About
AI & Data Engineer · Rawalpindi, Punjab, Pakistan
I build the layer between raw data and something a person can actually use: retrieval pipelines, LLM-backed services, and the backend and cloud plumbing that keeps them answering under real load.
Most of my work has started as a research idea and ended as a deployed system: an Urdu ASR correction pipeline that cut word error rate without retraining the acoustic models, an investment advisory platform built on LangGraph, an inference pipeline wired to Prometheus, Grafana and drift-triggered retraining.
I care about the unglamorous parts: clear interfaces, observable services, and delivery that stays incremental. Those are what decide whether a model still works six months after the demo.
Generative AI & LLMs
Prompt design, evaluation, and bounded refinement in production paths.
RAG systems
Chunking, embeddings, hybrid retrieval, and grounded answer synthesis.
Agentic AI
Stateful graphs, tool calling, and orchestration that stays debuggable.
Data engineering
Orchestrated ETL, warehousing, and data-quality checks that fail loudly.
Backend & cloud
Typed APIs, containerised services, and infrastructure as code.
MLOps
CI/CD for models, metric instrumentation, alerting, and retraining loops.
01
A clear problem statement and a well-named interface outlive any clever implementation. I write the boundary first and let the internals stay replaceable.
02
If a system cannot tell me it is degrading, it is not finished. Metrics, structured logs, and alerting are part of the feature, not a follow-up ticket.
03
Small, reversible releases beat a long-lived branch. It keeps feedback fast and keeps the blast radius of any single mistake small.
04
Latency against throughput, consistency against availability, build against buy. I would rather write the trade-off down than pretend it was not made.