Background
About
I'm an AI Engineer based in Islamabad, Pakistan, specializing in retrieval-augmented generation, agentic orchestration, and production ML pipelines.
My background spans the full stack of applied AI — from fine-tuning BERT-family models for multilingual NLP to shipping cyclic LangGraph agents integrated with WhatsApp, Instagram, Gmail, and calendar APIs. I've built systems that process thousands of documents daily with sub-3s latency and zero-downtime incremental indexing.
Bachelor of Science in Artificial Intelligence
Bahria University · Islamabad, Pakistan
Years Experience
Projects Delivered
Peak F1 Score
RAG Accuracy
Experience
Work Experience
Islamabad, Pakistan
AI Engineer
Stixor Technologies
- Developed and deployed backend and AI agent for autonomous appointment booking system integrating 6 third-party APIs, reducing vendor scheduling overhead by roughly 1.5 hours daily across active sessions.
- Delivered a Vision-RAG pipeline for mining report analysis, replacing fully manual lookup across 25 years of historical reports at sub-3s latency and 93%+ query accuracy.
Islamabad, Pakistan
AI Engineer
WebTech Fusion
- Designed invoice processing pipeline with PyTorch-based OCR and multithreading for concurrent batch operations, processing 3,000+ invoices daily with under 8s per-document processing time, eliminating 12+ hours of weekly manual extraction.
- Fine-tuned BERT, AraBERT, and RoBERTa-large for multilingual sentiment analysis across English, Arabic, and Urdu, achieving 96%+ F1 score across all three languages.
Capabilities
Skills
Generative AI & LLMs
Machine Learning
Backend & Infrastructure
Cloud & APIs
Reasoning
Stack Reasoning
Every tool choice is a tradeoff. Here's how I think about the technologies I reach for most.
LangGraph
OrchestrationI reach for LangGraph when the agent flow has conditional branches, cycles, or requires durable state across long-running tasks. The explicit StateGraph model makes control flow auditable. For simple linear chains, LangChain Expression Language is sufficient.
FastAPI
BackendAsync-first from the ground up. When you're running concurrent LLM calls, webhook handlers, and background tasks, synchronous frameworks add latency and complexity. FastAPI's Pydantic integration also gives you a validation layer at the API boundary without boilerplate.
Milvus
Vector DBChosen when the retrieval query requires scalar filtering on metadata at query time (date ranges, tenant IDs, document types). FAISS is faster and simpler for in-memory corpus sizes under ~1M vectors. Milvus pays its operational overhead when you need ANN search + metadata filters in one atomic query.
PyTorch
MLFor any work involving custom model architectures — modified attention heads, custom loss functions, LoRA adapter merging — PyTorch is the only practical choice. The dynamic computation graph makes debugging model internals tractable. For inference-only work on standard architectures, ONNX Runtime often wins on latency.
Celery + Redis
Task QueueWhen a web-facing endpoint triggers work that takes longer than 3–5 seconds (LLM calls, document processing), the work needs to be decoupled from the request handler. Celery + Redis is my default because Redis is already in the stack for caching and session state, so the broker is 'free'.
Docker Compose
InfrastructureThe sweet spot between 'just run it locally' and 'full Kubernetes overhead'. Most of my projects are multi-service (API + vector DB + task queue + cache) — Compose makes local dev and CI parity tractable without the cognitive overhead of container orchestration.
QLoRA
Fine-tuning4-bit base weight quantization + low-rank adapter training makes fine-tuning a 7B model on an 8 GB GPU tractable. The quality tradeoff (vs full fine-tune) is acceptable for domain adaptation tasks. For tasks requiring deep weight modification (not just output style), full fine-tuning on a larger compute budget is worth it.
Reciprocal Rank Fusion
RetrievalWhen combining dense and sparse retrieval signals, RRF avoids the score normalization problem — you don't need to calibrate cosine similarity scores against BM25 scores, which are in completely different ranges. Works reliably out of the box with no tuning required.