About

I build machine learning systems that run in production. My stock prediction engine hit a 4.25 Sharpe ratio in backtests. My PyPI package InSightAI has 1.2K+ downloads. I fine-tuned a Whisper model for Urdu and published it on HuggingFace.

Stack: Python, PyTorch, TensorFlow, FastAPI, Docker. I handle everything from data pipelines to model deployment. I write technical articles on Medium explaining what I build and why.

4.25
Sharpe Ratio
1.2K+
PyPI Downloads
18mo
In Production

Projects

Click any project to see the problem I solved, my technical approach, and results.


My flagship project maintained these metrics in production ↓

FLAGSHIP PROJECT

Stock Prediction Engine

Production ML system that maintained a 4.25 Sharpe ratio over 18 months. Built with ensemble methods (XGBoost + LightGBM + Random Forest), 73 engineered features, FastAPI backend, and Docker deployment.

4.25 Sharpe Ratio
18mo Production
73 Features
<3s Predictions
XGBoost LightGBM FastAPI Docker PostgreSQL
View Code Full Details Read Article
Returns Months ↑ 4.25 Sharpe

Additional Projects

InSightAI

2024

PyPI Package for ML Insights

1.2K+ PyPI downloads • Query data in plain English
GitHub PyPI

Climate Impact Predictor

2023

ML Model for Environmental Analysis

±1-6°C temperature accuracy • 0.89 F1-Score achieved
GitHub

Urdu Speech Recognition

2024

Fine-tuned Whisper Model

12.3% WER achieved • HuggingFace model
HuggingFace

DQN Atari Agent

2023

Deep Q-Learning Implementation

Stable training convergence • Deep Q-Learning implementation
GitHub

Ecommerce ML System

2023

Production ML Pipeline

100K+ orders processed • Daily ETL pipeline
GitHub

Experience

AUG 2025 — SEP 2025

AI & Automation Developer

Pexdel Solutions · Lahore, Pakistan

Built customer engagement system using n8n and OpenAI GPT, automating service inquiries. Integrated Twilio voice API and Google Calendar for automated scheduling, delivered in 4 weeks. Coordinated requirements across design and stakeholder teams.

SEP 2023 — SEP 2024

ML Engineer

Freelance via Upwork · Remote

Completed 8+ projects building document analysis systems with Flan-T5 and Llama 3.8B. Created PyTorch transformer prototype, generating $15K revenue. Implemented prompt engineering strategies, reducing client API costs by 40%.

OCT 2022 — DEC 2022

Data Automation Developer

HR Tech Clients via Upwork · Remote

Built automation systems for 5 HR companies processing 50K+ monthly resumes. Designed ETL pipelines with validation frameworks, collaborating with product teams. Developed Power BI dashboards presenting recruitment metrics to executives.

Labs

Some thoughts, reflections, & notes on design and development, along with some latest work in progress.

Fine-tuning Whisper for Urdu: Why 80% WER Happened
Common Voice dataset looked perfect. Model performance was terrible. Here's what went wrong.
01
Vector Database Showdown: 1 Week with 3 DBs
Tested Pinecone, ChromaDB, and Weaviate for RAG. Speed vs accuracy isn't what I expected.
02
n8n Scaling Disaster: When 1000 Workflows Break Everything
Automation worked perfectly at 50 workflows. At 1000, it all fell apart. Memory limits hurt.
03

Built by Sohaib Ahmed using HTML, CSS and JavaScript.

Sohaib Ahmed

Data Scientist & ML Engineer

Lahore, Pakistan · Open to remote opportunities
3+ years building production ML systems
PyPI package author · HuggingFace contributor

Project Title

About

The Problem

What I Built

Key Metrics

Technologies

View on GitHub

Fine-tuning Whisper for Urdu

Failed → Learning • 3 days • 2 weeks ago

The Problem

Initial Whisper fine-tuning for Urdu gave 80% WER (Word Error Rate) instead of the expected ~15%. Model was basically useless.

