LLM/RAG • Agentic AI • Computer Vision • NLP • FastAPI • MLOps
I am an AI/ML and Python engineer with 7+ years of experience building practical intelligent systems, machine-learning prototypes, APIs, data workflows, and applied AI applications.
My recent work focuses on taking AI beyond isolated experiments into complete engineering workflows: retrieval and RAG systems, LLM applications, agentic workflows, computer vision, NLP, model evaluation, Python APIs, databases, containerisation, and deployment-oriented architectures.
- Primary focus: AI/ML Engineering, Python Backend Engineering, Applied AI and Generative AI
- Current interests: production-oriented RAG, agentic AI, computer vision, NLP, ML pipelines and reliable AI services
- Engineering approach: combine ML capability with maintainable software, APIs, data infrastructure, testing and deployment
- Availability: open to international remote contract opportunities in AI/ML, Python, Applied AI and Generative AI engineering
Programming & Development
Python • C++ • SQL • PowerShell • .NET
AI / Machine Learning
PyTorch • TensorFlow • Scikit-learn • Machine Learning • Deep Learning • Computer Vision • NLP • Reinforcement Learning • Self-Supervised Learning
Generative AI / LLM Engineering
LLMs • RAG • GraphRAG • Agentic AI • LangGraph • MCP Server • OpenAI API • Hugging Face Transformers • Embeddings • SentenceTransformers • FAISS • Semantic Retrieval • Prompt Engineering
Backend & APIs
FastAPI • Flask • REST APIs • Pydantic • SQLAlchemy • API Integration
Databases & Retrieval
SQLite • PostgreSQL • pgvector • MongoDB • Neo4j • Redis • Vector Search • Graph Retrieval
ML Engineering / MLOps
ML Pipelines • Data Pipelines • Model Training • Model Evaluation • Inference Pipelines • Model Deployment • MLOps
DevOps / Delivery
Docker • Dockerfile • Docker Compose • CI/CD • DevOps • Git • Linux • YAML • PowerShell • Azure DevOps
Applications & Deployment
Streamlit • Gradio • Hugging Face Spaces • nginx
Repository: GRAPHRAG-MongoDB-Docker-Neo4j
This GraphRAG project directly implements Neo4j + MongoDB + FastAPI + Docker Compose + an OpenAI-compatible LLM, with a PowerShell-friendly setup. Neo4j stores people, projects and relationships, MongoDB stores the associated source documents, FastAPI exposes the service, and the retrieval layer combines graph facts with matching documents before the LLM produces a context-grounded answer. The repository therefore provides direct evidence of graph retrieval, Cypher-based relationship querying, document retrieval, API engineering, containerised database infrastructure and environment-based LLM integration.
Principal evidence: GraphRAG • FastAPI • Neo4j • Cypher • MongoDB • Docker Compose • OpenAI-compatible LLM APIs • PowerShell
Repository: AI-RAG-Chatbot
This project demonstrates a more production-oriented RAG infrastructure rather than only a notebook-level retrieval experiment. It uses PostgreSQL + pgvector for vector retrieval, Redis for caching and the ingestion queue, background document ingestion, retry handling, grounded retrieval, FastAPI services, Docker Compose and nginx. Uploaded PDF/TXT/Markdown documents are processed asynchronously, embedded and indexed, while the question-answering path returns source-aware answers from retrieved chunks and explicitly avoids guessing when relevant context is unavailable.
Principal evidence: RAG • FastAPI • PostgreSQL • pgvector • Redis • Background Ingestion • Retry Handling • Retrieval • Docker Compose • nginx
Repository: Multi-Agent LangGraph
This project provides commercially relevant evidence in multi-agent / agentic AI and LangGraph orchestration. A supervisor routes work between researcher, writer and reviewer agents through structured shared state; the reviewer can send work back for revision, while bounded revision and recursion controls prevent uncontrolled loops. It also demonstrates Pydantic-based structured routing, modular agent separation, environment-configured OpenAI-compatible LLM integration and fallback routing when structured supervisor output fails.
