The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
I built DocuMind — a financial document automation platform for Chartered Accountants
One thing I noticed while working on financial-document workflows was that a lot of a CA's time can go into repetitive document processing rather than actual professional decision-making.
So I started building DocuMind — a tool designed to automate the repetitive parts of working with financial documents.
💡 The idea
A CA uploads a financial document such as a bank statement, and DocuMind handles the mechanical processing:
Document → Extraction → Normalization → Transaction Processing → Rules → Validation → Structured Output → Excel
Instead of manually going through large documents and preparing the same data repeatedly, the system converts the information into structured, usable outputs.
What I built:
- PDF/document extraction
- OCR for scanned documents
- Transaction extraction and normalization
- Python-based rule processing engine
- Data validation and reconciliation
- Automated Excel generation
- Database-backed document management
- Authentication and user-level document access
- RAG/AI assistant architecture for document Q&A
- Local AI support through Ollama for privacy-sensitive workflows
- React frontend + FastAPI backend
Architecture:
React Frontend
↓
FastAPI Backend
↓
Document Processing Orchestrator
↓
Extraction / OCR
↓
Normalization
↓
Transaction Processing
↓
Rule Manager
↓
Validation & Reconciliation
↓
Structured Results
↓
Excel / CA-ready Output
The AI layer is intentionally not responsible for deterministic calculations.
Python handles things that should be predictable and reproducible, while AI/RAG is used where natural-language understanding and document interaction actually add value.
Why I'm building it:
I recently discussed the idea with a CA and got an important piece of feedback:
The goal shouldn't be to replace the CA's professional judgment — it should be to eliminate the repetitive work around it.
That changed the direction of the product.
I'm now refining DocuMind around workflow automation, accuracy, traceability, privacy, and useful outputs rather than trying to make AI “diagnose” financial information.
What's next:
I'm currently working on:
Improving the CA-focused workflow
Better document extraction
More robust validation
Source/page-level traceability
Better Excel outputs
Testing the workflow with real CAs
Expanding into additional financial-document workflows
Eventually building a more complete AI-powered document assistant
This project has been a great learning experience because I'm not just building features — I'm learning how to turn a real-world problem into a software architecture.
Tech Stack:
Frontend: React, Vite, Tailwind CSS
Backend: Python, FastAPI
Database: SQLite, SQLAlchemy
Data Processing: pandas, openpyxl
Documents: PyMuPDF, pypdf, Tesseract OCR, Tabula
AI/RAG: Ollama, ChromaDB, Sentence Transformers
Authentication: JWT
Deployment: Local-first architecture, with cloud deployment planned for later
I'm currently open to freelance and full-time opportunities in:
Full-Stack Development · Backend Development · Python/FastAPI · React · AI Automation · Document Processing · SaaS Development
Building products that solve boring, expensive problems is what I want to do more of. 🚀
#BuildInPublic #FullStackDevelopment #Python #FastAPI #React #AI #Automation #SaaS #Freelance #SoftwareDevelopment #ProductDevelopment
GitHub link: https://github.com/Tony-Starkb/DocuMind
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Trending
retrofuturism
Share your favorite designs from the past exploring the distant future. Or your own take on the aesthetic 🚀
halftone
Dots. Texture. Grain. Let's see your designs using the halftone technique.
Midjourney
Creatives are showing off what they're making with Midjourney.
colorgrade
Video editors sound off: How do you enhance the mood of your video in color grading?
Related posts
Before → After
Starting with an original product photo and turning it into a polished e-commerce visual through an AI-assisted workflow.
AI can speed up the production, but it doesn’t replace creative thinking. The creative direction behind the AI still matters; from composition, lighting, environment, and product placement to deciding what actually works for the brand.
These decisions come from human experience: everything we’ve learned, seen, and experienced, combined with our own creative taste and judgment. AI can execute the direction, but we’re still the ones defining it.
Original product photo → e-commerce-ready creative.
Faster production, without removing the human design decisions behind it.
0 voted
0%
1 voted
100%
1 vote
Closed
the red light makes the shot pop
I took on a creative challenge from FLORA
I found a few pieces through The Met’s Open Access collection with unique shapes and details and thought: "What if these works became handbags?"
I used the original artworks as the starting point, then reimagined their shapes, details, textures, and visual language into a series of bag concepts. I used Flora AI for the process from concepting the bag designs and generating video, to compositing the final pieces together–all in Flora's AI canvas.
This is my submission for the #FLORAMuseum challenge.
What I’ve learned after 1,500+ hours of working with AI.
AI understands cameras better than it understands “realism.” Lens, distance and lighting will get you much further than writing “ultra realistic” ten times.
A new AI tool rarely replaces an old one. Usually, it just becomes better at two or three very specific things.
One of the hardest things to ask AI is “keep everything exactly the same.” Changing the whole image is often easier than changing one tiny detail.
The more control you try to have, the fewer unexpectedly great results you get. Sometimes the best thing you can do to a prompt is delete half of it.
“Almost perfect” is the most expensive result. A bad generation gets deleted immediately. An almost perfect one can easily cost you another 20 attempts.
One image can go through four different AI tools, not because none of them worked, but because each one does one part of the job better.
After 1,500+ hours, my prompts have actually become shorter, while my briefs have become much more precise. Those are not the same thing. 1,500 hours ago, my main question was “How do I make this?”
Now it’s much more often “Is this even worth making?”
So good!