I build agentic systems
that survive contact with production.

Desktop AI learning platform with multi-user auth, RAG over PDFs and web, and persistent memory threads.

• Dual-vector retrieval (FAISS + ChromaDB)
• Multi-user concurrent sessions
• Memory-aware threads

1. Built secure authentication and conversation management for multiple concurrent users.
2. Designed a RAG pipeline over PDFs and web resources using FAISS + ChromaDB with contextual retrieval.
3. Shipped memory-aware chat — persistent threads, session tracking, conversation merging — plus an admin dashboard.

End-to-end recruitment platform — job posting through offer — with LLM-driven screening, matching, scheduling, and email.

• 4 third-party API integrations
• Subscription billing live
• Multi-tenant data isolation

1. Automated resume screening, candidate matching, interview scheduling, and email generation using LLMs, RAG, and agentic workflows.
2. Integrated Google Calendar, Gmail, Slack, and Stripe for scheduling, communication, team notifications, and subscription billing.
3. Stack: React, FastAPI, PostgreSQL, Redis, Docker, JWT.

A single FastMCP server that exposes 53 tools for GitHub repository management and LinkedIn posting, so an LLM agent can ship a PR and announce it on your feed in one workflow.

• 53 tools · 26 prompts · 4 resources
• Long-running task support
• 30s cache + rate-limit guard

1. Unified GitHub (43 tools — repos, branches, commits, files, issues, PRs) and LinkedIn (10 tools — text, image, multi-image, video, document, article, poll) under one FastMCP server with auto-discovery from app/.
2. Annotated every tool with MCP ToolAnnotations (readOnlyHint, destructiveHint, idempotentHint) so clients prompt for confirmation appropriately; gated read traffic through a 30s cache and a 10 rps / 20 burst rate limiter.
3. Long-running task support via fastmcp-tasks for video and multi-image uploads; tools return {success} / {error} dicts instead of raising, so agents dispatch on shape rather than catching exceptions.

Privacy-first AI assistant. Zero external API dependency. Fully local multi-modal with human-in-the-loop approvals.

• Fully local — zero cloud calls
• Human-in-the-loop approvals
• Vision + tool calling
• Persistent long-term memory

1. Built with FastAPI, LangGraph, Ollama, and SQLite — real-time streaming chat with no cloud calls.
2. Designed persistent SQLite + JSON long-term memory for coherent context across sessions.
3. Added vision input, tool calling, dynamic model orchestration, and explicit HITL checkpoints for sensitive actions — fully local multi-modal deployment.