8+ years on production backends. I own multi-tenant SaaS architecture from database schema to deployment, with a recent focus on production AI: a provider-agnostic LLM layer, RAG, and failure containment. Backend Lead Engineer at MicroAgility Services, Jan 2024 – Present. Based in Islamabad (UTC+5), with working hours aligned to EST/PST for US and Canada remote teams.
Important
Case study · Designing for AI Failure How the Talentnix AI gateway kept answering through live LLM outages: Retry across OpenAI, Gemini and Anthropic → RAG Fallback → Rule-Based Floor. Result: zero AI failures during live LLM outages.
| Result | System |
|---|---|
| Dashboard load from 12s to under 2s (83%) | PostgreSQL and MongoDB query tuning, new indexes, MeiliSearch rollout |
| Cut AI integration complexity by 60% | Talentnix AI gateway: OpenAI, Gemini and Anthropic behind one routing layer with automatic failover |
| Zero AI failures during live LLM outages | Retry → RAG Fallback → Rule-Based Floor |
| No cross-tenant data exposure incidents to date | Tenant isolation across REST and GraphQL APIs with OAuth 2.0/JWT |
| APIs under a second while thousands of async jobs run concurrently | Event-driven BullMQ and Redis job pipeline |
| Mentoring three associate backend developers | Learning & Development programme |
One agent: plan, act, observe, reflect. Each pass ends in Reflect: loop again with what was learned, or stop and respond. Illustrative run from the internal ATS (Claude Code + Gemini 2.5 Flash).
A team of agents on one bus. Planner, Retriever (RAG), Tool runner (MCP) and Critic exchange tasks through BullMQ and Redis until the Critic approves.
The fallback chain behind every agent call. Two retries per provider, a circuit breaker opens, the call moves on. With every LLM provider down, the RAG Fallback answers; the Rule-Based Floor answers last.
Personal builds, public and verifiable. Client systems are under NDA, so they appear only as architecture.
| Project | What it does | Link |
|---|---|---|
| Omni.io | Multi-tenant RAG support engine with a three-tier resilience ladder, Postgres row-level-security isolation, an embeddable widget and an MCP server. | Private repo |
| TalntFlow AI | Lightweight ATS with a Kanban pipeline, STAR scorecards, hiring analytics and an approval-based recruiting agent on Claude with a rules fallback. | Live ↗ · Source |
| Scanwise | Turns scans, PDFs and Office files into one cited summary and a source-grounded chat, with on-device OCR. | Live ↗ · Source |
| Catchbox | Human-in-the-loop safety net for web scrapers: failed scrapes land in an inbox, a person fixes them, and only approved records count as data. | Live ↗ · Source |
| Scrapefix | The first build of the scraper-triage dashboard: fix queue, record review and a personal ingest key for any scraper. | Source |
| Taskmaster Agent | Hackathon entry: an event-driven HR operations agent on Genkit and Gemini over Pub/Sub, with a manager-approval pause when it finds a policy violation. | Source |
| GitHub Glow-Up | Workspace for repository documentation: README drafts, code reports that separate verified facts from AI interpretation, and publishing to GitHub. | Live ↗ · Source |
| Verdict | Tamper-evident decision engine that issues hash-chained, signed reasoning receipts for EU AI Act record-keeping. | Live ↗ · Source |
| Lens Studio | Summarise, extract and compare across PDFs, DOCX and images on Gemini, with deterministic fallbacks. | Live ↗ · Source |
| Voice Studio | Real-time voice analyst on the Gemini Live API that answers from your own documents. | Live ↗ · Source |
| Shift | Semantic diffing for contracts and configs that ranks changes by materiality and explains business impact. | Live ↗ · Source |
| Cover Studio | Browser-based LinkedIn banner designer with an AI copywriting assistant, safe-zone guides and high-resolution export. | Live ↗ · Source |
| Qeloma Studio | In-browser photo and video editor on WebGPU, WebGL2 and WebAssembly, with local ONNX background removal. | Source |
| Qeloma Recruit | Angular ATS sandbox on Gemini: JD generator, STAR interview guides, scorecards and a five-stage pipeline. | Source |
| Extension suite | Ten single-purpose Manifest V3 browser extensions built with WXT, local-first. | Source |
Backend -
Data -
AI -
Architecture -
Cloud & Delivery -
- Backend Lead Engineer, MicroAgility Services (Pvt) Ltd. · Jan 2024 – Present · Talentnix (multi-tenant SaaS), APAC Management System (HRMS)
- Backend Software Engineer, MicroAgility Services (Pvt) Ltd. · Mar 2022 – Jan 2024 · iAgility and AgileBrains; 99.9% uptime with zero-downtime schema migrations on live multi-tenant databases
- Software Engineer, Market Icon · Aug 2018 – Mar 2022 · Distributed MEAN system on OpenStack, serverless AWS Lambda services
74 Verified Credentials · LinkedIn Learning 23 · Google Cloud Skills Boost 20 · IBM 14 · Anthropic 7 · Microsoft Learning 6 · 5 others → full list · in this repo
| Credential | Issuer | Completed |
|---|---|---|
| Claude Code in Action | Anthropic | Aug 2026 |
| Containers & Kubernetes Essentials | IBM | Sep 2026 |
| Gen AI: Beyond the Chatbot | Aug 2026 | |
| Introduction to generative AI and agents | Microsoft Learning | Sep 2026 |
Animations are plain SVG (CSS + SMIL), follow your GitHub light/dark theme, and stop when your system asks for reduced motion.