Mohammad Najeeb — AI Engineer: RAG pipelines, agentic systems and sovereign LLMs
Mohammad Najeeb
AI Engineer · M.Sc. Computer Science @ Saarland University
I build production AI systems combining Generative AI, agentic systems, LLM pipelines (RAG, fine-tuning), vector databases, and full-stack engineering.
Press Ctrl+P to jump to any file, Ctrl+Shift+P for the command palette, or type help in the terminal below.
# README.md
Mohammad Najeeb
AI Engineer | M.Sc. Computer Science @ Saarland University | Generative AI & LLMs | RAG Systems | Agentic Workflows | AI Applications
I build production AI systems combining Generative AI, agentic systems, LLM pipelines (RAG, fine-tuning), vector databases, and full-stack engineering.
What I do
At ignite AI GmbH I work on the unglamorous half of enterprise AI — governance policies, PII anonymisation layers, and on-premise LLM deployments for clients who cannot let their data leave the building. At EduTech UdS I am lead developer on the SRL Course Platform, an adaptive learning system built on React 19, Node/Express and PostgreSQL with pgvector.
The thread connecting them is retrieval. Most of my work ends up being some variant of the same question: how do you get the right context in front of a model, cheaply, provably, and without leaking anything you shouldn't. Hybrid search, Medallion ETL, semantic chunking, HNSW indexes, RRF — these are the tools I reach for most.
Before Saarbrücken I read Mathematics at Aligarh Muslim University, which is where the interest in the underlying statistics came from, and spent two years shipping front-ends professionally. I co-founded Olympia Academia in 2018 — now a ~200-member student scientific community that has hosted speakers from Caltech and NASA — and I still sit on its advisory board as technical lead.
Currently
Artificial Intelligence Engineer at ignite AI GmbH — Aug 2026 — Present
Student Assistant — Lead Developer at Educational Technology Department (EduTech UdS) — Apr 2026 — Present
Co-Founder · Advisory Board Member · Technical Lead at Olympia Academia — Aug 2018 — Present
Repository contents
11 projects · 6 roles · 22 posts · 1 publication. Use the explorer on the left, or press Ctrl+P.
At ignite AI GmbH I work on the unglamorous half of enterprise AI — governance policies, PII anonymisation layers, and on-premise LLM deployments for clients who cannot let their data leave the building. At EduTech UdS I am lead developer on the SRL Course Platform, an adaptive learning system built on React 19, Node/Express and PostgreSQL with pgvector.
The thread connecting them is retrieval. Most of my work ends up being some variant of the same question: how do you get the right context in front of a model, cheaply, provably, and without leaking anything you shouldn't. Hybrid search, Medallion ETL, semantic chunking, HNSW indexes, RRF — these are the tools I reach for most.
Before Saarbrücken I read Mathematics at Aligarh Muslim University, which is where the interest in the underlying statistics came from, and spent two years shipping front-ends professionally. I co-founded Olympia Academia in 2018 — now a ~200-member student scientific community that has hosted speakers from Caltech and NASA — and I still sit on its advisory board as technical lead.
Top skills
Sovereign AIAI SecurityMulti-agent Systems
Languages
English — Native or Bilingual (C1)
German — Elementary (A2)
Hindi / Urdu — Native
Certifications
Deep Learning Specialization — Coursera / deeplearning.ai
Enterprise AI that has to survive a compliance review — governance, PII handling, and LLMs that run inside the customer's own network.
AI Governance & Security
Design and enforce enterprise AI governance policies and guardrails to prevent data leakage and guarantee compliance across model workflows.
PII Protection & Data Privacy
Build robust data anonymisation and privacy layers that intercept and protect Personally Identifiable Information before anything is routed to an external or cloud LLM.
Sovereign & On-Premise LLMs
Deploy, fine-tune and optimise local LLMs on-premise, giving clients full data sovereignty and zero cloud-data exposure.
Enterprise AI Chatbots & RAG
Develop custom conversational assistants integrated with local corporate databases, document repositories and internal knowledge bases.
Cloud & Hybrid AI Architecture
Architect secure pipelines that bridge local workloads with cloud AI providers under strict Data Loss Prevention rules.
Stack
Sovereign AIAI SecurityOn-Premise LLMsRAGPII / DLPHybrid Cloud
# experience/edutech-uds.md
Student Assistant — Lead Developer
Educational Technology Department (EduTech UdS), Saarland University · Part-time
currentApr 2026 — Present
Saarbrücken, Saarland, Germany · On-site
Lead developer on an adaptive self-regulated-learning platform, plus two research pipelines: audio-to-knowledge-graph, and real-time gaze streaming.
