Manzar Anjum.

Full-stack AI engineer / Remote from Pakistan / US Eastern and EU hours

I have built AI agent systems for 29 businesses, and I support them after launch.

I'm Manzar Anjum. I design, build and run the whole system: the agents, the search they rely on, the approval steps that keep them safe and the servers underneath. Python, TypeScript and PostgreSQL. More than 10 years in systems and software.

Manzar Anjum
Manzar AnjumPakistan, UTC+5
  • 29businesses I have built AI agent systems for
  • ~300business documents a month in two of their offices, counted from the offices' own records
  • 3,285independent checks of finished work in five weeks. 934 sent the work back
  • <0.3skeyword lookup across about 1 million records. It used to take up to 6.6 s

01 / Selected work

Seven of the systems I have built.

Five are my own platform and tools. One is an agency's platform, shown without its name. One is a build for a company that agreed to be named. The other client systems stay private.

RequestTask, written downwhat / how / proofAgent works the taskone owner, start to finishCode checksread every commandPersonapprovesbefore it runsmessages, payments,deletes, client serversSecond agent verifiescold start, live systemclaims doneVerdict + recordpass tied to the filesthe record isthe next task's memory
Fig. 01 How a task moves: written down, checked by code, approved by a person, verified by a second agentDiagram

Production platform

The agent platform I build and support client systems with

Up to 14 agents work in parallel on this platform. I build and support client systems with it. Nothing reaches a customer or a bank account until a person approves it.

  • Python
  • PostgreSQL
  • pgvector
  • Docker
  • Claude
Recordsabout 1 millionVector indexpgvector: by meaningKeyword indexPostgreSQL: exact wordsRerankernewest decision winsAnswer+ sourceKEYWORD LOOKUPbeforeup to 6.6 snowunder 0.3 s
Fig. 02 Two searches over the same records, reranked, answered with the source attachedDiagram

Search and RAG

RAG search over a million records, with keyword lookups under 0.3 seconds

Hybrid vector and keyword search on PostgreSQL. Every answer cites the record it came from.

  • Python
  • PostgreSQL
  • pgvector
  • Embeddings
  • Reranking
TeamClientsPartnersOne web appinvites / roles / tenantseach sees only what is theirsGoogle sign-in broker11 products, read or writeCRMClaude works in it via MCPProbetests a connection
Fig. 03 One app for an agency's team, its clients and its partners, and what sits behind the loginDiagram

Agency platform

An agency's operating platform, with client portals and a CRM that Claude can work in

One web app for an agency's team, its clients and its partners: logins and roles, a sign-in broker for Google, and a CRM.

  • Node.js
  • OAuth 2.0
  • MCP
  • Caddy
  • Linux
Fig. 04 A quotation asked for in chat and returned as PDF and WordIllustration, made-up data

Client build / Phono Engineers (Pvt) Ltd

Quotations and tender papers from a WhatsApp message

A telecom engineering company asks in chat and gets the finished document back as PDF and Word.

  • Python
  • TypeScript
  • WhatsApp
  • Google Drive
  • LibreOffice
Fig. 05 A grid of live terminal sessions in one browser pageIllustration, made-up data

Infrastructure

Thirty terminals in one browser tab

The workstation I built to supervise long-running agents from a laptop or a phone.

  • Python
  • tmux
  • ttyd
  • systemd
  • Caddy
My serverCoordinatorPythonstate / queue / memoryreminders / limitsrecoverythe model reasons, code decidesprivate tunnelRented GPUone NVIDIA A40, 46 GBllama.cpp, CUDA build27B model, GGUF: servingvLLM32B coding model, AWQ: ready7B vision modeltrialled as a canaryControl scriptsstart / stop / status / rollback
Fig. 06 The setup as audited in August 2026: one rented GPU behind a private tunnel, three models, scripted controlDiagram

Self-hosted models

Open models on one rented GPU

In 2026 I ran a private assistant on open models on one rented NVIDIA A40: llama.cpp and vLLM, with scripted start, stop and rollback.

  • llama.cpp
  • vLLM
  • CUDA
  • Qwen
  • Bash
Fig. 07 The day's plan and a due item, posted by the botIllustration, made-up data

Internal tool

A Discord bot that runs a small team's day

Deadlines, the daily plan and the numbers, posted and chased by a bot instead of a person.

  • Python
  • Discord API
  • systemd timers

02 / How I work

Agents are only useful when you can trust the result.

  1. A person approves what leaves the building.

    Client messages and payments wait for a human yes. That rule lives in code, where no prompt can argue with it.

  2. A second agent checks the first.

    When an agent reports a task finished, a fresh agent with no memory of the work tests the claim on the running system. In five weeks that came to 3,285 checks, and 934 of them sent the work back.

  3. Problems surface before the customer sees them.

    Every morning, automated checks walk through two client offices' daily workflows before the staff log in.

  4. AI coding agents are my daily tools.

    I build with Claude Code and Codex every day. No change counts as done until a separate, fresh check passes it on the running system.

03 / Experience

From telephone exchanges to AI agents.

  • Mar 2023 to nowSenior Full-Stack AI EngineerSales Amplifiers LLCThe agent platform, its search, and AI agent systems built and supported for 29 businesses.
  • Dec 2025 to nowConvener, IT Automation and Digital TransformationLahore Chamber of Commerce & IndustryI lead the chamber's standing committee.
  • Jul 2016 to Dec 2022System EngineerPhono Engineers (Pvt) LtdPhone systems and networks across three cities, and the company's paperwork moved into software.
  • 2016BS, Computer ScienceLehman College, City University of New York

More about me

04 / Contact

Hiring a senior AI engineer?

I work US Eastern and EU hours from Pakistan.