Muzzammil Sarwar

Senior full-stack AI engineer

I design software systems end to end and build them quickly with AI-assisted development, keeping tests and code review as the quality gate. Over five years I've worked on product platforms, AI systems and data tooling in TypeScript and Python.

Right now I'm the sole engineer on Controla-VA, a workforce platform with five web portals, a NestJS API and a desktop agent running on about 50 machines. In my own time I maintain pgrecon, an open-source Oracle-to-PostgreSQL migration tool.

Islamabad, Pakistan. Open to remote roles.

Email GitHub LinkedIn CV (PDF)

Work

Controla-VA

Growth Rise Solutions · 2026 · controla-va.com

Time tracking, productivity monitoring, recruiting and finance for a managed staffing company, in one platform. I'm the only engineer, covering architecture, code, infrastructure and production support.

Next: in-app messaging with Hebrew–Urdu translation of text and voice, and an AI voice interviewer for first-round candidate screening.

Next.js · NestJS · PostgreSQL · pgvector · Prisma · Redis · Electron · C# · OpenAI · Gemini · Docker · Cloudflare

pgrecon

Open source · GitHub · PyPI

An Oracle-to-PostgreSQL migration assessment and schema conversion tool. It inventories an Oracle database from an offline dump, runs a deterministic rule engine that flags the constructs that drive migration cost, and emits PostgreSQL DDL for everything it can prove, with a named refusal for everything it can't. Published on PyPI with 16 releases so far.

Python · SQLite · PL/SQL · PostgreSQL · GitHub Actions

Box

Growth Rise Solutions · box.com.pk

A hybrid collaborative and content-based recommender built with LightFM and served through FastAPI. It powers product discovery in a live grocery app with thousands of products and shoppers.

Python · LightFM · FastAPI

Oracle-to-PostgreSQL migration tool

Bitnine Global

A Python migration tool with a CLI and a desktop GUI, covering DDL extraction, schema transformation and data migration. Its design lets DB2 and MySQL be added without rewriting the core.

Python · Oracle · PostgreSQL

Smart Cold Storage Dashboard

Final-year project · NUST

Real-time temperature monitoring for cold-storage automation. ML models forecast temperature trends and recommend storage conditions, supporting predictive maintenance and lower energy use.

Python · Machine learning · Forecasting

Kids Learning Platform

Freelance

An eye-tracking model measures engagement in real time, and a fine-tuned GPT-3 model adapts the learning content to each child. I built the front end, the back end and the model APIs.

Computer vision · Fine-tuned GPT-3 · REST APIs

Case studies

Problems from across my work, written up at the architecture level. Client code and data stay private.

Reliability · Controla-VA

A 17-minute session billed as 16.5 hours

Problem: A worker tracked 17 minutes and closed their laptop; by morning the session read 16.5 hours. Open sessions were billed as now − start, so any agent that died without a clean stop kept billing.
Approach: Billed time now has to be backed by evidence. A server-side clamp caps each session at its last recorded activity plus a grace window, and a reaper closes sessions whose agent has gone quiet. I later closed a second path where late offline data re-billed dead gaps, and designed an event-sourced model in which billed time is the union of verified intervals.
Result: Every installed agent version was corrected at once, without a client update. Under the new model, phantom hours can't be billed at all.

AI · Controla-VA

Keeping AI matching costs flat as data grows

Problem: Matching jobs to candidates meant an LLM scoring every CV in a keyword category. Cost grew with every upload, and keywords missed obvious matches such as "ML engineer" against "machine learning".
Approach: Retrieve, then rank. Each CV is embedded once into pgvector with an HNSW index, each job retrieves its nearest candidates, and only that shortlist goes to the LLM. Keeping vectors in PostgreSQL meant one query next to the existing filters and no new infrastructure.
Result: LLM cost per scan depends on the shortlist size, not on how many CVs exist. Candidates also surface for related roles they never applied to.

Desktop · Controla-VA

One desktop agent, about 50 machines, three operating systems

Problem: The agent runs on machines I can't touch. It has to survive sleep, crashes, power cuts and bad networks, update itself, and never lose or double-count paid time.
Approach: Electron with native modules. Every sample gets a client-generated ID, so the offline queue can retry safely. Updates install silently from a versioned channel, and the server enforces a minimum version. Browser URLs come from C# UI Automation over FFI. For macOS and Linux, a build-time platform layer keeps each bundle free of the other systems' code.
Result: The fleet updates itself, and most fixes ship on the server without a new build. Windows and Linux builds ship today; macOS is waiting only on Apple code signing.

