Regular guy working toward a computer vision engineer role. I care about two things: models that hold up on real data, and explanations that survive someone checking them.

• I love Linux and the terminal, and machine learning, especially computer vision, is where I want to build my career.
• Linux and terminals: my favourite environment, and I’m working toward making it my daily setup
• Medical imaging: 3D deep learning on chest CT scans (LIDC-IDRI)
• Tiny object detection: objects under 16 px in aerial images
• Explainable AI: Grad-CAM and occlusion maps, scored against ground truth
• Honest evaluation: multi-seed runs, bootstrap confidence intervals, written reality checks

2026-lung-nodule-detection

Finds nodules that 3 of 4 radiologists agreed on in full chest CT scans. A research prototype, not a medical device.

• Pipeline (FINDERS): blob detector → 3D CNN ensemble → occlusion maps
• Test CPM 0.823 (95% CI 0.767–0.868) on 135 unseen scans, with the test set opened once
• Honest weak spot: 3–6 mm nodules are found only 63% of the time at 1 false alarm per scan
• Ran a full scan end to end on an 8 GB MacBook CPU, and documented the cost

Hybrid-Explainable-AI-for-Pulmonary-Nodule-Malignancy-Classification-using-CT-Patches

A dual-branch model: a 3D ResNet-18 on the CT patch plus a tabular MLP on 15 features, fused through a learned gate. Explanations are scored on all 151 test nodules, and there is an interactive Streamlit app.

• The uncomfortable finding: a Random Forest scored AUC 0.942 with all features, but 0.666 without the radiologist ratings. Most of its score came from reading the answer key, so I reported that instead of the flattering number.
• Image-only 3D ResNet reaches about 0.90 AUC against radiologist consensus labels. The hybrid did not beat it.

literate-system

A literature-driven study of detecting objects smaller than 16 px in aerial images (AI-TOD benchmark, average object size 12.8 px).

• Surveyed 21 papers, and chose KLDet as the base paper
• Identified three failure modes the authors admit to: blurred features, occlusion, and missing context
• Proposed a hybrid: DN-FPN feature denoising used only in training (zero inference cost), plus a lightweight SCAM context module

ForgeLens

Flags suspicious payment screenshots. It raises suspicion and never confirms a payment. Only the receiver's bank or UPI app can do that.

• Currently: the PII redaction pipeline is done (Phase 0). The forgery checks are planned.
• Classical methods first: copy-move, noise inconsistency, template homography. Deep learning only as a measured baseline.
• Tests run on macOS and Ubuntu CI, and the same code runs on a CPU laptop or a GPU.

2026-Medical-Image-Morphological-Processing-Using-Brain-MRI

Erosion, dilation, opening, closing, and boundary extraction on brain MRI, with Otsu thresholding and kernel experiments.

• Languages: Python (primary), Bash, Java
• ML and CV: PyTorch, OpenCV, scikit-learn, NumPy, pandas, nibabel, pylidc
• Apps and demos: Streamlit, Gradio, Flask
• Testing: pytest, with CPU-only test suites
• Dev environment: Git and GitHub, Linux (RHCSA training), macOS terminal, SSH to GPU servers

• Verify everything. Key numbers come from scripts in the repo, and I check them against held-out data.
• Explainability over hype. A model that cannot justify its answer is not finished.
• Reality checks. Major projects include a document listing failures, limits, and trade-offs.
• Classical baselines first. Deep learning has to beat a simple method by more than random seed noise.

• ✉️ Email: me@shrinkhal.in
• 💼 LinkedIn: linkedin.com/in/shrinkhal-02761a2b0
• 🌐 Website: shrinkhal.in