Case study

Agentic Research Notebook

Local-first assistant for literature sprints

Back to ProjectsLive demo

Project Snapshot

A notebook app that spins up LangGraph agents to digest papers, enrich citations, and track experiment branches.

LangGraphNext.jsSQLite
Source

Skills Flexed

  • Azure OpenAI prompt engineering with JSON schema responses
  • Next.js 15 App Router UI + accessibility-first interactions
  • Adaptive learning logic: diagnostics, drift control, and retry flows
  • PromptOps harness for regression testing evaluation suites

What it is

Agentic Research Notebook is my weekend playground for combining structured note-taking with autonomous reading. I wanted a way to drop PDFs or arXiv links into a workspace and have focussed agents surface claims, run quick fact-checks, and tee up experiments while keeping everything reproducible.

How it works

  • Workspace graph · Each paper spins up a LangGraph flow that extracts claims, links them to citations, and logs follow-up questions inside a local SQLite knowledge base.
  • Tool belt · The agent can call an embeddings-powered search, a lightweight code runner (Pyodide), and a Zotero bridge for bibliography hygiene.
  • Trust layer · Every generated insight links back to the exact snippet or figure, and a “dissenting evidence” tool flags conflicting sources automatically.

Why it matters

  • Keeps me honest during fast research spikes by tracking assumptions and evidence in one place.
  • Lets collaborators replay my investigation by stepping through the stored graph state.
  • Works fully local-first - no cloud dependencies - so I can use it on client-sensitive literature too.