Alignment Graph Update, July 2026

Alignment Graph is a research tool that connects related papers to help AI Safety researchers discover relevant topics and manage their research.

The tool ingests and processes AI Safety and Alignment papers from the arXiv API, then visualises how they relate to each other as a graph in the browser.

The main technical challenges were:

• Building a service that can maintain and update the relationships between papers.

• Designing a UI that clearly expresses how papers connect.

Origins

The initial architecture centred on getting the backend data structure right. Before anything else, we needed a reliable way to source data and relate it accurately. We built an offline Python pipeline that runs in stages, in any order, from the CLI: it ingests, filters, embeds, clusters, and saves data to SQLite, ultimately producing the JSON consumed by a simple React SPA. This was sufficient up to roughly 4,000 papers, at which point browser graphics started to stutter and hindered the visualisation. That ceiling meant we couldn't yet test how useful the tool is against the full AI Safety corpus, but it let us validate the pipeline itself.

Current architecture

The architecture has since evolved to support over 10,000 papers. This meant moving from local SQLite to Supabase-hosted Postgres, served through a FastAPI REST API hosted on Railway. The frontend remains a React SPA deployed on Netlify. To address visualisation clutter, we introduced subgraph rendering in place of always rendering the full graph, alongside a new view with granular filter and search controls that improves paper discovery and supports creating and editing subgraphs.

App state falls into two categories: papers state and subgraph state. Papers state powers the dashboard (StatsView.tsx) and is fetched from FastAPI; since it can be large, requests use server-side filtering and search, and are paginated in pages of up to 200 papers. Subgraph state is currently managed via localStorage, with functions to store and fetch it locally- a fast-prototyping choice ahead of full auth support.

Because state ownership sits outside the React component tree- papers via FastAPI, subgraphs via localStorage- we added React Query to cache client-side data and avoid refetching when nothing has changed on the backend.

What's next

Development is currently steady. The next cycle focuses on migrating the heavy ML embedding workloads to hosted GPUs. Future cycles may add an AI layer letting users query the relationships between papers and surface undiscovered papers relevant to their custom graphs.

Other Considerations

Alignment Graph is built primarily as open-source software, for users to run and modify themselves. A production version ships regularly to alignmentgraph.com, though some of the heavier ML features are currently gated to local/self-hosted use.

I'm building out the key features first, UX refinements are an ongoing process. Apologies for any ugly divs in the meantime.