AI Engineer building production-grade LLM, RAG, and agentic automation systems.
2+ years shipping scalable enterprise AI at Liquid Technologies and Teknloje Solutions — from LLM orchestration and voice automation to computer vision and backend architecture.
44
MCP tools in production
7
Business domains shipped
4
Calendar providers
2+
Years shipping AI
GPT-4
RAG
LangChain
Pinecone
01 · Profile
About Me
I am a production-focused AI Engineer with 2+ years of experience designing, building, and deploying intelligent systems that solve real business problems. My work spans Generative AI, Large Language Models, Retrieval-Augmented Generation, voice AI automation, computer vision, and backend architecture — from system design to scalable production deployment.
I currently work at Liquid Technologies, where I am building Sally — a two-product AI meeting-intelligence suite. Sally Web is a multi-tenant B2B SaaS platform with a production RAG knowledge base, live tool-calling, and 16+ third-party integrations, while Sally Desktop is a privacy-first Tauri app that transcribes meetings entirely on-device. Before this, I spent 18 months at Teknloje Solutions shipping five production AI systems across enterprise RAG, banking document verification, multi-camera surveillance, and voice AI automation.
I focus on systems that drive measurable business impact, not prototypes. I take full ownership from architecture through deployment, combining strong backend engineering with AI automation to build out-of-the-box solutions that are reliable, scalable, and easy to integrate.
NestJS 11, PostgreSQL + pgvector, Redis/BullMQ, Socket.IO, React 18 — principal contributor across backend and frontend.
Designed governed agent execution: every agent action is risk-classified fail-closed, high-risk writes wait for owner approval and replay from the stored payload, and the audit log is append-only at the database, keeping every agent write accountable.
Built a bidirectional Model Context Protocol integration: an MCP server exposing 44 tenant-scoped tools to Claude Desktop behind an OAuth 2.1 authorization server, plus a multi-transport MCP client governed by the same gate.
Closed the tenant-isolation gaps the product is sold on: per-person feature switches on every route, a positions-based access model rolled out in shadow mode with zero regressions, and an audit that fixed cross-tenant reads.
Cut the cost of adding integrations and AI providers: a four-provider calendar connector platform with zero frontend code per provider, per-tenant encrypted OAuth apps, and bring-your-own-key LLM routing across OpenAI, Anthropic, Gemini and vLLM.
Shipped seven business domains end to end, from a PMO cockpit over Google Sheets, Odoo and Trello to a passwordless client portal, plus a meeting-intelligence pipeline turning Recall.ai and Meetstream.ai transcripts into Jira and Asana action items.
Sally Desktop
Product
Offline-first meeting recorder for Windows, macOS and Linux (Tauri 2, Rust, Next.js) with on-device Whisper/Parakeet transcription on ONNX Runtime, GPU acceleration and meeting sync to Sally, so meetings are transcribed without audio leaving the machine.
LeadGenPro
Product
Lead-generation SaaS (NestJS, PostgreSQL, React) that turns keyword searches on LinkedIn, Reddit, Twitter/X and Hacker News into CRM-ready leads via a Playwright worker, a BullMQ LLM-qualification pipeline and GoHighLevel sync.
Upwork Analytics Platform
Product
Solo-built read-only multi-account BI platform (Next.js, NestJS, SQLite/Drizzle) with incremental OAuth sync, AES-256-GCM token encryption and a CI-enforced read-only boundary that fails the build on any Upwork mutation.
AI Engineer
Teknloje Solutions
October 2024 — April 2026
Karachi, Pakistan · 18 months
Delivered 5 production AI systems across enterprise AI, computer vision, OCR, voice automation, and medical NLP in 18 months.
Built TekInsight — an enterprise AI platform with LLM-powered chat, FAISS + LangChain RAG document intelligence, and JWT multi-tenant authentication deployed on Flask/IIS, reducing retrieval time by 60% and query latency by 40%.
Engineered TekCondor — a hybrid YOLOv8 + ResNet Siamese network signature verification pipeline at 94%+ accuracy with sub-300ms PaddleOCR cheque digitisation inference, eliminating manual review for 75% of submissions.
Built a real-time multi-camera AI surveillance system using YOLOv8 + ArcFace embeddings with cosine similarity matching and polygon zone tracking, achieving 95%+ face recognition accuracy across distributed cameras.
Designed an AI Voice Agent Platform using Twilio + Retell AI + n8n that monitors inbound emails, extracts leads via LLM parsing, and triggers AI outbound calls — cutting lead response time from hours to under 2 minutes.