AI agents · Growth Rise Solutions

Multi-agent AI for hands-off business processes

Problem: Processes that spanned CRMs, messaging tools and data sources depended on people copying data across and pushing each step along.
Approach: CrewAI and LangChain agent crews with custom memory and tool orchestration, wrapped in FastAPI services and triggered by event-driven n8n workflows. More than ten went into production. Each service ships as a Docker image to Railway through GitHub Actions, with versioned releases and rollback.
Result: Multi-step tasks run end to end with no manual handoffs, and a bad release can be rolled back in minutes.

Machine learning · Box

Recommendations for a live grocery app

Problem: The app needed better product discovery, but new products and new shoppers have little or no purchase history to learn from.
Approach: A hybrid LightFM model that combines collaborative signals (what similar shoppers buy) with product metadata, so a new item can be recommended from its features alone. It runs as a FastAPI microservice that the app calls in real time.
Result: Live in production on box.com.pk, where it drives product discovery.

Open source · pgrecon

Pricing an Oracle migration before anyone commits to it

Problem: Oracle-to-PostgreSQL migrations go over budget because the expensive constructs (package-level state, autonomous transactions, interval partitioning, database links) surface late. Assessment tools usually want a live connection to a production database, which DBAs rightly refuse.
Approach: pgrecon never connects. The DBA reviews and runs a read-only SQL*Plus script, and only files cross the boundary. The dump loads into a local SQLite inventory, and a rule engine with stable rule IDs queries it, so the same dump always gives the same findings. A converter then emits PostgreSQL DDL for what it can prove and a named refusal for what it can't, and DDL that fails to parse becomes a finding instead of being skipped.
Result: An open-source tool on PyPI with 16 releases and CI. It ships a real Oracle 21c dump, so anyone can try it without an Oracle database.

Experience

Dec 2024 – now

Growth Rise Solutions · remote

Full-Stack Developer May 2026 – now

Sole engineer and architect of Controla-VA: five Next.js portals on a NestJS API, PostgreSQL and Prisma.
Built an Electron desktop agent used on about 50 workstations, with Windows and Linux builds; macOS is waiting on Apple signing.
Shipped AI features (RAG matching, vision classification, LLM document parsing) and a multi-currency finance platform.
Run staging and production on Docker, Caddy and Cloudflare, with CI/CD gated by automated tests.

Python / AI Engineer Dec 2024 – Apr 2026

Built multi-agent AI systems with CrewAI and LangChain, including custom memory and tool orchestration.
Shipped 10+ production automations connecting agents to CRMs, messaging and data sources through event-driven n8n workflows.
Built FastAPI backends serving AI features to hundreds of daily users, plus Next.js and Streamlit front ends.
Built the LightFM recommendation service behind the Box grocery app, and owned Docker deployments, releases and rollback.

Apr 2023 – Dec 2024

Software Engineer · Bitnine Global, Islamabad

Designed a Python test-automation framework for regression, smoke and sanity testing across Bash, PowerShell and CMD, automating over 70% of manual regression testing.
Built an Oracle-to-PostgreSQL migration tool with a CLI and a desktop GUI.
Structured it so new source databases (DB2, MySQL) plug in without touching the core.

Apr 2021 – Aug 2022

Full-Stack Developer (MEAN) · Zepto Systems, Islamabad

Built responsive Angular and AngularJS interfaces for a full-stack web application.
Designed Express.js REST APIs on MySQL, with Strapi as a headless CMS.
Used Pug templates to keep views maintainable and UI changes quick.

2019 – 2023

BS Computer Science · NUST (SEECS), Islamabad

Skills

Languages: TypeScript, JavaScript, Python, SQL, Bash, C#
Backend: Node.js, NestJS, Express, FastAPI, REST APIs, BullMQ, Redis, Auth.js, webhooks
Frontend: Next.js, React, Angular, Streamlit, Tailwind CSS, Radix UI
Data: PostgreSQL, MySQL, Prisma, pgvector, schema design, query optimisation, database migration
AI: OpenAI, Gemini Vision, RAG and embeddings, CrewAI, LangChain, LightFM
Automation: n8n, event-driven workflows, third-party API integration
Desktop: Electron on Windows, macOS and Linux, native modules and FFI, Windows UI Automation (C#)
DevOps: Docker, GitHub Actions, Caddy, Cloudflare, Railway, Linux
Testing: Jest, React Testing Library, Python test automation
How I work: I own the architecture: service boundaries, data model and failure modes. Claude Code and Cursor handle much of the implementation. Tests, type checks and my own review decide what ships.

Contact

Email is the quickest way to reach me: muzzammil@umaish.com

GitHub LinkedIn +92 335 9348075 CV (PDF)