👋 About Me

AI/ML Engineer and Data Scientist with 3+ years building production machine-learning and generative-AI systems. I work on LLM evaluation and safety, Retrieval-Augmented Generation, and AI agent systems, and I own the full path from data pipeline to deployment: architecture, delivery, and operations.

• 🔭 Building Bryge.io, a multi-tenant industrial IoT analytics platform, as AI/ML Engineer at Datality (London, remote)
• 🤖 I build LLM apps and I measure them. My portfolio has 12 evaluation and safety projects.
• 🌱 Working across RAG, AI agents, MLOps, and causal/classical ML
• 🎓 MSc Artificial Intelligence @ LUMS · BSc Data Science @ FAST-NUCES
• 💬 Ask me about LLMs · RAG · LangChain · Evaluation harnesses · MLOps
• 📫 Reach me at taimour.a.karim@gmail.com
• ⚡ Fun fact: I like Golf ⛳

🏆 Achievements & Recognition

Beginner-Data-Science-Projects · 2,900+ ⭐ · ranked Top 3 (Notebooks), Top 9 (Training), Top 10 (Educational) by growth acceleration

🚀 Featured Projects

🧪 clinical-llm-bias-audit Reproducible fairness-audit framework for clinical LLMs. Introduces the Geographic Disparity Index (GDI); from-scratch stats (Wilcoxon + BCa bootstrap). Multi-provider LLM · FastAPI · Streamlit	📄 sec-rag-analyst Production RAG over SEC 10-K filings: hybrid BM25+dense retrieval, RRF fusion, cross-encoder rerank, inline citations, labeled eval. Claude · FAISS · BGE · Docker
🧩 rag-architectures 13 RAG architectures benchmarked on one corpus, naive through GraphRAG, RAPTOR, and agentic. Query-transform methods score 0% on multi-hop; agentic and RAPTOR reach 83%. sentence-transformers · FAISS · Bedrock	⚙️ credit-default-mlops End-to-end MLOps: versioned data → tracked model → CI quality gate → drift detection → instrumented serving. DVC · MLflow · Evidently · Prometheus
🖥️ model-serving Production inference platform. Dynamic micro-batching measured at 8.0x throughput (121 to 966 RPS), plus canary and shadow deploys and load shedding, all load-tested over real HTTP. FastAPI · asyncio · Prometheus · Docker	🎯 uplift-targeting-engine Causal uplift modeling: S/T/X/R meta-learners predict per-user incremental effect, scored with Qini and policy value. Beats random targeting at a fixed budget. Meta-learners · XGBoost · econml · FastAPI

+ a deep LLM evaluation & safety cluster: llm-gateway · realtime-ml-pipeline · red-teaming · hallucination benchmarks · LLM-as-judge · prompt-regression CI · agent-eval-harness · llm-guardrails · llm-observability · MCP server. See them all at taimour-a-karim.vercel.app.

🛠️ Languages & Tools

📊 GitHub Analytics

🤝 Connect With Me

       

⭐️ From tkarim45. Every result stated against a baseline.