Muhammad Ahmad Mujtaba
I build production AI systems end-to-end, from raw data to deployed, optimized services.
Who I am
AI engineer with three years of experience building production AI systems: LLM fine-tuning, RAG pipelines, and full-stack ML backends with complex optimization layers. I've worked as the sole engineer on end-to-end projects for international clients, handling everything from raw data to deployed services. I hold a B.Sc. in Artificial Intelligence.
I care about the parts that don't fit on a slide: handling sparse and messy data, cutting inference latency, indexing high-dimensional vectors, and shipping services that hold up under real load, not just notebooks that run once.
Tools I build with
The technologies behind the systems below, grouped by where they sit in the pipeline.
Problems, and how I solved them
Not a list of job duties. Each item is a real constraint I hit, what I did about it, and the result it produced.
Contract AI Engineer
- constraint
A client needed data-driven pricing across many SKUs, with explanations they could trust.
approachSole ML engineer building the pricing recommender from the ground up: Supabase ingestion → SKU classification → STL decomposition.
resultA full pipeline carrying raw catalog data through to clean, model-ready demand series.
- constraint
Forecasting alone doesn't set a price; the recommendation has to be defensible per SKU.
approachLightGBM demand forecasting (Tweedie + quantile regression), elasticity simulation, and a two-phase price optimizer (LP + SLSQP).
resultA production price recommender with per-SKU explainability behind every recommendation.
Contract AI Engineer
- constraint
A RAG recommender was embedding raw objects, making retrieval slow: 15–20s on short queries and 50–60s on larger ones.
approachDesigned embeddings around semantic taxonomies instead of raw objects and re-architected the Qdrant retrieval layer.
resultCut query times to 5–7s (short) and 10–13s (large), roughly a 3–5× speedup.
- constraint
A monolithic setup capped throughput at 4–6 concurrent users.
approachMigrated the same system to a microservice architecture with async request handling.
resultScaled to 40–50 concurrent users.
- constraint
Source data was messy, unstructured, and scattered across the web.
approachBuilt end-to-end data pipelines: web scrapers, LLM-based JSON cleaning/structuring, and entity extraction (skills, education, tasks) before embedding.
resultClean, structured, embeddable data feeding the retrieval system reliably.
Junior AI Engineer
- constraint
Training was manual and didn't reproduce reliably.
approachDesigned and maintained automated training pipelines on DigitalOcean and physical GPU servers.
resultRepeatable, hands-off training runs.
- constraint
Clients needed custom NLP, generative, and computer-vision models behind real services, not off-the-shelf APIs.
approachBuilt and deployed NLP, generative AI, and computer-vision models behind APIs, including object detection, classification, and LoRA/QLoRA fine-tuning workflows.
resultShipped tailored models into client products.
- constraint
Products needed retrieval-augmented and chained-LLM features.
approachContributed LangChain and RAG-based product integrations alongside the core engineering team.
resultRAG features integrated into shipping products.
AI Engineer Intern
- constraint
Early-career: learning to ship ML into real services.
approachTrained and deployed models, handled data preprocessing, integrated vector databases, and built APIs with FastAPI.
resultFoundation in the full model-to-API lifecycle.
The numbers, and why they were hard
Real outcomes from production work. The pricing curve is illustrative; the latency and scaling figures are from shipped systems.
3–5× faster queries
Embedding around semantic taxonomies instead of raw objects, plus a re-architected Qdrant layer.
Monolith → microservices
Async request handling lifted a hard concurrency ceiling without rewriting the models.
concurrent users
Finding the profit-maximizing price
The optimizer doesn't scan every price. A two-phase LP + SLSQP search explores a bounded band around each SKU's current price (shaded), under margin floors and competitor caps, to maximize profit, not revenue. Demand falls as price rises; profit peaks where volume and margin trade off, and turns negative below unit cost.
Selected work
Projects I can show in full, and client work I can only describe at a high level.
IHS
Intelligent Home Surveillance: real-time fall detection, hazard recognition, child-safety monitoring, and RAG-powered first-aid guidance.
Demand-Forecasting Agent
Agentic ML pipeline for retail demand forecasting on M5-format data. Deterministic branch logic, data-relative thresholds, and a critic that gates acceptance — not a rubber stamp.
Taxonomy-Driven RAG Recommender
Semantic-taxonomy embeddings on Qdrant with async microservices: 3–5× faster retrieval, scaled from 4–6 to 40–50 concurrent users.
AI Profit Optimizer
Production pricing engine for an e-commerce catalog. LightGBM demand forecasting (Tweedie + quantile) over STL-decomposed history feeds a two-phase LP + SLSQP profit optimizer bounded by margin guardrails and competitor-aware caps. Adds champion/challenger model gating (WMAPE), a self-tuning feedback loop, and DB-backed runtime policy across sell- and buy-side pricing.
Generative · NLP · Vision Models
LoRA/QLoRA LLM fine-tuning, custom CV/NLP models, and automated GPU training pipelines delivered into client products.
Let's build something
Open to freelance projects and full-time AI roles. The fastest way to reach me is email.