SRL Course Platform — full-stack lead
Architected and deployed a production adaptive learning platform (React 19 + TypeScript, Node.js/Express, PostgreSQL 16 with pgvector) serving 50+ university course modules. Authored 18 architecture decision records covering authentication, LLM provider abstraction and data architecture.
Document-ingestion pipeline
Built a Medallion ETL pipeline turning lecture PDFs into searchable learning content — PDF parsing, LLM-driven semantic chunking, then vector embeddings in pgvector. Indexed 1,200+ documents with HNSW; cut manual content-labelling time by 40%.
Hybrid retrieval
Combined pgvector HNSW vector similarity with BM25 full-text search over a PostgreSQL GIN index, fused via Reciprocal Rank Fusion. 78% retrieval accuracy on the university knowledge base, served from a FastAPI endpoint at sub-200 ms latency.
Adaptive AI tutor
Backend supporting three content-depth levels per concept with pluggable LLM providers (OpenAI, Anthropic, Ollama). Per-model cost and latency tracking informs provider selection; response quality measured across 500+ student interactions.
AI Co-Design Platform — NLP pipeline
Built a system converting classroom audio into structured knowledge graphs: OpenAI Whisper speech-to-text, pyannote speaker diarisation, then LLM entity and relation extraction. Processed 45+ hours of lecture audio; outputs versioned in PostgreSQL for team collaboration.
Real-time gaze streaming
Optimised WebSocket-based gaze-data streaming from the Pupil Labs glasses API, reducing end-to-end latency on a custom behavioural data-collection tool.
Co-Founder · Advisory Board Member · Technical Lead
Olympia Academia · Volunteer
currentAug 2018 — Present
Remote / International · Remote
A student scientific community I co-founded in 2018. It now has ~200 members and has hosted speakers from Caltech, NASA and other leading research institutions.
Co-founder
Started the community in 2018 and served as Vice President; today I sit on the advisory board as technical lead.
~200-member scientific community
APS-funded, featuring invited speakers from Caltech, NASA and leading research institutions.
Technical lead
Responsible for the organisation's web presence and technical infrastructure, including a Progressive Web App and a RAG-based knowledge assistant over the community's archives.
Stack
CommunityTechnical LeadershipPWARAG
# experience/cix-saarland.md
Graduate Research Assistant
Computational Interaction Group (CIX), Universität des Saarlandes · Part-time
Sep 2024 — Mar 2025
Saarbrücken, Saarland, Germany · On-site
A research study on human-computer interaction patterns in mobile text entry — and the computer-vision pipeline that made it measurable.
Computer-vision pipeline
Built a MediaPipe + OpenCV pipeline tracking detailed hand movements during touchscreen typing, enabling precise analysis of finger positioning and movement patterns. Processed 50,000+ behavioural events at 96% hand-posture classification accuracy.
Eliminated the annotation bottleneck
Delivered a production-ready, reproducible dataset pipeline that reduced manual annotation effort by 30% and unblocked PhD-level forecasting research on user interaction patterns.
Statistical analysis
Implemented position-variability metrics, hand-switching detection and suggestion-bar interaction analysis to quantify typing behaviour.
Multi-modal comparison framework
Compared thumb-based against index-finger typing strategies, surfacing differences in spatial distribution, selection accuracy and ergonomic patterns.
Visualisation & data sync
Matplotlib heatmaps, distribution plots and comparative analyses; a processing pipeline synchronising video with keyboard event logs for temporal analysis of typing behaviour.
Letter of Recommendation on file.
Stack
PythonOpenCVMediaPipeNumPyPandasMatplotlibHCI
# experience/orion-solutions.md
Web Developer
Orion Solutions · Internship
Jul 2022 — Sep 2023
Lucknow, Uttar Pradesh, India · Hybrid
Fifteen months shipping production front-ends in a cross-functional team, and the place I learned what a design-to-dev handoff actually costs.
Front-end development
Developed components using HTML, JavaScript and React; built responsive interfaces for consistent experience across devices.
Design & handoff
Designed and improved user interfaces in Figma, and collaborated with designers to implement UI components aligned with company branding.
Backend integration
Assisted in integrating backend features using MongoDB and SQL, and wired up REST API and database layers.
Standards
Applied modern UI/UX principles, responsive design techniques and accessibility standards across delivery.
Certificate of Internship on file.
Stack
ReactJavaScriptHTMLCSSFigmaMongoDBSQLREST
# experience/taibah-university.md
Research Intern — Computer Vision & Deep Learning
Taibah University · Internship
Sep 2021 — Feb 2022
Al Madinah, Saudi Arabia · Remote
My first research role: compressing crack-detection CNNs small enough to be useful, without giving up the accuracy that made them worth deploying.