Principal evidence: Agentic AI • LangGraph • Supervisor/Worker/Reviewer Routing • Structured State • Pydantic • OpenAI-compatible LLM Integration • Bounded Revision Logic
Repository: PPG-Machine-Learning
The PPG/BCG project provides deeper ML-engineering and representation-learning evidence: self-supervised SimCLR pretraining, multimodal PPG/BCG architecture, 1D CNN or patch-Transformer encoders, an STFT computer-vision branch, signal-quality assessment, downstream heart-rate regression, configuration-driven train/evaluation pipelines and automated tests. Importantly, evaluation uses subject-independent splitting so windows from the same subject cannot leak across train/validation/test sets; the repository includes a test that locks this evaluation rule in.
Principal evidence: PyTorch • Self-Supervised Learning • SimCLR • Multimodal Architecture • CNN/Transformers • STFT • Training/Evaluation Pipelines • Subject-Independent Splitting • pytest
Repository: Card-Box-Video-Detection-Tracker
One of the strongest engineering projects in the portfolio, this system tackles persistent geometric tracking rather than simple per-frame rectangle detection. It combines Lucas–Kanade optical flow, RANSAC homography, persistent corner identity, occlusion handling, impossible-jump rejection, One-Euro temporal smoothing and recovery/reinitialisation logic, with OpenCV CSRT/KCF/MOSSE baselines and optional CoTracker, SAM2 and YOLO11n-seg backends. It also includes pytest coverage, synthetic ground-truth generation, IoU/identity-flip/corner-error evaluation and a compact browser export pipeline. The documented synthetic trial reports 0.96 mean IoU, zero identity flips and 5.3 px mean corner error for the micro-dot pattern, clearly labelled as synthetic evaluation rather than real-world performance.
Principal evidence: OpenCV • Lucas–Kanade Optical Flow • RANSAC Homography • Occlusion Handling • One-Euro Smoothing • CoTracker/SAM2/YOLO • Synthetic Evaluation • IoU/Error Metrics • pytest • Browser Export
Repository: Summarizer-AI-Meeting
This repository is particularly useful as evidence of a complete AI application rather than isolated model experimentation. It combines FastAPI REST endpoints, SQLAlchemy/SQLite persistence, faster-whisper transcription, Hugging Face Transformers, English/Farsi NLP, BART and mT5 summarisation, named-entity extraction, action-item and owner extraction, deadline parsing with dateparser/jdatetime, transcript-based speaker detection and ReportLab PDF reporting. The result is an offline-first workflow that accepts audio or transcripts and turns them into stored, structured and exportable meeting intelligence.
Principal evidence: FastAPI • REST APIs • SQLAlchemy • SQLite • faster-whisper • Hugging Face Transformers • English/Farsi NLP • Summarisation • Entity/Action Extraction • Deadline Parsing • PDF Reporting
Repository: semantic-RAG
Live demo: Hugging Face Space
The semantic-RAG project demonstrates a complete hybrid retrieval and evaluation pipeline using SentenceTransformers embeddings together with TF-IDF lexical reranking and a PyTorch/Hugging Face question-answering model. Beyond retrieval itself, the repository contains corpus preprocessing, reusable embedding caches, intent-sensitive retrieval logic, interactive inference, batch prediction, EM/F1 evaluation, smoke testing, top-level pipeline orchestration and a Streamlit deployment on Hugging Face Spaces.
