Project Concept & Core Goal: Built ALIBI GRAPH, a solo-developed multi-agent cyber-forensics tool designed to catch liars by automatically cross-examining suspect alibis against surveillance evidence and interrogation logs.
Building a hackathon project completely solo is always a chaotic rush,
Why I Made It
I wanted to solve the absolute nightmare of manual forensic analysis. When detectives have to sift through multi-page interrogation notes, vehicle tracking logs, and security footage, it takes hours to spot a lie. I wanted an AI app that could ingest these raw case files, map out the relationships, and instantly flag when a suspect's alibi falls apart.
How I Put It Together (The Real Tech Stack)
The Frontend: I used Streamlit to build out the interface because it let me spin up custom pages, file upload bays, and live terminal views super fast.
The Multi-Agent Pipeline: I split the logic into two main agents:
Agent 1 handles the dirty work of ingesting raw text evidence files, pulling out suspect names, and parsing out semantic topics.
Agent 2 acts as the cross-examiner, looking at the suspect's claims and matching them against physical evidence and surveillance timestamps.
The Database & API Setup: I used FalkorDB Cloud as the graph memory layer. Instead of dumping everything into a flat database where context gets lost, I needed a way to store suspects, evidence exhibits, and claims as nodes so I could run Cypher queries and map everything out in real-time.
The Real Struggles (Developer API Problems & Late-Night Coding)
Let's be honest—dealing with developer API problems, handling rate limits, and getting multi-agent loops to talk smoothly to a cloud database completely by myself under hackathon pressure was brutal. I ran into plenty of annoying connection drops, local server hiccups, and sudden Streamlit crashes right when I was trying to record. Trying to format Cypher queries dynamically while an agent is streaming data and wrestling with cloud configuration can drive you crazy at 3 AM.
But seeing it all finally click—dropping in a file for a suspect like Jordan or Morgan(Imaginary Entities considered only for projects), watching the live graph nodes stream in, and seeing those red contradiction edges light up on the visual canvas—made all the debugging completely worth it.
If you're building agentic apps with complex relational data as a solo dev, graph databases are awesome once you tame them. Check out my GitHub and demo video to see how it turned out!
Top comments (0)