Skip to content

Muhammad Etisam Ul Haq

I build LLM systems and put them in production.

AI/ML engineer and data scientist working on large language model systems — retrieval-augmented generation, agents and tool use, fine-tuning and evaluation — and on the data pipelines, APIs and deployment that carry them into production.

Role
Software Engineer — AI/ML
Focus
LLM systems · RAG · Agents · Backend systems · Applied ML
Study
BS Data Science, PUCIT — University of the Punjab
See the workRésumé
Portrait of Muhammad Etisam Ul Haq
Building Montlify, Clip Insights
About

Applied AI, built end to end — then measured.

A short account of what I work on, where I studied, and the problems I keep coming back to.

I work across the full path of an AI system: framing the problem, building the retrieval or training pipeline, evaluating it honestly, then wrapping it in an API and shipping it. Most of what I build ends up in front of real users — a browser extension with a live user base, an AI productivity platform, an MCP server mounted inside a production FastAPI service.

My background is a BS in Data Science at PUCIT, University of the Punjab, where I was a teaching assistant for Database Management Systems. Alongside that I have spent 230+ lessons teaching Python, SQL, machine learning and data analysis one-to-one, which is the fastest way I know to find out whether you actually understand something.

  • Shipped

    Four platforms live in production

  • Competitions

    Five wins · one international ranking

  • Taught

    230+ lessons, 40+ students

  • Hardware

    NVIDIA H100 / A100, AMD Instinct

Research interests

The questions I keep returning to, and the projects where I have actually worked on them.

  • Retrieval-augmented generation

    Grounding language models in documents that can be cited back. Built RAG pipelines over legal contracts (Sovereign), textbooks (Study Buddy) and video transcripts (Clip Insights) using LangChain with Pinecone and ChromaDB.

  • Knowledge graphs for structured retrieval

    Retrieval over an explicit graph rather than flat text. Built a Graph RAG system on Neo4j for UAE tax and legal regulation at Techtics.AI, where the relationships between provisions carry the answer.

  • Efficient fine-tuning and inference

    LoRA adapters and vLLM serving on NVIDIA H100 and A100 GPUs, and private inference on AMD Instinct GPUs for Sovereign — getting useful models onto hardware budgets that real teams have.

  • LLM evaluation

    Measuring what a model actually does rather than what a demo suggests: custom evaluation suites, MixEval and Needle-in-a-Haystack tests for accuracy, robustness and long-context recall.

  • Agents and tool use

    Giving models safe, typed access to systems. Designed an MCP server mounted inside a production FastAPI application with authentication, structured logging and OpenTelemetry tracing.

  • Applied ML for physical systems

    Machine learning where the target is a measurable quantity: ensemble and deep models for building energy consumption, and a YOLO detector for components in real-world electric circuit photographs.

Experience

Where the work has been done

Roles in order, most recent first. Certificates are linked where they exist.

  1. May 2025 — Present

    Full-time

    Software Engineer — AI/ML

    Linked Matrix

    Building LLM products end to end: retrieval and agent systems, fine-tuning and evaluation on multi-GPU hardware, and the Django/FastAPI services and CI/CD that put them in front of users.

    • Designed and deployed an MCP server mounted inside an existing FastAPI application, enabling tool-based interactions for the company product chatbot. Implemented secure authentication, structured logging and distributed tracing with OpenTelemetry; containerised and deployed with Docker.
    • Built an AI tool that evaluates and improves Upwork proposals using LLMs and historical win/loss proposal data, predicting success rate for new proposals. Integrated it with a Django backend and moved scoring onto Celery for asynchronous, parallel execution — a 60% improvement in speed and efficiency.
    • Designed a custom Python package that lets developers build AI chatbots quickly, and a multi-user chatbot platform on top of it (Django backend, React frontend, Supabase). Deployed bots to Google Cloud Run with CI/CD pipelines for automatic per-bot deployment.
    • Fine-tuned and evaluated large language models with LoRA and vLLM on NVIDIA H100 and A100 GPUs. Measured quality with custom evaluations, MixEval and Needle-in-a-Haystack tests to track accuracy, robustness and long-context understanding.
    • Trained a custom YOLO model on electric circuit images to 95%+ accuracy and built the inference pipeline that detects circuit components in complex real-world photographs.
    • Built an n8n workflow chatbot that pulls JSON from API endpoints and hands it to an LLM agent, producing context-aware answers grounded in live data.
    • Python
    • Django
    • FastAPI
    • MCP
    • Celery
    • LoRA
    • vLLM
    • OpenTelemetry
    • Docker
    • GCP Cloud Run
    • YOLO
    • n8n
    • React
    • Supabase

  2. February 2025 — April 2025

    Internship

    AI/ML Engineer — Internship

    Techtics.AI

    Applied knowledge graphs and Graph RAG to UAE tax and legal regulation, so that legal answers could be traced back to the specific provisions they came from.

