Skip to content
Uzair Azhar
Menu

Uzair Azhar, Python and AI engineer

Data, AI and computer-vision systems that are deployed, measured and maintained.

I take a messy, real problem, such as unreliable data feeds or documents nobody can search, and ship the system that solves it, with the tests, monitoring and documentation that keep it running after handover.

View projectsHire meGitHub

Latest pipeline run

March 2010 drop, finished 3 hours ago

Succeeded
Resolve3 msExtract1.02 sValidate430 msTransform212 msQuality gate31 msLoad1.38 s
rows read
41,511
rows published
40,683
rows quarantined
302
duplicates removed
526
quality score / 100
100.0
run time
3.13 s

Real invoices from a UK online retailer (UCI Online Retail II), processed on this server. 25 of 25 monthly drops loaded.

Open the live dashboard

Selected engineering projects

Each project has a working demo on real or clearly labelled data, a write-up of the design decisions, and its source code.

LiveData engineering

Data Pipeline Observatory

Problem
Monthly exports, a product master and a CRM API never agree. Duplicates, cancellations, stock write-offs and missing customer IDs silently corrupt every report built on them.
Solution
A Python pipeline that validates each row against a data contract, quarantines bad rows with reason codes, loads PostgreSQL idempotently and refuses to publish when a quality gate fails.
Technologies
  • Python
  • pandas
  • PostgreSQL
  • FastAPI
  • Procrastinate
  • Docker

Real data: UCI Online Retail II (CC BY 4.0)

Live demoArchitectureGitHub
LiveAI and LLM appsData engineering

AI CSV Analyst

Problem
Chat-with-your-data tools write SQL nobody reads, answer confidently when the question does not fit the data, and say nothing about the duplicates and gaps that skew totals.
Solution
Profiles the table first, turns each question into a validated query plan (rules first, an optional LLM second), runs it in a locked-down DuckDB sandbox and shows the SQL.
Technologies
  • DuckDB
  • Python
  • function calling
  • FastAPI
  • Next.js

Real data: UCI Online Retail II (CC BY 4.0)

Live demoArchitectureGitHub
LiveAI and LLM apps

Document Intelligence

Problem
Answers from long documents are only useful if you can check them, and invoice data is only useful if it is right. Chatbots over PDFs paraphrase without saying where; invoice readers rarely say when they are unsure; both break on scans.
Solution
Reads PDFs in an isolated process (text layer or OCR). Answers quote cited passages or abstain; invoices are read into fields, checked by arithmetic, and measured against the real records they were drawn from.
Technologies
  • PDFium
  • OCR (ONNX)
  • pgvector
  • PostgreSQL full-text
  • FastAPI
  • Next.js

Your PDF, a NIST publication, or real retail invoices drawn as documents

Live demoArchitectureGitHub
LiveAI and LLM apps

Customer Support Agent

Problem
Order questions are repetitive and costly to get wrong; prompt-only chatbots invent order details, promise refunds the policy does not allow and can be talked into things.
Solution
An agent loop over real invoices with validated tools, refund rules in code, cited policy passages, supervisor approval above £50 and a trace of every step.
Technologies
  • function calling
  • agent loop
  • PostgreSQL full-text search
  • FastAPI
  • Next.js

Real invoices; simulated shipments and policy

Live demoArchitectureGitHub

Next on the roadmap

Listed honestly: these are not live yet and have no demo until they are.

  • Edge AI Benchmark Lab

    PyTorch vs ONNX vs INT8 on CPU: latency, memory and accuracy, measured.

    Planned

  • Invoice Matching Agent

    Reconciles public-sector payments against contract awards and explains exceptions.

    Planned

  • Research Search Intelligence

    From a clinical research question to concepts, MeSH terms and ranked papers.

    Planned

  • Street Scene Intelligence

    Detection, tracking, pose and pedestrian crossing-intent on open street footage.

    Planned

What I build

Python data pipelines
Ingestion from files, databases and APIs, with validation, retries and audit trails.
AI and LLM applications
Question answering over documents, natural-language interfaces to data, tool-calling agents.
Computer vision
Detection, tracking and image analysis, optimised to run on ordinary CPUs.
Automation
Replacing manual spreadsheet and inbox work with scheduled, observable jobs.
APIs and backend systems
FastAPI services with typed contracts, migrations and sensible limits.
Dockerised deployments
Reproducible builds, health checks and rollbacks on a single VPS or the cloud.
Data processing
Cleaning, reconciling and profiling messy tabular data at millions of rows.
Research and scientific AI
Medical-imaging models and literature tooling, written up properly.

Engineering capabilities

Languages and data

  • Python
  • SQL
  • pandas
  • DuckDB
  • PostgreSQL
  • pgvector
  • JSON / CSV

Backend

  • FastAPI
  • Pydantic
  • SQLAlchemy
  • Alembic
  • REST APIs
  • job queues

Machine learning

  • PyTorch
  • ONNX Runtime
  • quantisation
  • computer vision
  • medical imaging

LLM systems

  • RAG
  • embeddings
  • function calling
  • Groq
  • llama.cpp

Delivery

  • Docker
  • Linux
  • nginx
  • Git
  • GitHub Actions
  • Ansible
  • TypeScript / Next.js

Research and specialised work

Medical imaging, skin-lesion classification

SHEL: A Knowledge-Guided Hybrid Representation Learning Framework for Efficient Multiclass Skin Lesion Classification

Manuscript under review at a Springer Nature journal. Details and results will be linked here once it is published.

Related work on this site: research-literature search with medical terminology (Research Search Intelligence) and computer vision on video (Street Scene Intelligence), both on the roadmap.

How I work

  1. Problem

    A written statement of what goes wrong today, who notices and what fixing it is worth.

  2. Architecture

    The smallest design that solves it, with the trade-offs written down before any code.

  3. Implementation

    Typed, reviewed code in small increments you can run from the first week.

  4. Testing

    Automated tests on every change, including checks against real data where it exists.

  5. Deployment

    Container images built and scanned in CI, released with a health check and automatic rollback.

  6. Monitoring

    Structured logs, health checks and alerts, so problems are noticed before users report them.

Have a data or AI problem that needs to work in production?

Freelance projects, contracts and full-time roles. Send a short description of the problem and I will reply with questions or a first plan.

Start a conversationuzairazhar@gmail.com