AI/ML Engineer & Data Scientist in Islamabad, finishing an MS in Data Science at Quaid-e-Azam University. I build LLM agents, RAG systems and AI products that people use, and I measure whether they work: held-out test splits, ablations, confidence intervals, and honest write-ups of what did not. 30+ projects so far, for clients and for this portfolio.
🟢 Open to AI/ML engineering, data science and ML research roles: onsite or hybrid in Islamabad/Rawalpindi, or remote.
VSARA, a healthcare platform of four connected apps on Google Play (built at Syslab Technologies): patients talk to an AI assistant by text or voice that answers from medical content (RAG), helps with symptom triage and schedules appointments through agent workflows. React Native · Django REST · Supabase · LLM APIs · speech · n8n. A/B tests of prompts and retrieval raised diagnostic response accuracy by 23%; data and model fixes cut the error rate by 18%. SARA Life · SARA Care · SARA Pharmacy · Flexa (published for Malaysia only) · demo video
| Project | What it is | Result |
|---|---|---|
| Data Analyst Agent | Self-correcting text-to-SQL agent (LangGraph): schema linking, read-only SQL guard, literal-value check, fix loop | 60.7% → 66.7% execution accuracy on held-out BIRD questions (+6.0 pts, 95% CI [+1.3, +11.3]) |
| E-mail Experiment & Uplift | Full readout of a 64,000-customer randomized experiment: SRM, Holm-corrected effects, CUPED, peeking, uplift models, IPW policy value | Men's e-mail +$0.77 spend per customer; blanket sending beats model targeting below ~$0.23/e-mail |
| Customer-Ops Agent | Support agent that takes real actions behind a gateway: policy in code, confirmation codes, human approval via interrupts, idempotent refunds | 30 held-out scenarios: 100% success, 0 policy violations vs 90% and 3 violations (incl. a double refund) for a prompt-only agent |
| Financial RAG Lab | Which RAG choices matter? Chunking, metadata scope, BM25/dense/hybrid, rerankers, HyDE on FinanceBench | Retrieval is the bottleneck: answer accuracy naive 24% → best config 47% → +HyDE 61% (+14 pts, 95% CI [+5, +23]); oracle 84% |
| Literature Review Agents | Multi-agent reviewer (arXiv + OpenAlex) where every finding quotes its source verbatim and a critic checks citations | 97% sentence support for both; citation precision 87% vs 80% (CI includes zero); judge fixed: wrong-citation catch rate 48% → 90% |
Built for clients, so the code is not public; happy to walk through any of them.
| Area | Projects |
|---|---|
| LLM apps & agents | AI e-mail template generator (React Native + GPT-4o-mini) · multi-agent content team (CrewAI) · autonomous data analyst (LLM-written pandas in a sandbox) |
| Computer vision | MedVision: chest X-ray pneumonia detection with Grad-CAM (CNN vs ResNet50 vs EfficientNetB0) · DetectPro: YOLOv8 object detection, 30+ FPS on GPU |
| Data science | Stroke-risk prediction (imbalanced, SHAP) · ChurnGuard (7,043 customers, 0.847 AUC) · FraudShield (1.7% fraud rate, SHAP + LIME) · TimeVision (ARIMA/SARIMA/Prophet/LSTM forecasting) · RecEngine (SVD, KNN, hybrid; NDCG) |
| NLP | TextMiner: TextRank/LSA vs T5 summarization (ROUGE), LDA vs BERTopic topics |
| Mobile | JSuiteCloud, an ERP app on the Apple App Store (React Native, TypeScript) |
Also: Qanoon Assist (private), a legal-information RAG over 117,365 sections of Pakistani law in English, Urdu and Roman Urdu (hit@5 1.00, faithfulness 0.944); and my MS thesis UrduComicLab, complete Urdu stories → right-to-left comics with open-weight models (spoken lines printed 0.49 → 1.00 vs baseline).
LLM agents & RAG: LangGraph, function calling, multi-agent pipelines, hybrid retrieval, cross-encoder reranking, evaluation & guardrails ML & stats: PyTorch, LightGBM/XGBoost, Vision Transformers, A/B testing, causal inference, uplift modeling Engineering: Python, TypeScript, SQL, FastAPI, Django REST, React, React Native, Docker, pytest
The quickest way is e-mail or LinkedIn. CVs are on my portfolio.