AI Documentation Generator (DocAi)

An AI-powered documentation agent designed to uncover and organize buried information, transforming scattered application data into clear, structured documentation.

Comment

Inspiration

Documentation is one of the most time-consuming, overlooked parts of any technical project. Researchers spend hours reformatting PDFs, developers write boilerplate docs from scratch, and students struggle to summarize dense papers into structured reports.

We asked a simple question: what if AI could look at any file — a PDF, a spreadsheet, an image — and instantly generate a professional, structured document from it?

That question became DOZAi — an AI documentation generation platform that transforms any uploaded file into a polished, structured output in seconds.

What It Does

DOZAi is a full-stack AI-powered documentation platform. Users upload any file (PDF, CSV, image, or text), and DOZAi:

  • Analyses the file using Google Gemini 2.0 Flash — extracting key themes, data, structure, and context
  • Generates a professional document tailored to the file content — complete with headings, summaries, and structured sections
  • Visualises the data using a custom Python Flask visualization engine with 50+ chart types
  • Stores and manages every generated document in a personal library, secured with JWT authentication and IP-based security alerts

The math behind our AI document scoring uses a weighted relevance model:

$$\text{DocumentScore} = \sum_{i=1}^{n} w_i \cdot \text{sim}(q, d_i)$$

where \( w_i \) represents the contextual weight of section \( i \) and \( \text{sim}(q, d_i) \) is the cosine similarity between the query embedding and document section embedding.

How We Built It

Frontend — React + Vite · Deployed on Vercel

  • Dynamic file upload with real-time AI analysis feedback
  • Document library with search and preview
  • Responsive design with custom DOZAi branding

Backend — Spring Boot · Huawei Cloud ECS

  • REST API with Spring Security + JWT authentication
  • IP verification interceptor for non-blocking security alerting
  • HikariCP connection pool to Supabase PostgreSQL
  • Systemd-managed service with structured logging

AI Engine — Google Gemini 2.0 Flash

  • Binary file analysis — files sent as base64 to Gemini vision API
  • Structured prompt engineering for consistent document output
  • Graceful fallback handling when AI analysis is unavailable

Visualization Engine — Python Flask + Matplotlib

  • 50+ chart types across 8 renderer modules: statistical · ml_ai · biology · engineering · business · network · physics · software
  • Custom DOZAi brand palette with professional typography at 150 DPI
  • Drop shadows, brand watermarks, and annotation helpers

Database — Supabase · PostgreSQL

  • Row-level security via app.current_user_id session config
  • IP verification tracking table
  • Generated file metadata and user document storage

Challenges We Faced

IP Verification Blocking Production Requests

Our IpVerificationInterceptor was returning HTTP 403 for any unverified IP, blocking legitimate users on every new network.

Fix: Redesigned as a non-blocking alert system — the interceptor always returns true, fires a background email notification, and auto-verifies the IP immediately. Zero friction for users, full security awareness for the account owner.

// Before — hard block
response.setStatus(HttpServletResponse.SC_FORBIDDEN);
return false;

// After — fire-and-forget alert, always allow
ipVerificationService.sendAlertEmailAsync(userId, ipAddress);
return true;

Gemini API 429 Quota Exhaustion

Our API key lived in a Google Cloud project (DOzAi) that was not linked to our billing account (Dozai-hosting). Large PDFs (~1.8 MB) exhausted the free-tier quota instantly.

Fix: Linked the correct billing account to the correct project. File analysis went from failing to completing in < 17 seconds.

Sending a 1.8 MB PDF to an LLM

Gemini's vision API accepts base64-encoded binary files. Getting Spring Boot's RestTemplate to correctly encode, POST, and parse a multi-megabyte PDF — with proper timeout handling — required careful tuning of the HTTP client pipeline.

Cross-Origin Authentication in Production

Getting CORS, JWT, and Spring Security to cooperate across Vercel → Huawei Cloud ECS required precise configuration of allowed origins, preflight handling, and token extraction order in the filter chain.

What We Learned

  • How to build a production-grade Spring Boot deployment on a cloud VM with systemd, structured logging, and health monitoring via /actuator/health
  • How to use Gemini's multimodal API for binary file understanding — not just text prompts
  • How to design a security system that informs without blocking — balancing UX and safety
  • How to build a Python visualization microservice with 50+ chart types that integrates cleanly with a Java backend
  • The real cost of not linking a billing account to the right Google Cloud project

Built With

Layer Technologies
Frontend React Vite JavaScript CSS
Backend Java Spring Boot Spring Security Maven
AI Google Gemini 2.0 Flash
Visualization Python Flask Matplotlib NumPy SciPy
Database Supabase PostgreSQL
Infrastructure Huawei Cloud ECS Vercel systemd
Auth JWT HikariCP iText PDF

Try It Out

Built With

  • css
  • frontend:-html
  • gemini
  • genai
  • javascript-backend:-python
  • sdk
Share this project:

Updates