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.

Open to Senior AI Engineer, Applied AI and ML Engineering roles · London or remote · mohammad.hijjawi1997@gmail.com

• AI systems that reach production. Agents, retrieval, evaluation and serving, delivered end to end and deployed on Cloud Run, AWS and air-gapped on-premise.
• Fast, careful work in codebases I did not write. 320+ merged pull requests across 75+ open-source AI/ML projects, each one reviewed and accepted by that project's maintainers, including vLLM, NVIDIA TensorRT-LLM, Hugging Face, Ray, Ultralytics and Langflow.
• Engineering discipline. I find the root cause, keep the diff small, and add a regression test that fails before the fix and passes after. I work inside each team's CI, linters, changelog and review process, and respond to review quickly.
• Measurement before claims. Evaluation harnesses, baselines and statistics, from production agents to two accepted EMNLP 2026 papers.
• Raising the level of the people around me. Architecture and code review for 130+ engineers building LLM, RAG and agent systems.

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.

320+ merged pull requests across 75+ open-source AI/ML projects (as of 31 October 2026), almost all bug fixes with regression tests. Credited by name in 40+ release notes, including Roboflow, Haystack, Ultralytics, Langflow, Unsloth, AnythingLLM, Prefect and xinference.

Most merged: Ultralytics (21) · Kornia (20) · Roboflow Supervision (15) · llm-gateway (15) · DeepTutor (11) · SAHI (11) · doctr (10) · SurfSense (9) · xinference (9) · Unsloth (8)

Also merged in Haystack, Prefect, statsmodels, Genesis, InvokeAI, Feast, darts, lancedb, ESPnet, FunASR, txtai, open_clip, Gymnasium, AnythingLLM, private-gpt and others.

All merged pull requests

• AI Instructor and Engineering Mentor, Multiverse, London. Hands-on architecture and code review with 130+ engineers building LLM, RAG and agent systems inside their own organisations.
• 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). Senior Machine Learning Engineer at Ideas Beyond Borders (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.