Work

Case studies

Two projects written up in depth, and the shipped work behind them. Each one starts with the constraint that shaped it.

Case study — full-stack AI

ShopAssist AI

A retail assistant that answers product and order questions without inventing anything. The hard part wasn't retrieval — it was making a 70B model refuse to guess when the catalogue had no answer.

Constraint
26K+ products, live order data, and zero tolerance for a hallucinated price or delivery date.
Approach
ChromaDB vector search for catalogue questions, Pandas order lookup for account questions, and an agentic router that picks between them per turn.
Grounding
Strict context grounding on Groq Llama 3.3 70B, order validation before any account response, PII masking, and source-verified answers.
Pipeline
Automated ingestion with quality checks over third-party APIs, containerised and deployed on Azure via GitHub Actions.
Demo video — assistant answering a product and an order question —
26K+
Products indexed
PythonFastAPIReactLangChainChromaDBGroqDockerAzure
Code
Case study — real-time ML

Real-time beat & tempo tracking

Beat tracking research runs offline on a workstation. I wanted it running live on a device, which meant replacing the inference method rather than optimising it.

Constraint
Particle filtering was accurate but far too slow and memory-hungry for embedded and edge deployment.
Approach
Replaced the particle filter with a learned temporal model — a TemporalConvNet using causal convolutions so it never looks ahead.
Result
104.86× speedup end to end, inference from 200ms down to 2ms, memory from 50+MB to under 10MB, model size 778KB.
Accuracy
F-measure held at 0.82 for beats, trained on 1,696 samples from the Ballroom and GTZAN datasets.
Demo video — live beat detection over an audio track —
104×
Faster than the baseline
PythonPyTorchTensorFlowNumPylibrosamadmom
Code

Experience

2020 — 2026
Cuba Buddy
12/2024 — 03/2026Berlin, Germany

AI Application Engineer

Built AI data pipelines automating leads, reducing manual work 70%.
Developed Python automation processing 500+ leads monthly via APIs.
Implemented behaviour-driven automation improving engagement 30%.
Created integrations connecting 6 platforms (payment, booking, CRM).
Built multilingual web apps serving 10K+ international users.
Set up GTM/GA4 analytics, reducing ad costs 20%.
PythonJavaScriptFastAPIDocker
Alchemative
12/2020 — 09/2023Lahore, Pakistan

Frontend Developer

Built 30+ Shopify Plus storefronts, improving performance by 40–50%.
Built a payment gateway integrated across 300+ merchants processing 85K+ transactions.
Automated order processing via Python APIs for 15+ merchants.
Built a multi-courier system supporting 9 logistics providers with automated routing, cutting processing costs 30% and delivery time by a day.
Analysed e-commerce data driving 10–15% revenue growth via product recommendations.
PythonDjangoJavaScriptSQLShopify

Other work

shipped at Cuba Buddy & Alchemative
Lead & integration pipelines
Behaviour-driven automation connecting six platforms — payment, booking, CRM — into one flow. Cut manual work 70% and lifted engagement 30%.
PythonFastAPIDocker
Live demo
Payment gateway integration
A gateway integrated across 300+ merchants processing 85K+ transactions, with automated order processing for 15 of them.
PythonDjangoSQL
Multi-courier routing
One dispatch system across nine logistics providers with automated routing. Cut processing costs 30% and delivery time by a day.
PythonJavaScriptPostgreSQL
Shopify Plus storefronts
30+ storefronts built and tuned, improving performance 40–50%, with product recommendations driving 10–15% revenue growth.
JavaScriptShopify
Multilingual web apps
Localised booking and content apps serving 10K+ international users, with GTM/GA4 analytics that reduced ad spend 20%.
JavaScriptPython
Live demo

Want the detail behind any of these?

The fastest way to reach me is email. I read everything.

Get in touch
Berlin, Germany · +49 155 6611 8078