ML engineer · Rahim Yar Khan

I make small language models useful.

Now
Independent — applied ML and product builds
Before
ML engineer at wAI Advanced Industries
Study
BS Information Technology, 2025
Based
Rahim Yar Khan, Pakistan

I fine-tune and quantize models small enough to run on modest hardware, then build the systems around them: evaluation pipelines, JSONL data flows, deployment. Lately, task-specific models for multilingual sentiment and structured output, plus the evaluation that decides what ships.

ML work

Selected ML work

Six ML projects · 2025—2026

Applied MLNLP

Multilingual Sentiment Engine

Sentiment across English, Urdu, Arabic and Roman Urdu.

A multilingual sentiment model optimized for speed: 4× faster inference with accuracy held. Built on Gemma and Unsloth, and evaluated across all four.

  • Gemma
  • Unsloth
  • Multilingual NLP
Languages
English, Urdu, Arabic, Roman Urdu
Result
4× faster inference
Accuracy
Held after optimization
Stack
Gemma, Unsloth

Applied MLAlara

Alara SLM Training

Task models for sentiment, structured output and reasoning across the Alara swarm.

Small task-specific models, trained and quantized to run inside a multi-agent system. The work covered sentiment, schema-valid output and reasoning, with the training and evaluation loop built around each task.

  • PyTorch
  • Unsloth
  • Quantization
Models
Task-specific SLMs
Stack
PyTorch, Unsloth
Optimization
Quantization
Context
Alara swarm

More ML work

  1. Structured Output LLM

    Gemma 270M producing schema-valid JSON for automation workflows, pipelines built end to end.

    Gemma · JSON Schema · Pipelines

  2. Automated Fine-Tuning System

    Ingests JSONL datasets, triggers fine-tuning runs, returns evaluation reports with version tracking.

    FastAPI · JSONL · MLOps

  3. Reasoning-Enhanced Model

    Chain-of-thought fine-tuning on Llama 3.2 3B, served through llama.cpp.

    Llama 3.2 · CoT · llama.cpp

  4. Research Assistant System

    Reads ArXiv and synthesizes answers, orchestrated by LangGraph.

    LangChain · LangGraph · RAG

Product engineering · one product build

Independent2026

Sabzi Mandi

An offline ledger for a vegetable commission shop.

One offline app for a real shop: a credit khata, owner-settlement bills printed as PDF slips, commission and expenses. Urdu interface, whole-rupee money, one SQLite file.

  • React 19
  • TypeScript
  • Fastify
  • SQLite
  • Playwright
Platform
Desktop app, runs offline
Interface
Urdu, right to left
Money
Whole rupees, no paisa
Output
PDF bill slips, one SQLite file
The Sabzi Mandi desktop ledger running offline.
Offline ledger · desktop

Experience

Where I have worked

2025—2026

  1. Nov 2025 — Sep 2026

    Machine Learning Engineer

    wAI Advanced Industries

    Fine-tuned small language models for enterprise products, and built the evaluation pipelines that decided what shipped.

  2. Jul 2025 — Sep 2025

    AI Engineer, intern

    Nexus Technologies

    Applied AI projects: training workflows, data pipelines, and the production habits that come with them.

  3. Sep 2026 — now

    Independent

    Applied ML and product engineering

    Small-model experiments, evaluation tooling and independent builds. Recent product: Sabzi Mandi, an offline ledger.

Toolkit

What I work with

The stack I reach for

Models & training

  • PyTorch
  • Scikit-Learn
  • Unsloth
  • LLaMA-Factory
  • Gemma
  • Qwen
  • Llama
  • 4-bit quantization
  • Chain-of-thought
  • Multi-task learning

Serving & systems

  • llama.cpp
  • GGUF conversion
  • Ollama
  • FastAPI
  • REST APIs
  • Docker
  • AWS EC2
  • HuggingFace Spaces
  • Gradio
  • MCP

Data & evaluation

  • Pandas
  • NumPy
  • JSONL pipelines
  • Matplotlib
  • Seaborn
  • Multilingual datasets
  • RAG systems
  • Evaluation reports

Specializations

  • Small language models
  • Edge AI
  • Model optimization
  • On-device inference
  • Privacy-focused AI
  • Agentic frameworks
  • Multi-agent swarms
  • Structured outputs

Writing

In the drafts

Three essays, not yet published

  1. Why Small Language Models Are the Future of Enterprise AI

    Efficiency and privacy beat parameter counts.

    Draft
  2. A Practical Guide to 4-bit Quantization

    Shrink a model four times without wrecking accuracy.

    Draft
  3. Building Automated Fine-Tuning Pipelines with FastAPI

    From JSONL ingestion to evaluation reports.

    Draft

Contact

Get in touch

Open to ML engineering roles, applied-model collaborations, and conversations about small models.

zamanayaz2025@gmail.com