Dr. Ordinary and Dr. Strange

Inspiration

Adverse drug interactions lead to millions of preventable medical complications annually. Most existing tools rely on static, rule-based systems that fail to capture the nuanced relationships between multiple drugs or patient-specific contexts. Inspired by AlphaFold’s success in modeling complex biological structures through deep reasoning, we set out to build a more intelligent, context-aware solution for detecting drug-drug interactions.

What it does

Dr. Ordinary is our baseline: a traditional drug interaction checker that mimics existing static tools, but adds a touch of convenience. Dr. Strange is the real innovation: a multi-agent, LLM-powered system that analyzes drug combinations using chain-of-thought reasoning, contextual patient data, and a gating network to weigh agent outputs. The final results are summarized into a readable, clinically useful interaction report.

How we built it

• Used a modular agent framework (Letta) to reason about drug combinations and interactions

• Built a frontend interface that accepts user input (drug list, patient context)

• Developed multiple sub-agents for specific reasoning tasks (toxicity, metabolic competition, contraindications, etc.)

• Introduced a gating network to weigh outputs from each agent

• Added a summarizing agent to compile a final, human-readable report

• Modeled our architecture loosely on AlphaFold’s iterative reasoning and relation modeling framework

Challenges we ran into

• LLMs hallucinating or providing vague/conflicting interaction details

• Balancing verbosity and accuracy in the summarizing agent’s outputs

• Creating a robust API flow between agents without excessive latency

• Designing the gating mechanism to properly weigh agent confidence

• Mapping drug names and dosages across inconsistent user input

Accomplishments that we're proud of

• Successfully replicated AlphaFold-like reasoning in a drug safety context

• Built an end-to-end pipeline that outputs richer, smarter DDI reports

• Created a modular agent-based architecture that is extensible and generalizable

• Delivered readable explanations for complex drug risks — not just raw flags

What we learned

• Chain-of-thought prompting significantly improves LLM reasoning in medical domains

  • Agent specialization + a gating mechanism creates better decisions than a single monolithic model
  • Patients and clinicians both benefit from human-readable AI explanations
  • Iterative architecture (like AlphaFold) can be abstracted into other domains

What's next for Dr. Ordinary and Dr. Strange

• Integrate more patient-specific parameters (e.g., renal function, allergies, vitals)

• Expand beyond drug-drug to include drug-food and drug-condition interactions

• Offer fine-tuned model variants for hospitals or pharmacists

• Build EHR plugins and a mobile interface for real-time use in clinical environments

• Allow feedback from users to refine and retrain the agent responses dynamically

Built With

• node.js
• python
• typescript