DocIntel Agent | Devpost

Inspiration

In Colombia and across Latin America, 162 million people live in poverty. Governments invest billions every year in social programs: food markets, youth employment, rural women's funds. But the only proof that resources actually reached a beneficiary is a handwritten paper record signed in a remote village with no internet.

Auditing those records is humanly impossible at scale. That gap is where corruption thrives.

We're LIFE·IN·CO, a technology company from Colombia. We asked ourselves: what if AI could be physically present at every single delivery, in real time, even where there is no connectivity? That question became DocIntel Agent.

What it does

DocIntel Agent is a multi-agent AI platform that operates at two levels:

Field App (Flutter + ML Kit, works fully offline): A mobile application used by field operators in remote areas. Google ML Kit runs on-device OCR with no internet required. Works in the Amazon, Chocó, and Cauca regions where connectivity is zero.

Multi-Agent Audit Platform (Vertex AI / Gemini + AlloyDB): Three specialized agents evaluate every record simultaneously:

Compliance Agent — record integrity
Impact Agent — quantities delivered
Anti-Fraud Agent — identity cross-validation

The platform triggers geolocated real-time alerts to regulators when anomalies are detected.

How we built it

Full Google-native stack: Flutter, ML Kit, Gemini 2.5 Flash, Gemini 2.5 Pro, AlloyDB, Vertex AI, Google Drive API, Google Sheets API, Python, FastAPI.

Pipeline: Ingestion (240ms) → OCR Flash (820ms) → Validation Pro (640ms) → QA (380ms) → Output (150ms) = 2.3 seconds end-to-end.

Challenges we ran into

Connectivity is a design constraint, not a bug. In Colombia's Amazonian and Pacific regions, internet infrastructure was never built. We designed an AI system that works first offline and syncs later.

Handwriting variability is extreme and no two field records look the same. We built a 3-layer OCR cascade that handles everything from clean printed text to difficult cursive in low-light conditions, reducing cost from $391 to under $1 per document.

Accomplishments that we're proud of

100,000+ records processed in production in 2 months for Colombia's JEP peace program
870 records processed in a single day at approximately $1 total cost
2 of Colombia's largest auditing firms as active clients
Direct impact on programs with over 8 billion COP (~$2M USD) under management

What we learned

The greatest technical challenge in AI for social impact is not the model — it's the last mile. Intelligence must work where infrastructure doesn't exist.

What's next for DocIntel Agent

Scale to +5M records by end of 2026, then LATAM expansion: Mexico (Bienestar, 30M beneficiaries), Brazil (Bolsa Família, 21M families). The same infrastructure gap exists across sub-Saharan Africa and South Asia, where 400M+ people are audited by paper records.