Research Mind

Inspiration

Every day, thousands of AI tools promise to "do your research for you" — but most just spit out generic, unverified paragraphs with no sources and no self-awareness about their own quality. We wanted to build something different: an AI system that doesn't just answer, but actually works like a real research team — searching, reading, writing, and then honestly reviewing its own output. The goal was simple: build something we'd actually use tomorrow, not just another AI wrapper.

What it does

ResearchMind takes any topic you give it and runs it through a 4-agent pipeline. A Search Agent finds recent, reliable sources on the web. A Reader Agent picks the most relevant source and scrapes it for deeper content. A Writer Chain turns all that raw research into a structured, professional report — Introduction, Key Findings, Conclusion, and Sources. Finally, a Critic Chain reviews the report it just helped create, scoring it out of 10 and pointing out strengths and gaps — just like a real editor would. The entire pipeline is visualized live in a clean, dark-themed Streamlit UI, and the final report can be downloaded as a Markdown file.

How we built it

We built ResearchMind using LangChain to orchestrate the agent pipeline, with Google Gemini as the underlying LLM for reasoning, writing, and critique. The Search Agent uses the Tavily Search API as a tool to pull fresh, relevant web results, while the Reader Agent uses BeautifulSoup and Requests to scrape and clean actual page content. The Writer and Critic stages are implemented as LangChain prompt chains with structured output parsing, so every report follows a consistent format. The entire experience is wrapped in a custom-styled Streamlit frontend with a live pipeline tracker that shows each agent moving from "waiting" to "running" to "done" in real time.

Challenges we ran into

Getting four independent agents to hand off context to each other cleanly was harder than expected — search results had to be summarized and trimmed before being passed to the reader agent, and the reader's scraped content had to be combined carefully with search data before reaching the writer, without blowing past context limits. Web scraping was another challenge — not every site returns clean, usable text, and we had to handle failures gracefully so one bad URL wouldn't break the whole pipeline. We also spent a lot of time tuning prompts so the Critic Chain gave genuinely useful, specific feedback instead of generic praise.

Accomplishments that we're proud of

We're proud that ResearchMind doesn't just generate a report — it critiques itself, which most AI research tools don't do. We're also proud of the UI: instead of a plain text box and output, we built a live pipeline visualization so users can actually see each agent working, along with transparent access to the raw search and scraped data (not just a black-box final answer). Getting all four agents to work together reliably, end-to-end, without manual intervention, felt like a real milestone.

What we learned

We learned a lot about multi-agent orchestration — specifically how important it is to control what context gets passed between agents, since LLMs perform worse when given too much noisy or unstructured information. We also learned how valuable a "critic" step can be in AI pipelines — having the system evaluate its own output caught weaknesses (like vague sourcing) that we wouldn't have noticed otherwise. On the engineering side, we got much more comfortable with LangChain's agent and chain abstractions and how to combine tool-using agents with simple prompt chains in the same pipeline.

What's next for Research Mind

Next, we want to expand the Reader Agent to scrape and synthesize multiple sources instead of just one, add PDF export alongside Markdown, and give the Critic Chain the ability to trigger an automatic rewrite when it scores a report too low — closing the feedback loop completely. We'd also like to add session memory so users can build on previous research over time, and eventually deploy a live public demo so anyone can try it without setup.