FracAdapt is a full-stack fleet intelligence platform for mission planning, vehicle readiness, and terrain-aware maintenance. It gives operators a dashboard to monitor military vehicles across bases, track engine, suspension, brake, and transmission health, review maintenance history, and identify which assets are ready, at risk, or nearing service thresholds before deployment.

For route planning, the app connects interactive mapping with real-world data. It uses OpenRouteService for routing, OpenTopography SRTM elevation for terrain profiles, Open-Meteo for weather, and physics-based calculations for slope, roughness, traction, traversability, fuel use, cargo impact, route risk, and predicted component stress across changing conditions.

Its ML service adds a Python and PyTorch RCAN super-resolution pipeline that improves 30m DEM inputs toward 10m analysis while preserving elevation, slope, and curvature. Together, the React frontend, Node/Express backend, SQLite database, security layer, and deployment tooling turn geospatial and mechanical data into clear readiness signals and maintenance priorities.

Managing appointment bookings manually is slow, error-prone, and pulls staff away from higher-value work. AI Scheduler automates the booking conversation through SMS: staff set availability and trigger outreach, then a tool-calling AI agent handles confirmation, rescheduling requests, waitlisting, and escalation, only looping in a human when the situation genuinely requires it.

When a slot opens, the system re-engages the waitlist and slots in the next best-fit client using LLM-ranked matching. Clients can also self-book through a branded booking portal with OTP authentication, while staff monitor live conversations through timeline and calendar views and can take over instantly.

I Skillora as a local-first AI platform for organizing the job search from discovery to application. The system pulls listings from seven sources, including LinkedIn, Indeed, Glassdoor, Google Jobs, ZipRecruiter, Himalayas, and Arbeitnow, using Playwright automation and local LLM relevance scoring against a candidate profile.

At its core, a YAML-driven resume engine turns one master resume into tailored versions for each role through per-job overrides. I also implemented persistent LLM memory so the platform can remember candidate facts across sessions and provide more useful coaching without repeated prompts.

The React job board supports save, dismiss, apply, source toggles, salary, experience, date filters, and CSV export. A Flask and Ollama coaching chatbot gives section-by-section resume feedback. The result is a private AI job search workflow with no SaaS dependency or third-party data sharing.

I built Codexa as a local, GPU-accelerated coding assistant designed for developers who want Claude Code-style workflows without API costs or external data sharing. The system runs Qwen3-Coder-30B through vLLM on DGX Spark hardware, reaching about 48 tokens per second on NVIDIA Blackwell GPU architecture.

The agent supports five working modes: default, explore, plan, review, and edit. I designed a three-tier permission system that controls when the agent can read, suggest, or modify files, with approval levels for automatic actions, one-time confirmations, and always-confirm workflows.

For safer code changes, I implemented an auto-planning gate that detects multi-file edit requests, creates a structured plan, asks for approval, and backs up files before changes are applied. The system also includes session audit logs, tool-call tracking, GPU memory capping, and multi-model support.

I built Lendly as a full-stack loan document matching platform for helping loan officers process lender program PDFs, extract key details, and turn complex lending documents into searchable knowledge. The system supports secure authentication, role-based access, PDF upload, metadata entry, file validation, status tracking, and document reprocessing so officers and admins can manage program documents with more confidence.

Uploaded files are converted into text chunks, enriched with embeddings, and organized through a searchable document dashboard with filters, pagination, bulk actions, and metadata controls. I also built a document detail workflow for content review, chunk editing, and processing control.

The backend uses Express, Prisma, PostgreSQL, and a local Llama AI server for retrieval augmented generation. The frontend is built with Next.js, providing responsive dashboards, profile controls, and secure API workflows for modern loan operations.