M. Ikram Ullah

About

Engineering that is measured, not just demoed

AI/ML engineer and data scientist in Islamabad, finishing an MS in Data Science at Quaid-e-Azam University. I like systems that take real actions safely, evaluations with held-out splits and confidence intervals, and results reported honestly, including what did not work. Open to onsite and hybrid roles in Islamabad/Rawalpindi and remote roles anywhere.

Selected work

Projects

Every public project has tests, a launcher for Windows and macOS, and an evaluation with held-out data.

Measured

Data Analyst Agent

Ask a database a question in plain English. The agent links the schema, writes SQL, runs it read-only, catches silent mistakes and fixes its own query.

60.7% → 66.7% execution accuracy on held-out BIRD questions
+6.0 points, 95% CI [+1.3, +11.3]
1.09 LLM calls per question

LangGraphtext-to-SQLGeminiDuckDBevaluation

Code

Evaluation in progress

Customer-Ops Agent with Guardrails

A support agent that cancels orders and issues refunds behind a gateway the model cannot talk its way past: policy in code, customer confirmation, human approval and idempotent refunds.

45 tau-bench-style scenarios incl. prompt injection and tool failures
Supervisor approval via LangGraph interrupts
No double refunds on payment timeouts

LLM agentsguardrailshuman-in-the-loopfunction calling

Code

Evaluation in progress

Literature Review Agents

A team of agents searches arXiv and OpenAlex and writes a short review where every sentence is cited, every finding quotes its source verbatim, and a critic checks the citations.

Planner → searcher → reader → writer → critic → reviser
Verbatim-quote grounding checked in code
Independent, validated LLM judge

multi-agentLangGraphresearchcitations

Code

Evaluation in progress

Financial RAG Lab

Which RAG choices actually matter? Chunking, metadata filters, BM25 vs dense vs hybrid, cross-encoder reranking and HyDE, measured on the exact evidence page in 84 company filings.

FinanceBench: 150 questions, 84 filings, 12,013 pages
45 retrieval configurations, dev/test split
Answers graded by a validated judge

RAGretrievalrerankingbenchmarking

Code

Measured

E-mail Experiment & Uplift Modeling

A full readout of a 64,000-customer randomized experiment: validity checks, effects with honest uncertainty, CUPED, peeking, uplift models and a dollar-valued targeting decision.

Men's e-mail: +$0.77 spend per customer (95% CI $0.49–$1.05)
Peeking 10 times → 18% false positives
Blanket sending beats targeting below ~$0.23/e-mail

A/B testingcausal inferenceupliftLightGBM

Code

MeasuredPrivate

Qanoon Assist

A legal-information assistant for Pakistani statutes in English, Urdu and Roman Urdu, citing the exact section behind every sentence.

6,664 laws · 117,365 sections from 5 government portals
Hybrid BM25 + BGE-M3 + reranker
hit@5 1.00, faithfulness 0.944

RAGUrdu NLPFastAPIReact

Private repository; demo video on request.

Research

Thesis and publications

MS thesis · in progress

UrduComicLab: complete Urdu stories → right-to-left comics

Turns complete Urdu children's stories into right-to-left comic books with correctly shaped Nastaliq dialogue, narration captions and sound effects, using open-weight models only (Gemma-2-9B parser, SDXL with IP-Adapter, Grounding DINO bubble placement, HarfBuzz lettering).

Baseline Thesis method

Spoken lines printed

0.49 → 1.00

Dialogue F1 vs gold

0.63 → 1.00

Speaker accuracy

0.83 → 1.00

Character recall

0.78 → 1.00

Face consistency (DINOv2)

0.54 → 0.67

15 paired comics (3 stories × 5 seeds), paired bootstrap confidence intervals; about 35 pinned open-weight models benchmarked.

Paper · pending publication

CNN vs Vision Transformer on imbalanced FER2013

Compared an enhanced ResEmoteNet CNN with squeeze-and-excitation blocks against ViT-Base/16 on FER2013 (35,887 images, 7 classes, 16.5:1 class ratio), with class-balanced focal loss, two-stage training and CutMix, Mixup and RandAugment.
Showed that ViT self-attention avoids minority-class collapse (64.9% vs 0% accuracy on Disgust), while the CNN with channel attention was stronger overall on low-resolution faces.

