Aetheria

Drug Discovery Multi Agent-System

Comment

Inspiration

In the high-stakes world of pharmaceuticals, the journey from clinical hypothesis to market readiness is incredibly fragmented. Decision-makers are forced to manually sift through biomedical literature, patent filings, supply chain reports, and market analytics to validate a single drug repurposing opportunity. This traditional process takes months of manual review.

We asked ourselves: What if the future of medical research was fully agentic? We were inspired to build Aetheria, a platform that mimics a 24/7 cross-functional R&D team. We wanted to create an intelligence orchestrator that could not only fetch data but dynamically reason through complex clinical and commercial variables to turn months of research into minutes of automated, deeply validated insights.

What it does

Aetheria is a production-ready, multi-agent Pharmaceutical Strategic Intelligence Orchestrator. It acts as an autonomous research team dedicated to analyzing drug repurposing, market viability, supply chain risks, and patent landscapes.

At its core, Aetheria features a Session-Based Routing Layer that dynamically switches between two cognitive modes:

  • Deep Research Mode (Slow-Thinking): A deliberate orchestrator that generates comprehensive, multi-section strategic dossiers. If you ask it to "Assess Minocycline repurposing for CNS indications," it deploys a swarm of specialized agents to investigate the market, patents, and clinical trials, synthesizing a citation-backed markdown report with automated data visualizations.
  • Lite Mode (Fast-Thinking): A rapid-response assistant that queries the context of the generated deep reports, internal secure documents (via RAG), and live web data for immediate facts (e.g., "What is the CAGR of the depression market?").

Key UI/UX Features: Beyond text, Aetheria features an Interactive Knowledge Graph that visually connects drugs, disease states, pathways, and companies. It also maintains a Research Timeline so users can track all nested queries, and provides an interactive 3D molecule viewer for compounds.

How we built it

Aetheria uses a Hub-and-Spoke architecture managed by a RouteLayer built on FastAPI.

  • Orchestration: We used LangChain, LangGraph, and DeepAgents to manage the complex state and multi-agent conversations.
  • Agent Ecosystem:
    • The Router (Session Manager): Tracks conversation history and routes broad queries to the Deep Agent and follow-ups to the Lite Agent.
    • Deep Research Agent (The Brain): Powered by OpenRouter (x-ai/grok-4.1-fast), it oversees specialized worker agents.
    • Lite Agent (The Assistant): Powered by Google Gemini for rapid, context-aware Q&A and embeddings.
    • Specialized Workers: We built custom tool-agents for specific domains. A PubMed Agent (BioPython/NCBI) and Web Intelligence Agent (Tavily) pull live real-world data. We also built Mock Enterprise APIs for IQVIA Insights, EXIM Trends, and Patent Landscapes to simulate enterprise data fusion.
  • Internal RAG & Visualization: We utilized Pinecone and Gemini Embeddings for secure document retrieval, while a dedicated Visualization Agent uses Groq (Moonshot/Kimi) to write and execute Python code (Seaborn/Matplotlib) to render charts on the fly.

Challenges we ran into

  • Cognitive Routing: Building a router that flawlessly understands when a user needs a 10-page analytical dossier versus a quick 2-sentence fact-check was tricky. We had to heavily refine our session manager to maintain context without triggering the heavy, token-intensive Deep Agent unnecessarily.
  • Agent Hallucination in Multi-Step Logic: When fusing real-world biomedical data (PubMed) with mock enterprise data (IQVIA/EXIM), the Deep Agent occasionally fabricated relationships. We solved this by strictly scoping the specialized worker agents' prompts and adding a validation step before final synthesis.
  • Safe Code Execution: Having an LLM generate and execute Python code for data visualization is inherently risky. We had to carefully sandbox the Matplotlib/Seaborn generation environment to ensure the app wouldn't crash from syntactical errors.
  • Dynamic Knowledge Graph Extraction: Parsing complex, multi-layered markdown reports into structured JSON nodes and edges for the frontend interactive Knowledge Graph required intense prompt engineering to ensure accuracy.

Accomplishments that we're proud of

  • The Hybrid "Fast/Slow" Architecture: We successfully created an AI that actually feels like a human research team. The seamless transition between the heavy-lifting Deep Agent and the conversational Lite Agent provides an incredible UX.
  • Intelligent Data Fusion: Blending external web search, live PubMed APIs, internal RAG (PDFs), and mock tabular databases into a single, highly readable, citation-backed markdown report.
  • The UI/UX Experience: We didn't just build a terminal script; we built a polished, production-ready frontend featuring interactive dynamic knowledge graphs, research timelines, and inline document querying that directly aids the researcher's workflow.

What we learned

  • Swarm > Zero-Shot: We learned the true power of Agentic workflows over standard LLM prompting. Delegating specific tasks (e.g., patent search vs. clinical trial analysis) to smaller, specialized sub-agents dramatically reduces hallucinations and increases the depth of the final output.
  • Model Optimization: Not every task requires the heaviest model. Routing tasks effectively—using Grok for deep orchestration, Groq for rapid sub-agent tasks, and Gemini for embeddings/lite Q&A—optimized our latency and token costs significantly.
  • Trust in AI: In pharmaceuticals, "black box" AI is useless. We learned that exposing the AI's "thought process" via the LangGraph execution steps and the Research Timeline is critical for user trust.

What's next for Aetheria

  • Live Enterprise Integrations: Replacing our Mock APIs with live enterprise data connections (real IQVIA, Cortellis, USPTO APIs) to make it fully deployable for existing pharma giants.
  • Collaborative Sessions: Allowing multiple researchers (analysts, scientists, and managers) to interact with and query the same dynamic dossier in real-time.
  • Interactive Dashboards: Upgrading the Visualization Agent to generate fully interactive web widgets (like D3.js or Plotly charts) rather than static base64 images.
  • Proactive Alerting: Implementing a chron-job feature where Aetheria proactively scans the web and patents for updates on a user's saved research targets, sending "Patent Cliff" or "Trial Failure" alerts automatically.

Built With

Share this project:

Updates