Abdul Samad SiddiquiSenior AI Engineer · Karachi
Torn collage: a lone figure, seen from behind, wading through painted cobalt water strung with code windows and node graphs, the sheet torn open on the right onto blueprint graph paper.

Karachi · Pakistan

Abdul SamadSiddiqui

Senior AI Engineer

I build AI systems that survive real use: agents, retrieval, and the infrastructure under them. What I learn goes back to open source.

About me

Abdul Samad SiddiquiBuilds things
that keep
working.

Senior AI Engineer · Karachi, Pakistan

I build AI products and the plain, reliable software that holds them up. The part I care about is what happens after the demo: whether it still answers well on the hundredth question, on a slow day, with input nobody planned for.

I learned how to do this in open source. People read my code in public, told me straight when it was wrong, and I got better. I still work that way, and I still prefer being corrected to being right.

Field kit

  • Python
  • PyTorch
  • Hugging Face
  • TypeScript
  • Ruby on Rails
  • Docker
  • AWS
  • GCP
  • Rust
  • GitHub
Abdul Samad Siddiqui as a young child, in a red hooded sweatshirt.
same person,
considerably fewer
stack traces
The three layers I work acrossA technical drawing of three stacked layers. The top layer is the agent loop , plan, act, observe. The middle layer is retrieval, question, index, ranking. The bottom layer is infrastructure, queues, storage and traces. Orange connectors run vertically between the layers, showing that a failure in any one of them surfaces in the others.one system01, agent loopplanactobserveretriescompaction02, retrievalaskembed+ sparsevectorsrankedmeasured,not assumed03, infrastructurepipelinesstoragetraceswhat actually ranfig. a, the layers are not independent
Working drawing · agent loop, retrieval, infrastructure· drag to read →

Selected experience

Where the work
became real.

Five chapters. Different industries. The same habit of taking useful systems all the way from an uncertain idea to working software.

  1. 01

    Product ownership

    HeyOz

    Senior AI Engineer

    I own an AI media product end to end: shaping the idea, building the system, and carrying it into production. The company has raised $4M.

  2. 02

    Enterprise delivery

    Radium Agency

    AI Engineer

    I built enterprise AI for Canon Europe and learned how to deliver with a multinational team, clear ownership, and real operational constraints.

  3. 03

    Scientific AI

    MetaSensing

    AI Engineer (part-time)

    I led the AI work behind chemical and material analysis, turning experimental data into useful software as the company raised $2.3M.

  4. 04

    Open source term

    Google

    Google Summer of Code

    INCF / HNN-Core

    With guidance from Brown University researchers, I built batch simulation workflows for a neuroscience tool used to study EEG and MEG signals.

  5. 05

    Open source term

    Linux Foundation

    LFX Mentorship

    Open Mainframe / Zowe

    Mentored by a Broadcom engineer, I improved the Python SDK for IBM Zowe and learned to ship changes through a mature open-source review process.

Selected work

Things I have
made & helped shape.

A small set of open-source systems, shown as working artifacts rather than a wall of technology names.

A gallery of nine visual styles rendered by animate-pls, including cut paper, riso, pixel art and blueprint frames

Open source · animation tooling

animate-pls

Animated explainers written in code and narrated in your own voice. I built it as an agent skill that turns one brief into repeatable video, without stock footage or an AI video subscription.

Canvas · local voice · MP4

View project
The BestRAG project mark shown as a small three-dimensional book

Open source · retrieval

BestRAG

A Python library I built for hybrid retrieval. It combines dense, sparse and late-interaction embeddings over Qdrant, so teams do not have to stitch three retrieval systems together.

Python · Qdrant · PyPI

View project
ArchData mark with the words Dataset Pipeline for LLM Routing

Open source · dataset engineering

ArchData

A pipeline I made for generating and augmenting conversational routing data. It is inspired by Katanemo’s Arch-Router, now part of Plano.

Inspiration: Plano

Multi-LLM pipeline · validation · augmentation

View project
HNN-Core interface showing a cortical simulation and its aggregate waveform

Google Summer of Code 2024 · INCF

HNN-Core

I built batch simulation workflows for a tool neuroscientists use to connect cortical models with EEG and MEG signals. The work was developed in the open with HNN-Core maintainers.

Computational neuroscience · batch simulation

View project
GraphNeT logo and the words Deep Learning for Neutrino Telescopes

Open source · scientific machine learning

GraphNeT

Contributions to an open-source deep-learning framework for neutrino-telescope event reconstruction. It is separate from HNN-Core: different science, with the same demand for reproducible software.

Graph neural networks · physics · Python

View project
A faceless silhouette reaching upward, half submerged in a torn collage of cobalt waves, orange arrows and a drawn network graph.

نغمۂ موج ہے، ہنگامۂ طوفاں ہو جا

علامہ اقبال · جوابِ شکوہAllama Iqbal · Jawab-e-Shikwa

a rough rendering:
you are the song of the wave
become the storm

Send
word.

For AI systems work, open-source collaboration, or a careful argument about why something keeps falling over. I read everything.

yours, Abdul Samad

Where to reach me

Karachi, Pakistan · working across time zones