What I Tried

  • Used Common Voice Urdu dataset as-is (seemed clean)
  • Default UnSloth training parameters
  • Standard learning rate of 2e-4
  • No audio preprocessing beyond basic normalization

What Was Wrong

Three critical mistakes destroyed performance:

  • Audio Quality: Dataset had mixed quality - some clips were 8kHz, others 48kHz
  • No Preprocessing: Didn't normalize amplitude or filter background noise
  • Learning Rate: Too aggressive for low-resource language fine-tuning
# What killed performance learning_rate = 2e-4 # Too high audio_preprocessing = None # Big mistake quality_filter = None # Even bigger mistake

The Fix

  • Filtered dataset by duration (3-10 seconds) and sample rate consistency
  • Added audio preprocessing pipeline: resampling → normalization → noise reduction
  • Reduced learning rate from 2e-4 to 5e-5
  • Implemented gradient accumulation for stable training

Results: WER dropped from 80% to 12.3% in first iteration. Still improving with better data cleaning.

Key Takeaway

Low-resource language fine-tuning needs more care than English. Data quality > data quantity. Audio preprocessing isn't optional.

Vector Database Showdown

Solved • 1 week • 1 week ago

The Experiment

Tested Pinecone, ChromaDB, and Weaviate for RAG system. Same dataset (10K documents), same embeddings (OpenAI ada-002), different databases.

Results

  • Pinecone: Fastest queries (12ms avg), but $70/month for production
  • ChromaDB: Slowest (85ms avg), but free and self-hosted
  • Weaviate: Sweet spot (28ms avg), moderate setup complexity

Winner: ChromaDB for MVP, Weaviate for scale, Pinecone for enterprise.

n8n Scaling Disaster

Failed • Weekend • 3 days ago

What Broke

At 50 active workflows, n8n ran smoothly. At 1000 workflows, memory usage hit 12GB and everything crashed.

The Culprit

n8n loads ALL active workflows into memory. Each workflow averaged ~8MB. Math killed me.

1000 workflows × 8MB = 8GB base memory + execution overhead = 12GB total = Dead server 💀

Learning

n8n is amazing for <100 workflows. Beyond that, switch to Temporal or custom orchestration. No-code tools have hard limits.

My Hobbies

Vol. 01 2024

Offline
Mode

Stories Beyond the Keyboard

A collection of hobbies, obsessions,
& side quests.

Introduction 01

The Side Quests

Data science is my profession, but these hobbies are my sanctuary. They are the "side quests" that keep my creativity alive and prevent burnout.

From the immersive worlds of Teyvat to the storytelling of anime, these interests shape how I approach problem-solving: with curiosity, patience, and a bit of grind.

"To be creative, you must be interested."
Gaming 02

The Gacha Trinity

Three open worlds. Infinite resin anxiety.

I don't just play; I optimize. Managing resources in these games is basically a second job in logistics.

Genshin Impact AR 60 v Luna III
Day 1
Zenless Zone Zero IK Lv. 60 v2.5
Wuthering Waves UL 80 v3.0
Gaming 03

Battle Stations

My setup is optimized for two things: high-FPS gaming and seamless coding sessions. (And maybe looking cool).

PS5 AAA Titles
Mobile Dailies

Gacha taught me probability better than any textbook.

Anime 04
I've tracked
1,914 titles

Since 2016, I've logged every single series, OVA, and movie. It's not just entertainment; it's a dataset of my own interests.

515.9 days
30,870 eps
Comedy Action Fantasy Drama
Now Watching 05

Current Queue

The seasonal rotation. I try to keep it balanced between shonen hype and slice-of-life healing.

One-Punch Man S3
One-Punch Man S3
Live via AniList API
K-Drama 06

Weekend Marathons

Sometimes you just need a 16-episode arc with a predetermined happy ending to reset your brain after debugging a race condition.

Romance Thriller Historical

16 episodes. One weekend. No regrets.

Side Projects 07

After Hours

Why do I code in my free time? Because there's no deadline, no PM, just an idea and a blank canvas.

New tech experiments
Personal automation
Creative web builds
The Loop 08

Creative Orbit

Code
Wall
Play
Idea
repeat

Games and anime spark creative project ideas.

Fin

"The best debugging happens after a quick domain run."

Click pages to flip