01 / Identity

Hi, I’m Hamza.
I build AI products that deliver past the pilot.

Currently atGEO TVExNielsenIQL’Oréal
03 / Experience
  1. Sep 2024 - Oct 2025NielsenIQData Quality AnalystChecked incoming retail data for errors, followed up on whatever did not match the source, and kept the panel accurate enough for clients to report from.
  2. Aug 2025 - PresentSehatlingsChief Product OfficerHead of product at a small health startup. Decide what gets built next, write the specs, work with the team until it ships, and drop what is not working.
  3. Sep 2025 - PresentGEO TV / Har Pal GeoSystem Developer & AnalystBuild and maintain the internal tools the newsroom uses, including the search and AI parts, and handle the data work that sits behind them.
  4. Apr 2026 - Jul 2026L'OréalBusiness Planning AnalystFound and fixed inconsistencies across several data sources and standardised them into one format the reporting could be trusted on. Built the reports and summaries the business made decisions from, and ran Salesflo as the only person on it.
05 / Status

Available
for work

Start a conversation →
06 / Time
--:--:--
PKT · UTC+5
04 / Work

EDI

An AI-native spreadsheet: natural language in, structured analysis out.

Role
Design and engineering
Year
2025
Stack
LLM Orchestration, Agents, React, TypeScript

Context

A spreadsheet is the most widely used analysis tool in the world and the most unforgiving one: the intent behind a formula lives in the analyst’s head, never in the file. EDI asks what a spreadsheet looks like if the model is part of the grid rather than a chat window bolted onto its side.

Approach

The question is stated in natural language; what comes back is structure: columns, formulas, a transformation you can inspect and re-run, not a paragraph describing what the answer might be.

That framing drove the architecture:

  • Output is verifiable. The model emits operations against the sheet rather than prose about the data, so every result can be traced to the steps that produced it.
  • The grid stays the source of truth. The model proposes; the spreadsheet holds state. Nothing is true because the model said it.
  • Agentic where it earns its keep. Multi-step analysis is decomposed into operations that can be checked individually, which matters more than fluency when the output is numbers.

What I’d do differently

Try it live ↗Repository ↗
02 / About

Rao Muhammad Hamza is an AI engineer in Karachi, three years into building GenAI systems that outlive the demo.

He builds retrieval pipelines, orchestrated LLMs and agents, and the whole stack around them: the retrieval layer, the evaluation harness, the interface the work is judged on. That range is what turns a promising prototype into something a team runs every day.

Currently System Developer and Analyst at GEO TV and Chief Product Officer at Sehatlings, after a year at NielsenIQ in data quality and a season in business planning at L’Oréal. Media, market research and beauty retail: three industries with very different ideas of a good answer.

He is most useful where the requirements are still moving, and will say plainly when a model is the wrong tool. Always open to a conversation about work worth building.

Email LinkedIn GitHub