MAAZ

Software & AI engineerSix years shipping production systems
Independent since 2025Previously senior at Astera Software
Karachi, PakistanAvailable for work
[01]// ApproachThree phases
How the work goes

I build systems that outlive my involvement

After 5 years at Astera Software where I grew from associate to senior, led teams, and built everything from microservices to AI-powered workflows. Then I went independent. Now I help clients build web applications and deploy AI agents that actually solve problems.

01

Find the real constraint

Most briefs describe a symptom. Before anything gets built I want the trace, the query plan, or the ledger — whatever the actual evidence is. The refund agent that started my last essay looked fine on the diagram and wrong in the log.

02

Build the boring version

Then make it fast. A dictionary lookup and an if-statement beat a model call for authorization, and they cost nothing. Cleverness goes where the problem is genuinely ambiguous, not where it is merely tedious.

03

Instrument, then leave

Log every gate that blocks and every gate that errors. The measure of the work is whether someone else can reconstruct what happened six months after I have gone.

[02]// Work5 shipped · 2024—2026
Selected

Things that are running

Products I own and systems shipped for clients. Each one is here because something in it was genuinely hard.

01

KhataFlow

Invoicing and bookkeeping on a real double-entry ledger, for businesses that bill across currencies.

  • Multi-currency invoicing with rates snapshotted per transaction and automatic realised FX gain/loss
  • 48 currencies, per-business base currency, with historical rates preserved on every document
Stack
TypeScript · Next.js · NestJS · PostgreSQL · Prisma · Better Auth
Kind
Client work · 2026 · active
Case study
02

Stencil

Invoice extraction where AI authors the rules once and a deterministic interpreter replays them for free.

  • Per-invoice model cost drops to zero once a supplier layout is known
  • Deterministic, reproducible extraction — the same PDF twice gives the same rows
Stack
Python · FastAPI · Celery · Redis · MySQL · pymupdf
Kind
Product · 2026 · active
Case study
03

Signal

Your job search, minus the searching.

  • Location fit scored separately from role fit, so a job you can't take never tops the list
  • One command starts API, scheduler and worker — no broker, no database server, no cron
Stack
Python · FastAPI · SQLite · httpx · BeautifulSoup · APScheduler
Kind
Product · 2026 · active
Case study
04

HotPlate

Taking a seventeen-year-old Karachi catering business off paper — quotes, proformas, invoices and payments.

  • Quote → proforma → invoice → payment as one tracked flow instead of WhatsApp and paper
  • Proforma invoices modelled as their own document, because that is what secures a booking
Stack
TypeScript · Next.js · Supabase · PostgreSQL · React PDF · Resend
Kind
Client work · 2026 · active
Case study
05

Raqam

Personal finance for Pakistan, built on the only data source that exists: the bank's SMS.

  • Every transaction points back at the alert it came from — nothing inferred is shown as fact
  • Parser uncertainty routed to a review queue instead of a confident wrong number
Stack
TypeScript · React Native · Expo · Supabase · PostgreSQL · Jest
Kind
Product · 2026 · wip
Case study
[03]// RecordAstera Software · 2020—2025
Measured, not claimed

Numbers with a source

Every figure below comes from a specific piece of work with a name attached. None of them are round, because real ones rarely are.

80%QA blockers

Claude-driven PR review on installer pipelines

−$10k/yrTooling spend

InstallShield → WiX migration

92%Install time

Installation Manager, rebuilt on Builder/Factory

67%Critical vulns

SonarQube audits across 12 sprints

C#.NET Core 8TypeScriptNext.jsReactPythonFastAPIPostgreSQLMongoDBSQL ServerRedisgRPCRabbitMQMassTransitDockerAzure DevOpsAWSGCPVercelAvalonia XPFClaudeLLM orchestrationSonarQubeWiX
[04]// Writing7 essays
In the open

What I got wrong first

Long-form on agents, ledgers and pipelines. Written after the fact, with the trace included.

The safety check that runs after the money moves is not a guardrail. It is a receipt.

Most of what I write comes out of getting something wrong in production and then having to explain it. That turns out to be the only reliable source of anything worth saying.

[05]// ServicesTwo client slots open
Engagements

What you can hire

I'm freelancing, currently working with two clients on projects ranging from web applications to deploying AI agents for their business workflows. If you need software built or AI integrated into your stack, let's talk.

01

AI agents in production

Tool authorization, input screening, evals, and the failure modes nobody chooses. Deployed, not demoed.

02

Full-stack product

Next.js, .NET Core, FastAPI, Postgres. End to end: architecture, build, deploy, iterate.

03

Systems that must not drift

Double-entry ledgers, multi-currency, reconciliation. Where a rounding error is a legal problem.

04

Legacy migration

WinForms to WPF. InstallShield to WiX. Monolith to microservices. Without stopping the business.

05

LLM pipelines & evals

Prompt testing as code. Deterministic extraction. Model cost that drops to zero after the first run.

06

Technical due diligence

Read the codebase, name the risk, price the fix. Written up so a non-engineer can act on it.

[07]// ContactReplies within a day

Got something that has to work?

Tell me what breaks, what it costs when it breaks, and what you have tried. That is enough for me to say whether I can help.