Principal evidence: SentenceTransformers • TF-IDF Lexical Reranking • PyTorch/Hugging Face QA • Hybrid Retrieval • Caching • Intent-Sensitive Retrieval • Preprocessing • Batch Prediction • EM/F1 Evaluation • Smoke Tests • Streamlit
Repository: Obstacle-Detection
Applied railway-safety AI prototype combining a PyTorch obstacle classifier with computer-vision and simulated camera/IR/occupancy sensor inputs. The broader system demonstrates Flask REST APIs, SQLite evidence logging, Folium GIS mapping, JSON/CSV reporting, TCP multi-train communication concepts and Flower/FedAvg federated-learning experimentation. It is presented as a prototype/simulation rather than as a deployed safety-certified railway product.
Principal evidence: PyTorch • Computer Vision • Sensor-Fusion Concepts • Flask • REST APIs • SQLite • Folium GIS • Networking • Flower/FedAvg
Repository: LLM-Chat-API
The LLM Chat API independently demonstrates FastAPI + the OpenAI Python SDK/API pattern + Docker + Dockerfile + Docker Compose. It provides a clean containerised API service with Pydantic request/response models, an OpenAI-compatible client, configurable base URL/model selection, environment-based API-key configuration and runtime injection of secrets rather than baking them into the image. This gives clear standalone evidence of LLM API integration and containerised Python API deployment.
Principal evidence: FastAPI • OpenAI Python SDK/API Pattern • Docker • Dockerfile • Docker Compose • Pydantic • Environment-Based Configuration • Containerised API Deployment
Repository: RAG-pdf-Project
Live demo: Hugging Face Space
The RAG PDF project demonstrates a modular document-QA architecture covering PDF ingestion, 300-word chunking, multilingual SentenceTransformer embeddings, FAISS vector search, top-k retrieval, FLAN-T5 answer generation, Gradio UI and Hugging Face Spaces deployment. Its repository separates ingestion, retrieval, generation, configuration and pipeline modules, making the end-to-end RAG architecture easy for a recruiter or technical interviewer to inspect.
Principal evidence: PDF Ingestion • Chunking • Multilingual Embeddings • SentenceTransformers • FAISS • FLAN-T5 • Modular RAG Pipeline • Gradio • Hugging Face Spaces
| Project | Focus |
|---|---|
| SENTIMENT-NLP- | DistilBERT • NLP • Hugging Face • PyTorch • Streamlit |
| Churn Risk Prediction | Machine Learning • Classification • Threshold Optimisation |
| MountainCar DQN | Reinforcement Learning • Deep Q-Networks |
| Catch Game DQN | Reinforcement Learning • DQN • Reward-Driven Learning |
| XGBoost Classification | Supervised ML • Gradient Boosting |
| UMAP Clustering | Dimensionality Reduction • Clustering |
| MNIST GAN | Generative Deep Learning • GANs |
| GPT Dataset Builder | Generative AI • Dataset Generation |
| PyTorch Implementations | Deep Learning • PyTorch |
| System Monitor | Python • Systems Programming |
| Autism Assistant | Applied AI • Software Engineering |
I also maintain forks and reference repositories for technical study, experimentation and exploration. These are kept separate from the authored project portfolio above and are not presented as work authored from scratch.
Areas explored include large-model inference and optimisation, transformer training, CUDA/GPU engineering, Python internals and graph data science. Examples include TensorRT-LLM, Megatron-LM, CUTLASS, cuDNN Frontend, CPython, LLMs from Scratch, Model Optimizer, Open GPU Kernel Modules and Neo4j Graph Data Science.
- Production-oriented Generative AI and RAG systems
- Agentic AI and multi-agent workflows
- AI-enabled Python backend services and APIs
- Computer vision and intelligent perception
- NLP and transformer-based applications
- ML pipelines, evaluation and deployment
- MLOps, CI/CD and containerised delivery
- Reliable AI for operational and safety-oriented systems
- GitHub: EbiAraz
- LinkedIn: linkedin.com/in/realebimhmdli
- Email: ebimahmudali@gmail.com
- Opportunities: INTERNATIONAL remote contract roles in AI/ML, Python, Applied AI and Generative AI engineering
Building AI that moves from models to meaningful systems.