As the pace of biomedical discovery accelerates at an unprecedented rate, every week hundreds of papers are published exploring the complex biological mechanisms of aging, cellular senescence, metabolic pathways, and potential therapeutic interventions. This information overload is overwhelming not only for the science curious public but also for scientists, domain experts, technologists, and investors. What if one could fully automate the research, curation, critical analysis, illustration, and publication of compelling scientific stories from expert content in academic journals on a regular basis? What if this system didn’t just summarize papers in isolation, but actually “remembered” past discoveries, weaving them together into a living, interconnected map of scientific progress in a field?
This was the motivation for me to build Bio By AI – a fully autonomous, agentic system that runs on the cloud, powered by a swarm of specialized AI agents that scan and filter open access literature and create a weekly newsletter of interesting newsworthy stories on these topics. As a start, I have focussed on longevity/healthspan domain which is a long standing interest of mine. So there you have it, a fully autonomous agentic AI generated longevity newsletter and knowledge graph – Bio By AI {Longevity Edition}. Like the concept? Check it out here
In this article, I’ll dive deep into the architecture of this system, exploring the methodology and “algorithm”, the distinct AI agents that bring it to life, the mechanics of its deep knowledge graph, the community feedback loop, and the robust autonomous deployment pipeline that ensures zero human intervention is required.
The Methodology: An Agentic Approach to Curation
When building an automated content engine, a monolithic script attempting to do everything—fetching, summarizing, drawing, and publishing—quickly becomes brittle and unmanageable. Instead, I adopted an agentic architecture.
In an agentic system, complex workflows are broken down and delegated to specialized AI “personas.” Each agent has a singular, well-defined role, a specific set of tools, and a distinct charter. These agents operate sequentially, passing their outputs down the pipeline like a highly efficient, automated newsroom.
The entire pipeline is orchestrated by a central controller that wakes up every week, it triggers a cascade of agentic actions: researching the latest literature, writing the stories, updating the historical database, generating the infographics, and deploying the website.
Let’s meet the team of agents working autonomously behind the scenes.
A Swarm of Agents
1. The Lead Researcher Agent
The pipeline begins with the Researcher agent. This agent interfaces with massive biomedical literature from ArXiv, PubMed, PMC and other open access journals. It isn’t just pulling random papers; it runs targeted queries focusing on longevity, aging, senescence, and other relevant keywords.
The Agent then filters for papers published within the last week, ranks them by relevance and impact, and selects the top three. It parses the XML responses to extract the title, authors, publication date, abstract, and DOI, formatting this raw data into a clean JSON structure ready for the next stage.
2. The Science Journalist Agent
Once the raw data is collected, it is handed over to the journalist agent. Powered by LLM (Gemini in this case), the writer is instructed to adopt the tone of a high-end feature journalist—think Wired or The Atlantic. It eschews clickbait in favor of compelling, accessible storytelling, with a critical lens.
However, it is not just a summarizer; it is a critical thinker who has been instructed to maintain a critical lens that points out the study’s shortcomings, methodological limitations, and the challenges ahead before the research can be translated. This is to ensure the newsletter remains grounded.
3. The Graph Builder Agent
Perhaps the most innovative agent in the swarm is the Graph Builder. As the Journalist finishes a story, the Graph Builder reads the text and performs deep Information Extraction (IE).
Using an LLM, it extracts the core biological entities and the relationships between them to build a comprehensive knowledge graph. It merges these new nodes and edges into a global knowledge graph known here as the Longevity KG
4. The Illustrator Agent
The IIustrator agent takes the generated headline and body of the story and crafts a highly specific prompt for an image generation model
It is instructed to strictly generate sleek, modern scientific infographics and data visualizations, actively avoiding abstract art or irrelevant imagery. These infographics provide a visual anchor for the complex biological concepts being discussed.
5. The Publisher Agent
Finally, all these assets—the text, the insights, the images, and the graph—are handed to the Publisher. The Publisher renders the static HTML site.
It generates a styled front page and individual story pages. It also organizes each week’s content into a perfectly self-contained folder, crawls backward through the repository’s history to retroactively update the older editions, ensuring that the entire site remains interlinked and effortlessly navigable.
The Knowledge Graph: Memory, Insights, and Exploration
One of the greatest limitations of standard automated newsletters is their lack of memory. Bio By AI breaks this paradigm through introduction of its Knowledge Graph (KG).
The Knowledge Graph serves as the long-term memory of the AI swarm, transforming a chronological list of newsletters into a highly structured, interconnected web of biological data.
