Agentic Research Assistant

Inspiration

Researchers, students, and inventors lose days on literature and prior-art searches. Before filing my own patent (an AI-based smart irrigation and nutrient management system), I spent hours digging through patent databases and papers, then manually comparing them against my idea. I wanted an assistant that behaves like a small research team instead of a single chatbot.

What it does

Agentic Research Assistant takes a research question or an invention idea and runs it through a pipeline of specialised AI agents:

1. Prior Art Search Agent searches patent and paper sources for related work.
2. Novelty Diff Agent compares your idea against each result and highlights what overlaps and what is genuinely new.
3. Claim Drafter Agent drafts structured claims or a cited research brief from the findings.
4. Examiner Simulator Agent critiques the output the way a patent examiner or reviewer would, so weaknesses show up before submission.

You watch each agent's progress live, then get a cited brief with key findings, gaps, and suggested next steps in minutes.

How we built it

• Backend: Node.js, TypeScript, and Express, with each agent as its own module and prompt
• LLM: Anthropic Claude API for reasoning, comparison, and drafting
• Data: PatentsView API for prior art retrieval
• Frontend: React, Vite, and Tailwind CSS
• Streaming: Server-Sent Events, so users see live agent progress instead of a loading spinner
• Validation: We tested it on our own irrigation patent, and it surfaced known prior art references (CN120345527A and CN119586529A) that we had found manually.

Challenges we ran into

• Keeping agents from repeating each other's work, which we solved with strict input and output schemas between agents
• Reducing hallucinated citations by forcing every claim to link back to a retrieved source
• Handling slow multi-step pipelines gracefully, which is why we stream progress

Accomplishments that we're proud of

• A real end-to-end pipeline validated against a real patent case
• Live agent-by-agent progress that makes the reasoning transparent
• Output that is cited and traceable, not just fluent text

What we learned

Multi-agent design works best when each agent has one narrow job and a structured hand-off. Grounding outputs in retrieved sources matters more than prompt cleverness.

What's next for Agentic Research Assistant

• Add academic paper sources (arXiv, Semantic Scholar) beside patents
• Export briefs and claim drafts to PDF or DOCX
• Let users save projects and compare multiple ideas side by side
• Add a human-in-the-loop review step before final output

Built With

• anthropic-claude
• express.js
• llm
• multi-agent-systems
• node.js
• patentsview-api
• react
• restapi
• server-sent-events
• tailwindcss
• typescript
• vite