Engineering the
systems behind AI.

An offline-first assurance platform for checking the integrity of computer vision datasets, model artifacts and inference records. Its dataset pipeline validates YOLO and COCO annotations, detects duplicate files using SHA-256 hashes, compares labels against trusted references and flags unusual image distributions with statistical checks. It also scans image patches for low-entropy trigger patterns. For model and inference integrity, it verifies PyTorch, TorchScript and ONNX artifacts against reference manifests and uses HMAC-SHA256 signatures to detect altered or replayed inference records. FastAPI exposes the workflow, while SQLite stores audit events and generated reports provide reviewable evidence.

Stack Python, FastAPI, SQLite, NumPy, PyTorch, ONNX, HMAC-SHA256, Pytest

An unsupervised analysis workflow for identifying unusual patterns in NASA DASHlink flight telemetry. It loads Parquet flight recordings, selects configured sensor channels, handles missing values, re-references time and resamples signals into comparable representations. Standardization and PCA reduce the feature space before DBSCAN groups recordings by density, with an estimated neighbourhood radius derived from the nearest-neighbour distance curve. A separate diagnostic stage ranks sensor deviations against robust median and median absolute deviation baselines. Results are exported as CSV reports and explored through an interactive Streamlit dashboard. The scores support investigation; they are not calibrated probabilities of aircraft faults.

Stack Python, NumPy, pandas, scikit-learn, PyArrow, Plotly, Streamlit

A policy-bounded financial remediation system designed to handle payment failures that may require more than another retry. Incoming payment events are normalized and enriched with customer and transaction context before safety rules check fraud flags, late-authorization risk and intervention limits. When an event passes those checks, a Gemini-powered decision layer can select from constrained actions such as creating a payment link, proposing instalments, applying a policy-limited discount or scheduling a promise to pay. A deterministic fallback, mocked payment integration, idempotent execution and audit logging keep the workflow inspectable. A React dashboard receives live updates through FastAPI and WebSockets. The repository documents evaluation on synthetic events, not live payment recovery.

Stack Python, Gemini API, FastAPI, React, WebSockets, SQLite

An edge-oriented diabetic retinopathy screening workflow for settings where specialist access and reliable connectivity may be limited. Before classification, it assesses fundus image focus, illumination and retinal field-of-view coverage, then enhances usable images with CLAHE. A ResNet-50 model imported through ONNX classifies images across five diabetic retinopathy severity levels. The workflow pairs the result with attention visualizations and a screening report, applies a referral rule for referable cases, and stores referral records locally for later synchronization when connectivity returns. MATLAB handles image processing and inference, while Simulink models patient flow and deployment resources. It is a screening support prototype, not a replacement for clinical diagnosis.

Stack MATLAB, Simulink, ResNet-50, ONNX, OpenCV