Entities, Relationships, and the Tech Stack
Under the hood, the Graph Builder agent parses the text and strictly categorizes discoveries into precise Entity Types, such as:
Gene(e.g., FOXO3, SIRT1)Protein(e.g., mTOR, AMPK)Drug/Compound(e.g., Rapamycin, Metformin, NAD+)Disease(e.g., Alzheimer’s, Sarcopenia)Biological Process(e.g., Autophagy, Cellular Senescence)
It then connects these entities using strictly typed Relationships, such as INHIBITS, ACTIVATES, ASSOCIATED_WITH, TREATS, or CAUSES. For example, it might extract the triple: [Rapamycin] -(INHIBITS)-> [mTOR].
This extraction is built on a Python stack using LLM and leveraging the networkx library to manipulate the graph topology and calculate node centrality. The backend serializes this graph into a lightweight knowledge_graph.json file. On the frontend Cytoscape.js—a highly optimized, open-source graph library— is used to render a beautiful, interactive, force-directed network diagram directly in the browser.
Generating Insights from the KG
This graph isn’t just for pretty visualizations—it actively feeds back into the writing process.
Before the Journalist drafts a new story, a preprocessing script compares the raw abstract of the new paper against the existing knowledge_graph.json. For example, if the new paper mentions “mTOR”, the system queries the graph for all historical context, previous discoveries, and relationships associated with that node.
This historical context is injected into the Writer’s prompt. The Writer then synthesizes this data and outputs a dedicated paragraph titled “Insights from Longevity KG”.
This creates a enhanced reading experience. If a new paper discusses a breakthrough in Senolytics, the newsletter doesn’t just summarize the paper—it explicitly connects it to an article published three months ago, pointing out how this new discovery contradicts or reinforces the previous findings. The newsletter evolves from a simple aggregator into a continuous, longitudinal narrative of scientific progress.
A New Paradigm for Reading: Exploring the Graph
Perhaps the most powerful feature of the Knowledge Graph is how it transforms the user experience. You don’t have to read Bio By AI chronologically like a traditional newsletter. You can read the publication entirely by exploring the graph.
Every node and edge in the Graph stores the exact URLs (paper_slugs) of the articles that established those relationships. A reader can open the Knowledge Graph page, visually identify a massive, highly-connected cluster around “Autophagy”, and click directly on the connecting lines to jump straight into the historical stories that explain those pathways.
It can help researchers conduct non-linear, exploratory research, visually identifying the most critical focal points in longevity science and tracing the literature back to its source.
The Feedback Loop: The Rating System
An autonomous system is only as good as the value it provides to its audience. To ensure the agents are generating high-quality content, I implemented a rating system to collect feedback from readers.
Every story has comes with a 5-Star Rating Scale. Built on top of a NoSQL database, the rating system allows users to simply click a star to instantly record their feedback.
Deployment and Automation
A truly autonomous system must run entirely without human intervention. To achieve this, I leveraged GitHub Actions as a serverless execution environment, deploying the final rendered static site directly to GitHub Pages.
The deployment pipeline is governed by a YAML workflow file configured to run weekly Every Friday at 5:30 PM.
However, building an autonomous deployment pipeline that retains historical memory is fraught with edge cases. GitHub Pages typically relies on a “build and push” mechanic that overwrites the deployment branch. If not handled carefully, the pipeline will delete all past editions every time it publishes a new one. For this I implemented a state-restoration architecture that handles, safe extraction, self-contained generation, retroactive patching, and autonomous deployment using Github Actions.
The result is a flawlessly automated CI/CD loop. The system wakes up every Friday, downloads its own memories, learns new information, generates fresh content, links it all together, publishes the updated site, and goes back to sleep.
Conclusion
This project serves as a blueprint for automation of scientific knowledge and information curation, and presentation in a compeling format. By moving away from single-shot LLM queries and embracing a multi-agent architecture equipped with long-term memory structures (Knowledge Graphs), interactive reader feedback loops, and robust CI/CD pipelines,AI systems can act as true autonomous researchers and publishers. This is not intended to be a product but just sets the baseline. Conceivably, a fully featured product could be built with rich functionality and a much wider scope.
These systems won’t just regurgitate facts; they would build context, identify trends, apply critical analysis, and write stories autonomously. As the volume of scientific literature continues to grow exponentially, autonomous agentic swarms like this will become not just a novelty, but an absolute necessity to keep pace with exploding discoveries.