Data Engineer · Lahore, open to remote

Muhammad Nouman

Six hours of compute, rewritten into ninety seconds.

Single-family pipeline · rows in flight Legacy · disk between every stage Rebuilt · one plan, columnar
6 h1.5 min Per-ticker compute
6 GB18 MB Raw storage footprint
10.6 M Price observations per run
7sectors·~35tickers Data portal coverage

Selected work

What actually changed.

Rebuilding a 13-stage analytics pipeline

2025 – 2026

Technical lead · single-family, then multifamily

94 modules, about 16,500 lines, and a CSV file written to disk between every one of thirteen stages — so multi-gigabyte files were re-read and re-parsed roughly twenty-one times per run. Six hours per ticker. But the disk round-trips were only half of it. One algorithm was doing far more arithmetic than the data justified:

One home’s rent over a year 0 rows evaluated
Rent only moves on lease events, so the old daily sampling restated a known value hundreds of times over. A DuckDB ASOF join evaluates the events alone: 195× fewer rows, 90× faster.
6 h → 1.5 minCompute per ticker
6 GB → 18 MBRaw storage
94 → 42Modules
195× / 90×Rows / speed
PythonDuckDBPolarsApache Arrow ParquetMySQLpytestGitLab CI

Path

How I got here.

Nov 2024 — present

Data Scientist / Data Engineer

Big Byte Insights · Lahore, Pakistan

Rebuilt the single-family analytics pipeline end to end, then carried the same rewrite through multifamily. Own the aggregation layer feeding the client-facing data portal across seven sectors.

Jul 2025 — Aug 2025

Manhwa translation, end to end

Personal project · open source

A comic page goes in, a translated comic page comes out. YOLOv8 finds the speech bubbles, PaddleOCR reads them, an open-weight LLM translates, and the result is typeset back into the original bubbles — the part everyone underestimates, because text that fits in Korean rarely fits in English. Source →

Jul 2023 — Jan 2024

Generative AI for software engineering

COMSATS University Islamabad · unpaid research with a faculty supervisor

Trained an RNN sequence model in TensorFlow over ~350 software projects carrying 400–1,000 requirements each, generating structured requirement statements at 92% accuracy against ground truth. Built a companion LLM pipeline producing UML class and architecture diagrams, with structured-output validation so nothing unparseable shipped.

Graduated Dec 2024

BS Software Engineering

COMSATS University Islamabad, Lahore Campus · CGPA 3.5 / 4.0

Plus the Machine Learning Specialization from DeepLearning.AI and Stanford Online, August 2024.

Toolkit

What I actually build with.

Languages

Python, SQL, Bash

Engineering

ETL and ELT pipeline design, incremental ingestion and backfill, analytical data modelling, columnar storage and partitioning, schema contracts, data quality and reconciliation

Data

DuckDB, Polars, pandas, Apache Arrow, Parquet, NumPy, SQLAlchemy, pandera

Stores

MySQL, Snowflake, BigQuery, AWS S3, Google Cloud Storage, Azure Blob, SFTP

ML & AI

scikit-learn, XGBoost, DBSCAN, PCA, TensorFlow, LLM application development with open-weight models, prompt engineering, computer vision (YOLOv8, PaddleOCR)

Practice

pytest, mypy strict, ruff, structured logging, memory and performance profiling, Git and GitLab CI, Linux

Contact

Open to remote data engineering roles.

If you have a pipeline that costs more than it should — in hours, in memory, or in the nerve it takes to change it — that's the work I like most.