I built a company brain that asks YOU questions
By Akash Amalan, team Nexis 3.0. Solo project for Graph Hacks 2026 (WeMakeDevs x FalkorDB).
Try it: https://spine-app-teal.vercel.app (free server, so the first load can take a minute)
Code: [https://github.com/akashamalan/spine-app] ยท
Video: [https://youtu.be/tcZuIt4tGgs]
Imagine the person who owned your billing service quits. Nobody updates anything. Months later something breaks and nobody knows who to call.
A normal AI chatbot won't help here. It finds the old document and tells you the old answer, super confidently.
So I built SPINE. Most company brains wait for you to ask. SPINE notices problems on its own, then asks a real person.
What it does
- Ask it anything. Like "Who owns the billing service?" It walks through the graph one step at a time and you can watch it.
- It patrols by itself. Small agents look for things like: only one person knows this, nobody owns this, this fact is old, two documents disagree.
- Vera asks a human. When a fact looks doubtful, Vera asks the right person to confirm it.
The 3 layers (the main idea)
Everything lives in one FalkorDB graph, split into 3 layers:
- Source: what exists. People, teams, services, documents.
- Claims: what SPINE believes. Like "Priya owns billing", with a confidence number.
- Meta: what SPINE noticed and did. Findings, questions, answers.
The rule I'm proudest of: the agents can only write to layer 3. They can flag stuff, but they can never change a fact. Only a real person's answer can change what SPINE believes. We save their exact words, mark the old fact as "replaced", and delete nothing.
So an AI can't quietly rewrite the company's memory. It has to ask.
Old facts lose trust
Facts get less trustworthy over time. Ownership facts lose half their trust every 120 days. Old ones turn amber, then red, and SPINE raises a flag. You can right-click any finding to see the exact math.
What I learned about the database
Before this I thought a graph database was just a fancy picture. It's not. Questions like "who owns this?", "what breaks if this goes down?" and "who's the only person who knows this?" are about connections, and in FalkorDB they're just short Cypher queries. Doing the same thing with normal tables would be way messier.
I also learned the data matters. I made up a company and planted faults in it on purpose: an owner who left, a service with only one owner, a service with no owner, and two documents that disagree. Without those, the agents had nothing to find.
The demo
I ask "Who owns the billing service?" The graph draws the path and shows Priya Sharma as a faded ghost, because she left. Nobody replaced her.
Then Vera asks a person: "Is this still true? Priya Sharma is the primary owner of the billing service." They answer in their own words, and SPINE saves it.
I also made a Simulation window. Click "someone leaves" or "six months pass" and watch SPINE notice on its own, no refresh.
Stuff that went wrong
- My main demo question once answered about the wrong service. Lesson: finding the right fact matters as much as having the data.
- A Chrome extension caused a scary red error that wasn't my bug. Incognito hid it, which is how I figured it out.
- Deploying took a while. My site couldn't talk to my backend until I fixed CORS and set the backend link.
How I used AI
I used AI coding assistants (Claude Code and Claude), and I'm saying that openly. They helped me write code fast. The idea, the trust rule, the testing and what to cut were my decisions.
Limits
All people and data are invented. The login is demo-only (no passwords), though the server still blocks viewers from admin things. It's tested on one fake company only.
Next
Slack and GitHub connections, real login, and Vera asking people where they already work.
Built for Graph Hacks 2026, Track 03: Company Brain.
#falkordb #graphdatabase #ai #hackathon
Top comments (0)