Senior AI Engineer · agentic systems, LLM architecture, applied NLP

London · LinkedIn · ORCID · mohammad.hijjawi1997@gmail.com

I design, build and deploy AI systems end to end: data pipelines, model work, multi-agent orchestration, retrieval, evaluation, and production serving on Google Cloud Run and AWS.

My core specialism is agentic AI: LangGraph state machines and checkpointers, MCP tool layers, retrieval-augmented generation, guardrails as architecture, human-in-the-loop interrupts, and the evaluation harnesses that make an agent defensible in production. I have shipped a six-agent enterprise assistant to Cloud Run and re-delivered it air-gapped on-premise.

since-cutoff finds which of your exact dependency versions your coding model writes wrong, and fixes them with short AGENTS.md notes that a type checker verifies. It turns the idea behind our EMNLP 2026 temporal-isolation paper (what a model knows about the world after its cutoff) into a tool for real code.

• The problem, measured across vendors: of 36 widely used Python AI libraries, even a model trained up to June 2026 predates a public API break in 20; for early-2025 models it is 33. Results for 21 models from 8 vendors.
• The fix, measured per project: on a sample project, 8 type-checked notes took Claude Haiku 4.5 from 14% to 57% correct on held-out tasks (one project, measured with 0.1.0); 0.3.0 adds the statistics and baseline comparisons to check such results on your own code.
• Where it runs: CLI (uvx since-cutoff scan), an MCP server in the official MCP Registry, a Claude Code plugin, a GitHub Action and a pre-commit hook. Python, MIT, documented in English, Chinese, Spanish and French.

120+ merged pull requests across 40 open-source AI/ML projects (as of October 2026), mostly bug fixes with regression tests.

• Haystack (deepset): credited among the contributors to Haystack 3.3.0, with fixes to OpenAI-format tool-call parsing in ChatMessage, QueryExpander and TextCleaner.
• Kornia (18 merged) and Ultralytics (14): computer vision geometry, image I/O and utilities.
• Docling, Roboflow Supervision, Genesis, Feast, SAHI, ONNX, statsmodels, AnythingLLM and others.

All merged pull requests

• AI Instructor and AI Engineering Mentor, Multiverse, London. Hands-on architecture and code review with 130+ engineers building LLM, RAG and agent systems inside their own organisations, including the NHS and the University of Cambridge.
• Instructor, Great Learning, with Johns Hopkins University and UT Austin. Designed the reference implementations for a 14-week agentic AI programme: LangGraph agentic RAG, MCP-native ReAct agents, red-teaming, multi-agent systems, HITL workflows and evaluation.

Previously Data Scientist at EMCOR UK (now OCS). Lead AI and Data Engineer at Ideas Beyond Borders since 2019 (Bayt Al Hikma NLP ranking; content from that programme has been read 500 million+ times).

Insurance claims graph. Orchestrator, specialised A2A worker agents, aggregator, critic and verdict nodes. Policy and document tooling from two FastMCP servers over streamable HTTP, with per-node checks and a full audit trail per claim.

Sentinel Finance. ReAct agent over five MCP tools (market data, news, sentiment, private-document RAG) with dual input/output guardrails, per-claim source attribution and an audit log.

Returns and refunds agent. Red-teamed an unguarded action-taking agent (prompt injection, PII exfiltration, tool poisoning, over-refund), then rebuilt safety as graph architecture: guardrail nodes, DeBERTa-v3, Detoxify, Presidio and server-side caps.

Clinical data assistant. Natural-language SQL with LangGraph interrupts: safe reads auto-execute, writes pause for human approval, unsafe queries are rejected, state is checkpointed across the interrupt.

SCREENDEX. Text-first OCR and keyframe index replacing raw video frames. Cut multimodal input cost by 75.6% on one frontier model and 82.8% on another at comparable accuracy, with a frozen evaluation harness (EMNLP 2026 Industry Track).

• EMNLP 2026, Main Conference (CORE A)*, accepted. What Does a Language Model Know If It Has Never Seen the Future? Temporal Isolation as a Natural Experiment in Knowledge Attribution. Co-author.
• EMNLP 2026, Industry Track (CORE A)*, accepted. Cheaper Than Frames: An OCR+Keyframe Index as a Perception Layer for GUI-Video Question Answering. Co-author.
• IEEE/ACS AICCSA 2025. TextAge: Evaluating Large Language Models for Historical Text Dating. Led the LLM strand: zero-shot and chain-of-thought evaluation, plus LoRA fine-tuning of LLaMA 2 7B and Mistral 7B.
• MICAD 2025, Springer. A Multimodal Deep Learning Framework for Mycetoma Classification: Integrating Vision Transformers, Medical Language Models, and Transfer Learning. ISBN 978-981-95-7425-4.
• WikiJournal of Humanities 2025 (sole author, peer-reviewed): Integrating Generative AI in Wikipedia Classrooms · Beyond Basics: Advanced PetScan Techniques for Improving Wikipedia Articles.

Teaching beyond the day job: 8-week agentic AI course at the University of Hertfordshire; guest lectures at the University of Birmingham; AI strategy adviser to Wikimedia UK; AI panel, Wikimania 2026, Paris.

MSc Data Science, High Distinction (81.4%, top 3 of 150+), University of Birmingham, 2023-2024. Chevening Scholar (UK Government). Elected Best Student Representative, College of Engineering and Physical Sciences.

BSc Engineering, An-Najah National University, 2019.