    • Built an AI/ML solution for UAE tax and legal regulations using Neo4j, knowledge graphs and Graph RAG.
    • Enhanced legal analysis with LLM-powered, graph-based retrieval over the regulation corpus.
    • Neo4j
    • Knowledge Graphs
    • Graph RAG
    • LLMs
    • Python

    Certificate

  3. May 2024 — January 2025

    FreelanceRemote

    Python and Databases Instructor

    Freelance

    Taught Python, SQL, machine learning and data analysis one-to-one to an international student base.

    • Delivered 230+ online lessons covering Python programming, SQL, machine learning and data analysis.
    • Taught and mentored 40+ students from diverse international backgrounds.
    • Held a consistent 5-star rating for clear explanations, a practical teaching approach and student support.
    • Python
    • SQL
    • Machine Learning
    • Data Analysis

    Profile

  4. February 2023 — April 2023

    Internship

    Internship

    Pak Launch

    Market research and competitive analysis for an early-stage platform.

    • Conducted in-depth market research and competitive analysis to identify growth opportunities and industry trends.
    • Market Research
    • Competitive Analysis

    Certificate

Projects

Systems I have designed, built and shipped

Each one is written up the way a model card is: what it does, how it is built, and what came out of it. Every link goes to something you can open.

Full-stack SaaS

Live in production

Montlify

AI productivity platform

A full-stack SaaS platform bundling 100+ AI-powered tools for writing, research and workflow automation, with AI chatbots, user authentication and subscription management.

How it is built

Django and Django REST Framework serve a React + TypeScript frontend. Celery with Redis runs generation jobs off the request path so long-running LLM calls never block a page. Groq and OpenAI sit behind a single provider layer. The whole stack is containerised and deployed to AWS EC2 behind Nginx and Gunicorn, with GitHub Actions handling CI/CD.

Scope
100+ AI tools in one platform
Async
Celery + Redis job queue
Deploy
Docker on AWS EC2, GitHub Actions CI/CD
  • Django
  • DRF
  • Celery
  • Redis
  • React
  • TypeScript
  • Tailwind CSS
  • PostgreSQL
  • Gunicorn
  • Nginx
  • Docker
  • Groq API
  • OpenAI API
  • GitHub Actions
  • AWS EC2
montlify.com

RAG · Browser extension

Live in production

Clip Insights

AI browser extension for video

A browser extension that turns watching a video into working with it: real-time screenshots, timestamped notes, instant summaries, key-point extraction, a chat that answers questions about the video you are on, and PDF export of everything captured.

How it is built

A Django + PostgreSQL backend serves a React extension UI. Transcript chunks are embedded and indexed in Pinecone, and LangChain runs retrieval over that index so the chat answers from the actual video rather than the model prior. Captures are held locally in IndexedDB and synced to AWS S3. Deployed with Docker behind Nginx, and used by real users in production.

Retrieval
LangChain over a Pinecone transcript index
Offline-first
IndexedDB capture, S3 sync
Output
Notes, summaries and PDF export
  • Django
  • PostgreSQL
  • React
  • JavaScript
  • LangChain
  • Pinecone
  • AWS S3
  • IndexedDB
  • Nginx
  • Docker
clipinsights.comChrome Web Store

RAG · Private inference

Live

Sovereign

AI compliance auditing platform

A self-hosted platform that audits contracts against GDPR, HIPAA and SOC 2 and returns a cited, severity-ranked risk report — every finding pointing back to the clause and the regulation it came from.

