A GitLab-native autonomous DevSecOps agent that converts dependency vulnerability reports into validated, human-review merge requests — no external hosting, no webhooks.

• Designed a multi-stage remediation pipeline: GitLab dependency scan → Claude CVE analysis → Pydantic validation → file patch → GitLab MR creation
• Implemented patch safety guardrails enforcing semver downgrade blocking, file allowlist checks, regex-validated semver, and one retry on malformed AI output before fail-fast
• Separated AI reasoning from application logic via a structured JSON output contract: Claude produces a RemediationModel that is validated before any file is touched
• Configured GitLab Duo Agent and Flow YAML definitions for native platform integration, eliminating external hosting requirements

Architecture: Claude produces a structured RemediationModel (Pydantic) → patch_validator.py enforces semver safety rules → gitlab_client.py creates a remediation branch, commits the patched package.json, and opens an MR with full security context for human review.

A real-time physically-inspired black hole simulation running at 60fps in the browser, built with Three.js and a custom GLSL raymarching pipeline.

• Implemented an Euler-integrated GLSL raymarcher with modified Newtonian gravity (a = −M·r / (|r|² + ε)^1.5) and a softening parameter to prevent singularity precision errors
• Applied a plane-crossing disk intersection heuristic (Y-sign flip detection per step) instead of volumetric raymarching, enabling large step sizes and sustained 60fps
• Modeled relativistic Doppler beaming per fragment: computes Lorentz factor γ and Doppler shift D = 1/γ(1−β·cosθ), raised to 3.5× for asymmetric accretion disk brightness
• Implemented Keplerian shear on disk turbulence using anisotropic 3D fBm noise with radially decaying angular velocity (ω ∝ r⁻¹·⁵) and blackbody temperature-to-RGB mapping (inner ~6000K, outer ~1500K)

Architecture: Three.js renders a 200,000-star Points layer as a background texture. A fullscreen ShaderMaterial raymarches each pixel: rays curl through the gravity well, pierce or miss the accretion disk via plane-crossing, and escaped rays reproject their final bent velocity vector to UV space to fetch the lensed starfield texture.

A full-stack financial analytics dashboard for loan risk visualization, integrating a FastAPI backend with an XGBoost credit risk model exposed as a REST endpoint.

• Built a FastAPI backend exposing 10 REST endpoints for customers, loans, risk reporting, branch performance, and authentication — backed by PostgreSQL
• Integrated an XGBoost credit risk model with SMOTE class imbalance correction, serializing the trained model and feature encoder for inference at /ml/predict-risk
• Designed dual risk scoring: a rule-based DB-computed endpoint (/compute_risk/{id}) alongside ML inference, maintaining clean separation of heuristic and model-driven scoring
• Containerized the backend and PostgreSQL service with Docker Compose for single-command local deployment

Architecture: React (Vite + shadcn/ui) → Axios → FastAPI routers → either PostgreSQL (via async query) or the serialized XGBoost model. Auth flows through bcrypt-hashed passwords; risk scoring routes branch between rule-based aggregation and ML inference independently.