// projects
Technical case studies about the decisions, constraints, and lessons behind applied AI, local systems, DevOps, backend, and robotics work.
A local-first, multi-agent RAG pipeline for the first-pass desk assessment of EFQM-based organizational self-assessment reports. It retrieves assessment guidance, links findings to source evidence, flags gaps and conflicts, and produces structured draft feedback for human review.
A private operations dashboard that unifies project health, GitHub work, dashboard tasks, bounded Codex telemetry, alerts, and analytics. The hosted application runs on Vercel with Supabase, while an authenticated outbound Windows service provides permitted local summaries without exposing the workstation or sync credentials to the browser.
A local-first personal AI system designed around source-backed memory, persistent agent identities, policy-enforced actions, and controlled recovery. Milestone 1 is complete with encrypted backups, attended restore, integrity checks, and scheduled execution; the FastAPI kernel, searchable memory, agent runtime, and mobile PWA are the approved next milestones.
An internal on-device LLM routing layer that classifies each prompt by task type and complexity before selecting a local model or cloud provider. My operational estimate was that roughly 65 percent of requests stayed local, reducing cloud AI spend by about 45 percent.
Built and maintained CI/CD pipelines for Dockerized C++ and Python backend services on embedded Linux across ARM and x86 targets. I taught myself Rust on the job and implemented a REST-based OTA update mechanism on three test vehicles with automated version checks, fallback logic, health monitoring, and service orchestration.
Real-time object detection prototype using GStreamer and YOLO on embedded ARM and x86 hardware. The modular pipeline separates preprocessing, inference, post-processing, and visualization, then produces visual overlays and structured detection metadata for downstream feedback experiments.
Researched and benchmarked more than ten LLM approaches, AI tools, and hardware options for an early-stage internal evaluation. I tested selected options, documented evidence and uncertainty, and prepared decision-focused reports.
ML-based mood tracker with a Random Forest model that reached about 87 percent accuracy after feature engineering and SMOTE class balancing. A FastAPI backend runs the preprocessing and prediction path, while Flutter, SQLite, and Firebase support the application.
A booking, membership, payment, attendance, and operations platform deployed for one operating yoga studio. I owned most of the Flutter member and coach app plus its API integration, contributed to the React admin client, and coordinated with an ASP.NET Core and MySQL backend secured by Identity, JWT, and five-role access control.
AI-driven robotics prototype combining computer vision and sensor fusion. YOLO for real-time visual recognition, LiDAR for obstacle avoidance and mapping, custom path-planning logic using detected objects as waypoints.
Retail automation prototype for autonomous product handling. YOLO-based fruit recognition on a conveyor/shelf, grasp-point detection for robotic-arm manipulation, closed-loop camera-driven automation.
Dual-interface logistics app with separate driver and admin views, real-time task management via Firebase, role-based authentication, Material Design UI.