Built a LLaMA 3.2 + LangChain FastAPI medical summarisation service converting doctor-patient conversations to structured clinical JSON at sub-3s latency, reducing documentation effort by 70%.
LangChainOpenAILLaMAFAISSYOLOv8PaddleOCRArcFaceTwilioRetell AIn8nFlaskFastAPIDockerPyTorchSQL Server
04 · Case Studies
Case Studies
Deep dives into the systems I have shipped — the problem each one faced, how I approached it, and the impact it had in production.
Agentic AI · MCP · Tenant Isolation
Sally Web — Governed Agent Execution in a Multi-Tenant SaaS
Liquid Technologies · 2026
Challenge
Sally's AI agents act directly inside tenants' workspaces — writing to calendars, CRMs and project tools. Every action had to be accountable, tenant-isolated, and reversible without blocking the product's speed.
Approach
Risk-classified agent execution: every action is classified fail-closed before it runs, and high-risk writes wait for owner approval, then replay from the stored payload.
Bidirectional Model Context Protocol: an MCP server exposing 44 tenant-scoped tools to Claude Desktop behind an OAuth 2.1 authorization server, plus a multi-transport MCP client governed by the same gate.
Tenant isolation hardening: per-person feature switches on every route and a positions-based access model rolled out in shadow mode with zero regressions.
Integration economics: a four-provider calendar connector platform with zero frontend code per provider, per-tenant encrypted OAuth apps, and bring-your-own-key LLM routing across OpenAI, Anthropic, Gemini and vLLM.
Sally Desktop — Meetings That Never Leave the Machine
Liquid Technologies · 2026
Challenge
Enterprise buyers wanted meeting intelligence, but refused to stream sensitive audio to the cloud. Transcription and summarisation had to run entirely on-device across Windows, macOS and Linux.
Approach
Tauri 2 desktop app with a Rust core and Next.js UI, shipped to Windows, macOS and Linux.
Real-time dual-source audio pipeline: EBU R128 loudness normalization, RNNoise neural noise suppression and Silero VAD.
Fully on-device Whisper/Parakeet speech-to-text on ONNX Runtime, GPU-accelerated across Metal, CoreML, CUDA, Vulkan, ROCm and optimized CPU.
Authentication handled entirely in the Rust process, so bearer tokens never enter the webview.
LeadGenPro — From Keyword Search to CRM-Ready Leads
Liquid Technologies · 2026
Challenge
Sales teams needed a steady flow of qualified leads from LinkedIn, Reddit, Twitter/X and Hacker News, but each source has its own rate limits, anti-bot defenses and data shapes.
Approach
A Playwright browser-automation microservice performing stealth LinkedIn authentication with security-checkpoint recovery, persistent session reuse and structured profile extraction via LinkedIn's internal API.
An 11-queue BullMQ/Redis pipeline with per-platform rate limiting to keep every source inside its limits.
A provider-agnostic AI qualification layer (OpenAI, Anthropic, Google Gemini) scoring and classifying leads automatically.
A two-stage contact-enrichment waterfall — BetterContact, then Apollo.io as fallback — with automatic sync to GoHighLevel CRM.
AI Voice Agent Platform — Cutting Lead Response from Hours to Minutes
Teknloje Solutions · 2025
Challenge
Inbound leads sat unanswered for hours over weekends and after hours. The business needed 24/7 automated first contact without adding headcount to the sales floor.
Approach
An n8n automation layer monitors inbound emails around the clock and extracts leads via LLM parsing.
Twilio + Retell AI trigger AI outbound calls the moment a lead is detected, with qualification handled conversationally.
Every interaction syncs to the CRM, keeping the sales pipeline current without manual entry.
A selection of AI systems I have designed and shipped — meeting intelligence, lead generation, recruitment, and conversational platforms. Each card opens a short breakdown of what the product does, the impact it had, and the technology behind it.
Featured projects
06 · Background
Education
Bachelor of Science in Computer Science
FAST National University of Computer and Emerging Sciences (FAST NUCES)
Karachi, Pakistan · August 2021 — May 2025
One of Pakistan's top-ranked CS programs.
07 · Learning
Certifications
📜
Supervised Machine Learning
Coursera · Andrew Ng · 2024–2025
📜
Advanced Learning Algorithms
Coursera · Deep Learning, TensorFlow · 2024–2025
📜
Data Cleaning & Processing
DataCamp · Python · 2024
📜
Astera Bootcamp: Data Integration
2023
08 · Contact
Let's Work Together
Open to full-time AI Engineer, Generative AI Engineer, LLM Engineer, and Applied AI Engineer roles. Available for both on-site (Karachi) and remote positions.