How it is built

Retrieval-augmented generation over the regulation corpus, with multi-provider LLM orchestration so the same audit can run against a hosted API or a locally served model. Private inference runs on AMD Instinct GPUs through vLLM, which keeps client contracts inside the deployment. Django REST Framework, PostgreSQL and Celery handle the audit pipeline; the React frontend is deployed on Vercel and the backend on Railway.

Frameworks
GDPR, HIPAA, SOC 2
Output
Cited, severity-ranked risk report
Privacy
Private inference on AMD Instinct GPUs
  • Django
  • DRF
  • PostgreSQL
  • Redis
  • Celery
  • Python
  • LangChain
  • RAG
  • vLLM
  • AMD Instinct
  • Groq API
  • Docker
  • Railway
  • Vercel
  • React
  • TypeScript
  • Tailwind CSS
sovereign-compliance.vercel.app

Full-stack commerce

Live in production

Agento Wear

Production e-commerce platform

A production e-commerce platform with guest checkout, secure authentication, dynamic product management, order processing, payment integration and an admin dashboard.

How it is built

Built on Next.js and TypeScript with Supabase as the backend. Access control is enforced at the database with Row-Level Security rather than only in application code, backed by server-side validation. Transactional email runs through Resend, and the app is deployed on Vercel.

Security
Postgres Row-Level Security + server-side validation
Checkout
Guest and authenticated flows
Ops
Admin dashboard, automated email workflows
  • Next.js
  • TypeScript
  • React
  • Supabase
  • PostgreSQL
  • Supabase Auth
  • Resend
  • Tailwind CSS
  • Vercel
agentowear.com

RAG · Education

Open source

Study Buddy

AI-powered learning platform

A learning platform for 9th–12th grade students preparing for Board and entrance tests. Language models generate and evaluate mock tests, answer questions from the textbook, and adapt to the student.

How it is built

A React frontend over a Django + MySQL backend. Textbook content is chunked, embedded and stored in ChromaDB; LangChain runs retrieval so answers are grounded in the specific textbook a student is studying, not general knowledge. Assets are stored on AWS S3.

Audience
Grades 9–12, Board and entrance prep
Retrieval
LangChain over a ChromaDB textbook index
Recognition
1st of 35+ at the PUCon FYP Exhibition
  • React
  • Django
  • MySQL
  • ChromaDB
  • LangChain
  • AWS S3
GitHub

Machine learning · MLOps

Open source

Energy Optimization ML System

Predicting and reducing building energy use

A machine learning system that forecasts building energy consumption and recommends changes to reduce it, cutting consumption by 15–30% in evaluation.

How it is built

Real-time data integration feeds ensemble models and deep learning models trained and tracked with MLflow. Predictions are served over a FastAPI prediction pipeline and surfaced in an interactive Streamlit dashboard with actionable energy-saving recommendations. Packaged with Docker.

Result
15–30% reduction in consumption
Models
Ensemble + deep learning, tracked in MLflow
Serving
FastAPI prediction API, Streamlit dashboard
  • Python
  • Scikit-learn
  • Pandas
  • NumPy
  • FastAPI
  • MLflow
  • Streamlit
  • Docker
  • Matplotlib
  • Seaborn
  • Plotly
GitHub

Statistics · Visualisation

Open source

Netflix & NBA Data Analysis

Exploratory data analysis

Exploratory analysis of the Netflix catalogue and NBA player datasets, looking for the trends that survive scrutiny rather than the ones that look good on a chart.

How it is built

Statistical analysis and feature engineering in Pandas and NumPy, with the findings developed through Matplotlib, Seaborn and Plotly visualisations and written up as actionable insights.

Datasets
Netflix catalogue, NBA players
Methods
Statistical analysis, feature engineering
Output
Visual analysis and written insights
  • Python
  • Pandas
  • NumPy
  • Matplotlib
  • Seaborn
  • Plotly
  • Scikit-learn
GitHub

More work

Skills

The toolkit, grouped by what it is for

Listed rather than scored. A percentage on a skill bar tells you nothing a project or a repository cannot tell you better.

