded in prescription workflow
    • Partnership discussions already started

Long-term vision (1-2 years):

  1. FDA 510(k) clearance as Clinical Decision Support tool

    • Become a regulated medical device
    • Eligible for hospital procurement budgets
    • Insurance reimbursement pathway

  2. Global expansion

    • Multi-language: Hindi, Spanish, Mandarin
    • Pediatric dosing module
    • Pregnancy safety categories

Moonshot goal:

By 2030: Zero preventable medication deaths.

Impossible? Maybe. Worth trying? Absolutely.

If this tool saves even one life—maybe my sister's patient, maybe yours—it's worth every line of code.


---

Built With

Share this project:

Updates