Generative AI & Agents
- LangChain
- LangGraph
- LlamaIndex
- RAG
- OpenAI API
- Claude API
- Prompt Engineering
- LLM Fine-tuning
- Hugging Face
Available for select AI engineering work
Building intelligent systems
that actually ship.
I'm Ismail Qayyum — software engineer, deep learning researcher, and builder of production-grade AI systems. I design agentic LLM workflows, end-to-end data pipelines, and neural networks that move real metrics for real businesses.
AI Engineer · Karachi
0
AWS assets covered on Secure.com platform
0
faster query performance
0
cut in deployment time
0
accuracy improvement on 300K+ records
Python
PyTorch
TensorFlow
Spark
Airflow
AWS
MongoDB
Neo4j
Docker
Kubernetes
FastAPI
Redis
Hugging Face
OpenAI
Anthropic
Git
GitHub
Linux
01 About
An engineer who treats AI like infrastructure — designed, deployed, and accountable.
I graduated from FAST-NUCES with a B.S. in Artificial Intelligence and have spent the past year shipping production AI at Disrupt.com (formerly Gaditek), where I architect Medallion-Architecture pipelines that move over a million events a day across Secure.com's 10,000+ asset security platform.
Before that, I built autonomous LangChain/LangGraph agents in live client environments at Imperium Dynamics, and led computer-vision research at FAST that produced a published Zenodo dataset now cited in autonomous-vehicle simulation work.
My focus is the unglamorous middle of the AI stack: pipelines that don't break, agents that recover from their own mistakes, and APIs that make production data actually usable. I care about latency, observability, and the moment a feature ships — not the demo.
02 Toolkit
The stack I build with, daily.
Data Engineering
- Apache Airflow
- Apache Spark
- Medallion (Bronze/Silver/Gold)
- AWS S3
- MongoDB
- Neo4j
- OpenSearch
- Indexed Views
- ETL/ELT
Deep Learning
- PyTorch
- TensorFlow
- CNNs
- RNN / LSTM
- GANs
- VAEs
- Diffusion Models
- Attention
ML & Data Science
- scikit-learn
- NumPy
- Pandas
- Matplotlib
- NLTK / NLP
- OpenCV
- YOLO
- Computer Vision
Cloud & DevOps
- AWS
- Azure
- Docker
- Kubernetes
- FastAPI
- Redis
- Git
- n8n
- Make (Integromat)
- OpenAPI
Languages
- Python
- Java
- C++
- Dart
- SQL
03 How I Build
End-to-end Medallion Architecture — from raw events to indexed APIs.
Source
Raw telemetry from S3, OpenSearch, Neo4j — 1M+ events/day per tenant.
Bronze
Append-only landing zone. Spark ingestion jobs orchestrated by Airflow.
Silver
Schema-aligned, deduped, type-safe. LLM-driven self-correction across 300K+ records.
Gold
Business-grade aggregates & metrics, indexed for query latency. 45% faster.
Serving
MongoDB indexed views + FastAPI REST endpoints powering analytics dashboards.
04 Research
Models I've designed, trained, and shipped.
Diffusion · Image Restoration
Published · Zenodo
Conditional Diffusion for Sensor-Level Image Restoration
Extended a diffusion model with stage-wise diffusion, adaptive conditioning, and attention to recover clean images from real-world degradations — built for autonomous-vehicle sensor simulation. Designed a novel composite loss (Contextual + Perceptual + L1) that measurably improved reconstruction fidelity over the L1-only baseline.
DCGAN · Generative Modeling
Deep Convolutional GAN on CelebA
Trained DCGANs from scratch on the CelebA dataset at both 32×32 and 128×128 resolutions. Generator (transposed-conv stack) learns to synthesize face crops from a 100-D Gaussian noise prior; the discriminator (strided-conv stack) judges authenticity in a min-max game until the generator's samples become indistinguishable from real faces.
- z (Gaussian)100-D
- FC + reshape1024 · 4²
- conv† · BN · ReLU512 · 8²
- conv† · BN · ReLU256 · 16²
- conv† · BN · ReLU128 · 32²
- conv† · BN · ReLU64 · 64²
- conv† · Tanh3 · 128²
- input image3 · 128²
- conv · LReLU64 · 64²
- conv · BN · LReLU128 · 32²
- conv · BN · LReLU256 · 16²
- conv · BN · LReLU512 · 8²
- flatten + FCscalar
- sigmoidD(x) ∈ [0,1]
CNN · Image Classification
Custom CNN on CIFAR-10
Designed a custom convolutional architecture from scratch in PyTorch to classify CIFAR-10 across 10 categories — airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck. Iterated through regularisation, augmentation, and architectural changes to push test accuracy to 81%.
Code Sample · LangGraph Agent
How I wire up an agent.