Concrete crack detection at scale
Processed the SDNET2018 dataset — 56,000+ concrete crack images — for CNN-based detection, benchmarking nine architectures.
Knowledge distillation
Transferred knowledge from large teachers (ResNet-50, VGG-19) into lightweight students, improving Matthews Correlation Coefficient by up to 20% while shrinking the deployable model.
Interpretability
Applied Grad-CAM to verify the models were attending to actual cracks rather than background texture; achieved 96% classification accuracy on the test set.
Stack
PythonTensorFlowKerasGrad-CAMLaTeX
# projects/
Projects
11 projects — 4 featured. RAG systems, multi-agent platforms, computer vision, and one LED strip.
Ten specialised AI agents solving real business problems behind one FastAPI surface.
Jun 2026Hackathonfeatured
A multi-agent platform built for a hackathon against real company problem statements. Each agent owns one business task end to end — parsing the input, calling the model, and persisting a structured result into a Postgres gold layer so downstream agents can reuse it.
Shared agent base class handling model calls, retries and structured-JSON coercion, so a new agent is roughly one file.
Gemini-backed generation with results cached in a PostgreSQL gold schema, seeded at startup from the supplied problem datasets.
The marketing agent encodes real platform constraints — TikTok/Reels 1080×1920 safe-zone geometry — directly into the generation prompt so output is usable without manual reframing.
A Model Context Protocol server that turns the NCBI Entrez API into a first-class LLM tool.
Aug 2025Personalfeatured
Model Context Protocol client integrating with the NCBI Entrez API for gene and protein metadata retrieval. Built so that an LLM application can look up authoritative biological data in real time instead of hallucinating it.
Symbol-based queries and flexible search across two interfaces — a REST API and a CLI.
Production-ready MCP server speaking JSON-RPC, compliant with NCBI's rate-limiting rules.
Deployed on Vercel with a modular design and automated testing.
RAG chatbot answering 500+ university queries, with a 35% lift in retrieval relevance.
May 2025Personalfeatured
An AI-powered chatbot using Retrieval-Augmented Generation to answer university questions accurately, built on scraped and cleaned data from Aligarh Muslim University's public sites.
Vector embeddings and document-chunking strategies tuned for context-aware retrieval — a 35% improvement in query relevance over the naive baseline.
Local LLM inference optimised with configurable context windows, keeping the whole thing runnable without a paid API.
LangChain orchestrates question processing, context retrieval and response generation over the university corpus.
Containerised with Docker and deployed through Streamlit for easy knowledge-base updates.
Full-stack learning platform that makes Transformers, LLMs and neural nets legible to Master's students.
Apr 2025 — Jul 2025Saarland Universityfeatured
Engineered a full-stack AI-driven learning platform that simplifies complex ML concepts for Master's students, built by a four-person agile team with clearly defined roles (Tech Lead, Co-Lead, Evaluation, Design).
Integrated state-of-the-art NLP models (Hugging Face, PyTorch) to power an interactive AI peer-teaching agent, boosting engagement and conceptual retention.
Scalable React + TypeScript frontend against a Flask backend, delivering multimodal learning — visuals, animations, guided exercises and active-recall checkpoints.
Applied evidence-based pedagogy — chunking, scaffolding, multimodal presentation — to reduce cognitive overload and improve learner outcomes.
Deployed on Vercel with GitHub CI/CD.
Usability testing with 20+ students produced a 30% improvement in navigation efficiency.
Cross-platform study assistant, designed and validated with NASA-TLX workload analysis.
2025Saarland University
A cross-platform application designed to streamline studying by organising, summarising and testing lecture materials, tailored for university students.
AI-powered summarisation of notes, personalised quizzes targeting weak areas, and progress tracking through graphs and metrics.
Led usability testing with semi-structured interviews and NASA-TLX workload analysis, identifying navigation and feedback-visibility as the major UX gaps.
Iteratively refined the interface on user feedback — folder categorisation and weak-area quizzes came directly out of testing.
Navigation efficiency improved 30% across iterations.
Does CLIP transfer to OCT scans? A controlled comparison against ViT and ResNet50.
2025Saarland University
A comparative analysis of the CLIP model against Vision Transformers and ResNet50 for zero-shot classification of Optical Coherence Tomography images.
Zero-shot classification with custom textual prompts, evaluated on Precision, Recall, F1, ROC-AUC, Accuracy, Cohen's Kappa and MCC.
Demonstrated CLIP's strengths in zero-shot transfer while surfacing concrete domain-adaptation limits for medical imaging.
Extensive testing across the OCT2017 and RFMiD datasets.
