Back
Year
2026
Tech Stack
Python, PyTorch, OpenCV, Pillow, NumPy
Description
Reverse-engineered an undocumented, pre-trained Pix2Pix GAN from raw
checkpoint weights with no available source code or documentation, then built a
consultation tool around the reconstructed model. The system combines the
reconstructed inference pipeline with patient records, before/after visual
simulation, versioned revisions, and PDF export.
Why This Project Matters
The project required reconstructing the model architecture and
inference pipeline from the checkpoint weights before the application layer could
be developed. It demonstrates the ability to investigate an undocumented
machine-learning system, reproduce its inference process, and integrate the
resulting model into a functional tool rather than stopping at a research or
notebook prototype.
Technical Highlights
- Achieved 1.37s mean CPU inference time during model inference
- Recorded 523MB peak memory usage during CPU inference
- Built the inference workflow with zero GPU dependency
- Reconstructed the full Pix2Pix GAN architecture and inference pipeline from raw checkpoint weights alone, without source code or documentation
Key Features
- Patient record management
- Before/after visual simulation
- Versioned revision management
- CPU-based model inference
- PDF export for consultation outputs
My Role
- 🧩 Worked as the sole developer from model reverse-engineering through deployment
- 🔍 Reconstructed the Pix2Pix GAN architecture from raw checkpoint weights without source code or documentation
- 🖼️ Built the CPU-only inference pipeline and image preprocessing workflow
- 🗂️ Designed the patient record and version-history data model
- 📄 Built the PDF export and doctor-revision workflow, while benchmarking CPU inference performance