Excerpt from the Teaching Assistant Agent — a typed LangGraph node that grades a single answer against a rubric, with structured output and stateful retries.
from langgraph.graph import StateGraph
from langchain_anthropic import ChatAnthropic
from .state import GradingState
from .prompts import build_grading_prompt, parse_grader_output
def grade_answer(state: GradingState) -> GradingState:
"""Score one student answer against the rubric."""
llm = ChatAnthropic(model="claude-opus-4-7", temperature=0)
prompt = build_grading_prompt(
question=state.question,
rubric=state.rubric,
answer=state.answer,
)
score, feedback = parse_grader_output(llm.invoke(prompt))
return {**state, "score": score, "feedback": feedback}
graph = StateGraph(GradingState)
graph.add_node("grade", grade_answer)
graph.add_node("verify", verify_against_rubric)
graph.add_conditional_edges("verify", route_on_confidence)
graph.set_entry_point("grade")
agent = graph.compile()05 Experience
Where I've shipped.
-
Software Engineer II — AI & Data Systems
Disrupt.com (formerly Gaditek) · Karachi- Architected end-to-end Medallion Architecture pipelines (S3 → Bronze → Silver → Gold) on Apache Airflow + Spark, delivering features from requirements to production across optimised MongoDB indexed views and REST APIs.
- Owned full feature lifecycle — ingestion, orchestration, schema design, API delivery, shipping — across Secure.com's 10,000+ AWS asset security platform.
-
Software Engineer I — AI & Data Systems
Disrupt.com · Karachi- Engineered real-time ingestion pipelines processing 1M+ daily events per tenant from AWS, OpenSearch, and Neo4j, powering multi-tenant security analytics for Secure.com.
- Integrated LLM-based self-correction across 300K+ vulnerability records — improved data accuracy by 30%.
- Designed multi-environment databases for 300K+ CVE/CWE entries — 45% query performance gain.
- Built staging→production sync tooling — cut deployment time by 80% and manual review workload by 60%.
-
Machine Learning Engineer Intern
Imperium Dynamics · Karachi- Designed and shipped LangChain + FastAPI autonomous agents with chained LangGraph multi-step reasoning, deployed in live client environments.
- Implemented a Redis-backed stateful memory layer for low-latency, high-availability LLM interactions.
- Automated internal email workflows via Microsoft PowerApps — 40% less manual ops.
- Authored FastAPI architecture following enterprise standards: security, modularity, versioning, full OpenAPI docs.
-
Student Research Lead — Computer Vision & Generative AI
FAST-NUCES · Karachi- Led creation of NBF_StreetView — a 19,000 image-pair dataset for autonomous-vehicle sensor simulation; published on Zenodo (DOI: 10.5281/zenodo.14552482).
- Extended a diffusion model with stage-wise diffusion, adaptive conditioning, and attention for real-world image-degradation restoration.
- Designed a novel composite loss (Contextual + Perceptual + L1) that measurably improved reconstruction fidelity.
06 Selected Work
Things I've built & published.
Production · Internal
Secure.com Data Platform
Multi-tenant Medallion-Architecture data platform handling 1M+ daily events per tenant from AWS, OpenSearch, and Neo4j. Owns the path from raw security telemetry to indexed Mongo views powering dashboard analytics across 10,000+ AWS assets.
- Airflow
- Spark
- S3
- MongoDB
- Trino
- Iceberg
1M+daily events
45%query gain
80%deploy time cut
LLM Agent · Open Source
Teaching Assistant Agent
Autonomous LLM agent that grades student assignments end-to-end — parses PDFs, evaluates answers against rubrics, and generates structured feedback reports. Eliminates hours of manual grading per session.
- LangChain
- LangGraph
- FastAPI
- LLM Agents
Published Dataset · Research
NBF_StreetView
Computer vision dataset of 19,000 degraded/clean street-view image pairs for autonomous-vehicle sensor research. Published on Zenodo and cited as Qayyum, I., Badar, A., & Rizwan, A. (2024).
- Python
- OpenCV
- imgaug
- Diffusion
Enterprise Automation
LangGraph Autonomous Agents
Production LangChain + FastAPI agents with chained LangGraph multi-step reasoning and a Redis-backed stateful memory layer — deployed in live client environments at infrastructure-grade reliability.
- LangGraph
- FastAPI
- Redis
- OpenAPI
More on GitHub.
Notebooks, agents, infra experiments — pinned and updated regularly.
07 Writing & Publications
Published, cited, read.
-
Zenodo · Published Dataset 2024
NBF_StreetView
Qayyum, I., Badar, A., & Rizwan, A. (2024). Computer-vision dataset of 19,000 degraded/clean street-view image pairs.
DOI: 10.5281/zenodo.14552482 -
AI Monks · Medium Jun 2024
Autoencoders: Basics and Beyond
PyTorch implementation of autoencoders from scratch — published in the AI Monks publication.
Read on Medium -
AI Monks · Medium Oct 2023
Understanding the Neural Network Model
Deep dive into neural-network architectures and forward propagation.
Read on Medium
08 Contact
Have a hard problem
that needs an AI engineer?
I'm selectively available for AI engineering work — agentic LLM systems, data platforms, and ML automation. Drop a line and let's talk.
Email
ismail.qayyum4@gmail.com
Phone
+92 334-020-2266
LinkedIn
/in/ismailqayyum
GitHub
@IsmailQayyum
Medium
@ismail.qayyum4
Location
Karachi, Pakistan