  • LLM & generative AI

    Retrieval, agents, fine-tuning and evaluation

    • Large Language Models
    • RAG
    • LangChain
    • LangGraph
    • MCP
    • AI Agents
    • Vector Databases
    • LoRA
    • vLLM
    • MixEval
    • Needle-in-a-Haystack
    • Graph RAG

  • Machine learning

    Classical models, deep learning and vision

    • Machine Learning
    • Deep Learning
    • NLP
    • PyTorch
    • Scikit-learn
    • YOLO
    • Feature Engineering
    • Model Evaluation

  • Languages

    What the work is written in

    • Python
    • SQL
    • JavaScript
    • TypeScript

  • Backend & APIs

    Services that carry models to users

    • Django
    • Django REST Framework
    • FastAPI
    • Flask
    • Celery
    • Redis
    • REST APIs
    • Nginx
    • Gunicorn

  • Data & databases

    Relational, document, graph and vector stores

    • PostgreSQL
    • MySQL
    • SQL Server
    • MongoDB
    • Neo4j
    • Pinecone
    • ChromaDB
    • Supabase
    • Database Design
    • Data Warehousing

  • Analysis & visualisation

    From raw data to a defensible conclusion

    • Pandas
    • NumPy
    • Matplotlib
    • Seaborn
    • Plotly
    • Power BI
    • Excel
    • Streamlit
    • Statistical Analysis

  • MLOps, cloud & DevOps

    Shipping, tracking and keeping things up

    • Docker
    • CI/CD Pipelines
    • GitHub Actions
    • AWS
    • GCP
    • MLflow
    • OpenTelemetry
    • Linux
    • Git
    • GitHub
    • n8n
    • Vercel
    • Railway

  • Frontend

    Interfaces for the systems above

    • React
    • Next.js
    • Tailwind CSS
    • HTML
    • CSS
Recognition

Competitions, awards and credentials

Results from hackathons and coding competitions, plus the certifications behind the skills listed above. Everything here links to its proof.

Awards

Award certificates
  • Winner

    AI/ML Hackathon — TechFest

    University of the Punjab

    Won the AI/ML competition at TechFest, held at PUCIT, Lahore. Built a machine learning model that outperformed the other entries on accuracy and efficiency, plus a Streamlit chatbot using the OpenAI API and LangChain.

    Machine learning

  • Winner

    Data Quest Hackathon — PUCon

    University of the Punjab

    Won the Data Quest hackathon at PUCon, competing on agents, retrieval-augmented generation and generative AI.

    Agents, RAG and LLMs

  • Winner

    Web Hackathon — NUTEC

    FAST NUCES

    Won the web hackathon at NUTEC, building a responsive web application under competition conditions and taking recognition for its design and functionality.

    Web development

  • 1st place

    Final Year Project Exhibition — PUCon

    University of the Punjab

    Placed 1st among 35+ projects in the Final Year Project Exhibition at PUCon with Study Buddy, an AI-powered learning platform.

    Applied AI

  • Ranked

    Code Sprint LA 2026 — UCLA

    UCLA

    21st place out of 392 participants across 244 teams in UCLA’s international coding competition.

    Competitive programming

  • Winner

    Harvard CS50x Puzzle Day

    Harvard University

    Winner of Harvard’s CS50x Puzzle Day, an algorithmic problem-solving competition.

    Problem solving

    Harvard certificate

  • Participant

    International AI hackathons

    lablab.ai

    Took part in “The INTERNET OF AGENTS” and “Co-Creating with GPT-5”, building and submitting work on autonomous agents and generative AI.

    Agents and generative AI

    lablab.ai profile

  • Receiving the Data Quest hackathon award at PUCon

    Data Quest — PUCon

    Agents, RAG and LLMs

  • Presenting the final year project at the PUCon exhibition

    FYP Exhibition — PUCon

    1st of 35+ projects

  • Receiving the NUTEC web hackathon award at FAST NUCES

    Web Hackathon — NUTEC

    FAST NUCES

  • CS50x Puzzle Day 2026 certificate from Harvard University, issued to Muhammad Etisam Ul Haq

    CS50x Puzzle Day — Harvard

    Signed by David J. Malan

Certifications

Writing

Notes on what I have built

Write-ups of systems I have shipped and techniques I use, plus articles published on Medium.

On Medium

Contact

Get in touch

Whether it is about a role, a collaboration or a question on something here — send a message and I will reply.

Email
etisamhaq24@gmail.com
Location
Lahore, Pakistan
Languages
English, Urdu, Hindi, Punjabi
Time zone
PKT (UTC+5)

Elsewhere

or email directly