Stack
PythonPyTorchCLIPViTResNet50
# projects/concrete-cracks.ipynb
Efficient Models for Concrete Crack Detection
Distilling nine CNN architectures into something small enough to deploy on infrastructure.
2021 — 2022Taibah University
A study finding the best deep-learning models to detect cracks in architectural infrastructure, then compressing them without losing the accuracy that justified them.
Trained and benchmarked nine CNN architectures on SDNET2018 — 56,000+ concrete crack images.
Used knowledge distillation to transfer knowledge from ResNet-50 and VGG-19 teachers into lightweight student models, improving MCC by up to 20%.
Grad-CAM interpretability confirmed the models attend to actual cracks, not background texture; 96% classification accuracy on the test set.
Results presented with SciPy-stack visualisation — Matplotlib and NumPy.
MERN health platform with Hugging Face inference and a Gradio prototyping loop.
2023Personal
A web application providing automated AI-based medical diagnosis and live global medical insights, letting users manage health profiles and receive personalised recommendations.
End-to-end ML pipeline covering preprocessing, inference and deployment behind a MERN stack.
Hugging Face pre-trained models for prediction; Gradio cut prototyping cycles by roughly 40%.
Interface designed for non-technical healthcare staff, with user registration, profile management and secure data handling.
Screen-reactive bias lighting — capture the edge pixels, drive the strip, stop squinting.
2022Personal
An Arduino-based ambient light for a monitor that samples the pixels around the screen edge and illuminates an LED strip behind it to match — immersive, and easier on the eyes during long sessions.
Host-side capture loop samples edge regions of the framebuffer and streams colour data over serial.
WS2812b RGB strip driven with C++ and the FastLED library on an Arduino Nano.
Built primarily to reduce eye strain; the immersion was the bonus.
Zea Mays Leaf Disease Classification Using Swin Transformer
2024
Zim, A. H., Ashraf, A., Iqbal, A., Najeeb, M., Malik, A., Kuribayashi, M., Khan, A.
International Conference on Signal, Machines, Automation, and Algorithm (SIGMAA 2024) · Springer · pp. 827--838
Applies the Swin Transformer's hierarchical shifted-window attention to leaf-disease classification in Zea mays (maize), comparing it against convolutional baselines for agricultural disease diagnosis from field imagery.
@inproceedings{zim2024zeamays,
title = {Zea Mays Leaf Disease Classification Using Swin Transformer},
author = {Zim, A. H. and Ashraf, A. and Iqbal, A. and Najeeb, M. and Malik, A. and Kuribayashi, M. and Khan, A.},
booktitle = {International Conference on Signal, Machines, Automation, and Algorithm (SIGMAA 2024)},
publisher = {Springer},
pages = {827--838},
year = {2024},
doi = {10.1007/978-981-97-6352-8_59}
}
# skills.json
Skills
36 across 7 categories. “used in” counts the projects and roles on this site where each one actually appears.
Design-system-driven styling for Skill Lyft and internal tooling.
★★★★☆used in 2
Figma
Figma Inc.
Interface design and prototyping — and running the handoff from the other side too.
★★★★☆used in 4
A
Accessibility
W3C
Semantic markup, keyboard navigation, WCAG contrast. Including this site.
★★★★☆used in 3
Data & DevOps
PostgreSQL
PGDG
16 with pgvector. HNSW vector indexes, GIN full-text, Medallion-layered schemas.
★★★★★used in 5
Docker
Docker Inc.
Compose stacks wiring frontend, backend, KB API and database for reproducible local dev.
★★★★☆used in 5
Git / CI-CD
—
GitHub Actions pipelines, ADR-driven change management, automated test gates.
★★★★☆used in 10
Linux
—
Deployment target for everything above. EC2, Apache path routing, service management.
★★★★☆used in 8
N
NLP Tooling
Whisper · pyannote
Speech-to-text, speaker diarisation and entity extraction over 45+ hours of lecture audio.
★★★★☆used in 2
# contact.sh
Contact
Open to AI/ML engineering roles and research collaboration
#!/bin/sh
# contact.sh — every way to reach me, in one place.
NAME="Mohammad Najeeb"
ROLE="AI Engineer"
BASE="Saarbrücken, Saarland, Germany"
mail --to "mdnajeeb.cs@gmail.com" # personal
mail --to "mona00002@uni-saarland.de" # university
open "https://www.linkedin.com/in/mohammadnajeeb/"
open "https://github.com/mohammadnajeeb"
curl -O "assets/Najeeb_Resume.pdf"
echo "$NAME — $ROLE, based in $BASE"
echo "Open to AI/ML engineering roles and research collaboration"