0%

ViT accuracy on the rarest class (Disgust)

0%

CNN on the same class

Career

Experience & education

AI/ML Engineer · Syslab Technologies
Oct 2024 – Nov 2025

Islamabad, Pakistan
- Architected VESARA, an agentic healthcare voice and chat assistant combining a RAG pipeline (FAISS + Hugging Face sentence-transformers) with Whisper speech recognition and neural text-to-speech.
- Built n8n automation workflows orchestrating multi-step agent actions (appointment scheduling, symptom-triage routing, EHR lookup), improving response throughput by 31%.
- Improved diagnostic response accuracy by 23% through systematic A/B testing of model configurations, prompt strategies and RAG retrieval settings (chunk size, retrieval depth).
- Built an automated model-monitoring pipeline (accuracy, precision, recall, F1, latency) with real-time Power BI dashboards for early detection of model drift.
- Investigated model failures through error logs, edge cases and interaction data; targeted retraining and data-quality fixes reduced the error rate by 18%.
- Processed 50K+ patient interaction records with Python and SQL; validated outputs with product managers, QA and healthcare providers against clinical standards in an Agile team.
Freelance AI/ML Engineer · Self-employed (direct clients)
Feb 2023 – Present

Remote
- Delivered 15+ end-to-end AI/ML projects for healthcare, e-commerce and fintech clients, from scoping and data collection to deployment and handoff.
- Built RAG chatbots and document Q&A systems (LangChain, FAISS, Hugging Face embeddings) that answer from clients' internal knowledge bases with cited, grounded responses.
- Built churn, demand-forecasting and fraud-detection pipelines (scikit-learn, XGBoost, LightGBM) with MLflow experiment tracking; reported results through Tableau and Power BI dashboards.
- Developed n8n workflows connecting LLM agents to CRM, e-mail and scheduling APIs, cutting clients' manual operational work by 30–45%.
- Fine-tuned Hugging Face models (BERT, DistilBERT, Mistral-7B) for text classification, sentiment analysis and named-entity recognition; deployed solutions with FastAPI and Docker.
MS Data Science · Quaid-e-Azam University, Islamabad
Feb 2025 – Present

Coursework: Machine Learning, Deep Learning, NLP, Big Data Analytics, Statistical Modeling, Data Visualization. Thesis: UrduComicLab.

Certifications: Machine Learning Specialization · Deep Learning Specialization · Natural Language Processing · Big Data Analytics

Toolbox

Skills

LangGraph (state machines, interrupts, checkpointers)LangChainfunction callingmulti-agent pipelineshybrid retrieval (BM25 + dense, RRF)cross-encoder rerankingHyDEQdrantFAISSChromaDB

Contact

Let's talk

Hiring for an AI/ML engineering, data science or research role? I would be glad to walk you through any of these projects.

LinkedIn GitHub

AI/ML Engineer CV Data Scientist CV ML Researcher CV

Engineering that is measured, not just demoed

Projects

Data Analyst Agent

Customer-Ops Agent with Guardrails

Literature Review Agents

Financial RAG Lab

E-mail Experiment & Uplift Modeling

Qanoon Assist

Thesis and publications

UrduComicLab: complete Urdu stories → right-to-left comics

CNN vs Vision Transformer on imbalanced FER2013

Experience & education

AI/ML Engineer · Syslab Technologies

Freelance AI/ML Engineer · Self-employed (direct clients)

MS Data Science · Quaid-e-Azam University, Islamabad

Skills

Let's talk

Data Analyst Agent

What I built and measured

Customer-Ops Agent with Guardrails

What I built and measured

Literature Review Agents

What I built and measured

Financial RAG Lab

What I built and measured

E-mail Experiment & Uplift Modeling

What I built and measured

Qanoon Assist

What I built and measured