I'm a software engineer. I spend most of my time on the boring-but-critical side of AI: the part where an LLM has to work with financial statements, transaction data and compliance evidence, and a real person has to be able to trust the output.
Right now I'm at DeFa by Invoicemate, building an agent that helps risk officers review payment-service providers before money moves. Before that I built AI backends for a non-profit finance platform, WhatsApp agents for a real-estate company, and MCP servers that let agents talk to QuickBooks, Salesforce and HubSpot.
- The model talks, the code decides. LLMs are great at intent, terrible at bookkeeping. I let them handle the conversation and keep every balance, state change and decision in plain, testable code.
- Humans stay in the loop. Anything that touches money gets citations, an audit trail and a person who can say no. I design for that from day one instead of bolting it on.
- Ship it to a real user early. My favourite projects are the ones somebody outside my laptop actually depends on. A desktop invoicing app I built is used daily by a handful of small businesses, which taught me more than any tutorial.
- Local-first when possible. SQLite, plain files, offline modes. Fewer moving parts, fewer 3 AM pages.
- Small, readable, done. I'd rather delete a feature than ship one I can't explain in a paragraph.
expense-tracker-skill is my take on personal finance for the agent era. A deterministic SQLite ledger with a portable Agent Skill on top, so Claude Code, Codex or Hermes can run it without ever being allowed to make up a number.
nuces-flex-MCP started as a joke ("roast my transcript") and turned into a proper MCP server for my university's student portal. Ask Claude how many classes you've bunked and it'll tell you.
quickbook-OAuth-demo and mcp-server-test are small, deliberately minimal reference implementations I keep coming back to when I need multi-tenant OAuth or a deployable MCP server done right.
CForge-mini-compiler is a compiler for a made-up C-like language, all the way from lexer to LLVM IR and NASM. University project, but the one I'm proudest of from that era.
LoLQueueAssist auto-accepts League of Legends queues and handles champion pick/ban so I stop missing games. lazy-post-detector flags LinkedIn posts with em dashes. It's a meme. It's not that deep.
Off GitHub: FBR Invoicer, an offline-first Tauri desktop app for Pakistan's digital tax invoicing, and FuzzSeer, an LLM-guided fuzzer for Solidity contracts that prioritises risky execution paths instead of brute-forcing.
- Building agentic RAG pipelines on LangGraph and Azure AI Search for financial due diligence
- Going deep on MCP and Agent Skills as the interface between models and real systems
- Thinking about what "explainable" should mean when an agent recommends a credit decision
Plus LangGraph, LangChain, the OpenAI and Gemini APIs, Auth0, and whatever the problem actually needs.