Folio

Inspiration

When companies try to deploy AI agents, they know nothing. Consultants are paid millions to try and make companies AI-native, yet often fail. We believe this is because of the lack of organisation of data. With correct organisation and formatting of the data, it would be possible for AI agents to be so much more effective. Companies data is scattered across their file system, in the cloud, in email inboxes. OpenClaw deploys agents in a smart way to adapt to the task at hand, and Moltbook creates a forum of agents all contributing to each other. Using these ideas to streamline the direction of our agents, and keeping the human in the loop, they organise the filesystem into a structure that is optimised for companies to deploy AI on top of. We liked the way Obsidian stores the notes, so we decided to create a graph visualisation tool to help the customer understand what is happening to the data.

What it does

Transforms messy, structured data from different sources into a verified, hierarchical knowledge base that is easy for agents to navigate. We split it up into 'phases':

│ PHASE 1: ONBOARD
│
│ Company details → Connect data sources → Set goals │ (Google Drive, OneDrive, Gmail, Outlook, etc.)

▼

│ PHASE 2: EXPLORE
│
│ AI agents connect to sources → Crawl & discover
│ → Data visualization (volume, types, character)
│ → Agents consult the Forum for how-to guides

▼

│ PHASE 3: STRUCTURE
│
│ AI builds hierarchical KB with READMEs
│ → Progressive disclosure (overview → detail)
│ → Cross-references between domains
│ → Agents write new guides to the Forum

                   │
                   ▼

│ PHASE 4: VERIFY
│
│ AI surfaces ambiguities & contradictions
│ → Human answers structured questionnaire
│ → KB gets corrected and finalized
│ → Result: verified, navigable knowledge base

                   │
                   ▼

│ PHASE 5: USE
│
│ Any AI agent can now navigate the KB precisely
│ → Build domain-specific agents on top (email,
│ client comms, compliance, reporting)
│ → Agents stay informed by the company's context

And at the end of it we get a data infrastructure that agents can navigate. At frequent intervals, new data is fetched, and updated into this knowledge base. A D3 graph updates in real time (with convex, a dynamic database that is built into the software itself).

How we built it

Agents: A master agent coordinates four sub agents : explorer, structurer, verifier, knowledge writer), through the pipeline. LLM responsibilities are split deliberately, Claude handles the orchestration, reasoning, tree design (of knowledge base), and questionnaire generation. Gemini handles multimodal extraction from PDFs, images, and spreadsheets. We use 5 design patterns: •⁠ ⁠capabilities focused prompts (describe only what you can see) •⁠ ⁠mechanical loop detection (a hash based sliding window that catches stuck agents) •⁠ ⁠agent forum (a moltbook-style communication between ages, each new engagement with a thing updates the agent's internally stored paths) •⁠ ⁠verification questionnaire (we want to keep the human in the loop with hard decisions that could affect them rather than just making assumptions, e.g. if there is a conflict in files) -navigable knowledge tree (hierarchal structure, which beats RAG for this use case)

UI was built with lovable. Icons were iterated over by hand.

Graphs were created Obsidian-style with D3, mirroring the styling from the UI.

Challenges we ran into

Getting agents to use pre-configured interactions were harder than expected. Often they would try things from scratch again rather than use tools already available to them. The specifications for what their new 'skills' were had to be extremely precise. Seeding data at different stages was tricky for the demo - the agent pipeline and the UI / graph base needed mock data, which is hard to represent a whole company that has been running for 30 years. D3 force simulation posed a challenge as it had to be adjusted to not look too gimmicky, adjusting the gravitational effects. Agents can't subscribe to retroactive queries, so the verification phase uses an exponential backoff polling loop waiting for human responses. We focused on accounting as a niche to first deploy our product in. But we do not have domain expertise in this. When we are using agents to build up this forum and help them evolve, we need to be careful not to do so using specific company data, as this could be a huge security concern.

Accomplishments that we're proud of

We like the idea of keeping the humans in the loop. Most AI platforms auto-guess and often make assumptions the humans don't want. We do not want that, as we are potentially dealing with sensitive company data. Asking a human deals with the ambiguities the AI cannot solve itself. Operational guides written on the forum are a good way to ensure that the system continues to improve itself and evolve new capabilities and potential emergent behaviour. Real time graph visualisation is really cool - as soon as new data is added, it is populated in the original data graph instantly. Then as soon as the pipeline of agents is run, the changes are reflected instantly in the UI.

What we learned

RAG is broken for complex domains. Similarity search returns hundereds of fragments of data that could be completely separated in the file system, and lack context. The chunking destroys the hierarchical meaning that is often so important for these systems. We help to optimise the agents precision. It is really important to keep the human in the loop or the agents could behave unexpectedly at tines. When we listed what was forbidden, the agents were unimaginative. So we had to make sure to make sure we pay attention to what is available.

What's next for Folio

Expanding to other areas other than accounting: such as Lawyers, doctors Hopefully we can make it smarter about conflict resolution. E.g. if there is the same data 20 years ago and 1 month ago, then it is very likely that the one month ago is correct. We could make a MCP to help external third party agents plug into Folio's knowledge base, and let them contribute to the forum as well. Pennylane is specific accounting software that we could integrate. The more sources, the bigger the knowledge base. We ideally want to be the infrastructure layer for enterprises moving towards agentic workflows.