Math Captcha Forge
MobileNetV3-based mathematical CAPTCHA solver. It brings data preparation, hard-example mining, ONNX export, and a low-latency FastAPI inference chain together in the same product flow.
Overview
The goal was to turn math-based CAPTCHAs not into something merely solved correctly, but into an ML pipeline that is exportable, scalable, and practically serviceable. The challenge wasn't just training a model; it was building a single chain from training all the way to serving.
I split the pipeline into data cleaning, hard-example mining, model training, quantization, and API serving layers. I used a lightweight backbone with MobileNetV3 and brought inference cost down through ONNX export and INT8 optimization. At the end of the line, I made it easy to integrate via a FastAPI-based prediction service.
What came out was not just a model but a reusable application: the dataset, the training logic, the optimized export, and the serving layer all gathered under the same project. That turned the project into a real engineering case study rather than an academic demo.
MobileNetV3-based mathematical CAPTCHA solver. It unifies training, export, and the inference pipeline into a single case study.
Highlights
Technical Architecture
FinGuard AI
Auditable, AI-powered compliance assistant that manages banking and HR compliance processes with a multi-agent RAG architecture.
Compliance Assistant