AI Engineer · LLM & RAG Systems · Data Science

Co-Founder & CTO at Vexonova · LLM Team Lead

Building practical AI systems from data and documents to retrieval, reasoning, APIs, and deployment.

I am a Computer Engineering student and AI Engineer focused on building LLM, RAG, Ai Agent and practical AI systems that move beyond experiments into usable software.

My work covers the full path from document ingestion and retrieval to grounded generation, APIs, user interfaces, testing, and deployment.

I am also a Co-Founder and CTO at Vexonova, where I contribute to the technical direction of the organization and lead the LLM team across RAG, LLM, and AI-agent projects.

My engineering background also includes Machine Learning, Deep Learning, Computer Vision, and Data Science, giving me experience working with both structured data and real-world unstructured inputs.

Build systems. Understand the trade-offs. Ship what works.

My main focus is practical AI engineering, particularly systems that connect models with real data and real users.

• LLM Applications — Turning foundation models into useful software products.
• RAG Systems — Building retrieval pipelines that ground model responses in trusted documents.
• Document Intelligence — Processing PDFs, DOCX, TXT, Markdown, and scanned documents.
• AI Agents — Exploring systems that can reason, retrieve information, and execute multi-step tasks.
• Machine Learning & Data Science — Building reliable ML pipelines and analytical applications.
• AI Deployment — Packaging AI systems behind APIs, interfaces, and containers.

I intentionally build RAG systems in progressive levels.

Each project introduces a new retrieval concept rather than simply adding more technologies.

A practical baseline for asking questions over personal documents.

Pipeline:

Documents → Chunking → Embeddings → ChromaDB → Retrieval → Grounded Answer

Supports PDF, DOCX, TXT, and Markdown files with local embeddings, persistent ChromaDB storage, provider-independent LLM generation, source citations, Streamlit UI, and Docker Compose deployment.

Stack: Python · LangChain · ChromaDB · Streamlit · Docker

A more structured RAG architecture introducing metadata-aware retrieval and filtering.

The goal is to make retrieval more precise by allowing the system to use document metadata alongside semantic similarity.

Stack: Python · LangChain · Vector Database · Docker

Status: ✅ Completed

The next stage of the retrieval pipeline, focused on combining different retrieval strategies to improve search quality across different query types.

Status: 🚧 In development

ChemLib-AI is an intelligent RAG assistant designed for a university chemistry library.

Instead of manually searching through books and documents, students and researchers can ask questions in Persian or English and receive answers grounded in the library's own content.

The system is designed around a simple principle:

If the documents cannot support the answer, the system should not invent one.

• 📄 Multi-format ingestion — Processes PDF, DOCX, TXT, and other document sources.
• 🔎 Semantic retrieval — Uses local BGE-M3 embeddings with ChromaDB for multilingual retrieval.
• 🖨️ Scanned-document support — Uses vision-based OCR when standard text extraction is insufficient.
• 🛡️ Pre-generation grounding check — Avoids unnecessary LLM calls when retrieved context is insufficient.
• 📚 Source citations — Answers can reference the underlying books and pages.
• 🔄 LLM provider fallback — Automatically switches from Gemini to OpenRouter when necessary.
• 👤 Private document mode — User documents remain session-scoped instead of being added to the shared knowledge base.
• ⚙️ Admin management — Administrators can add, replace, and remove shared documents.
• 🔀 Cross-source comparison — Supports comparing information across multiple sources.
• 🌐 REST API — Exposes functionality through FastAPI.
• 🧪 Automated testing — Includes mock-based pytest coverage.
• 🐳 Containerized deployment — Supports Docker-based execution.

My contribution: Technical Lead at Vexonova and LLM Team Lead, contributing to the design and development of the LLM/RAG system and its engineering workflow.

Stack: Python · RAG · LangChain · ChromaDB · BGE-M3 · Gemini · OpenRouter · FastAPI · Streamlit · Docker · Pydantic · pytest

IAI-001 Safety Monitoring System connects Computer Vision, rule-based analysis, and LLM generation to transform workplace video events into management-ready HSE reports.

Video → YOLOv8 → ByteTrack → Rule Engine → Structured Alerts → Local LLM → HSE Report

• 🎥 Detects relevant events from worksite video.
• 🧠 Converts visual observations into structured safety events.
• 📋 Uses a rule engine for deterministic safety logic.
• 🤖 Uses locally served Qwen2.5 for report generation.
• 🌍 Supports Persian and English reports.
• 🧩 Uses Jinja2 templates for controlled prompt generation.
• ⚡ Exposes the LLM reporting layer through FastAPI.

LLM Team Lead

I led the LLM workstream, including prompt engineering, local model inference with Hugging Face Transformers, Pydantic report schemas, and integration behind the API layer.

Stack: FastAPI · Docker Compose · Gradio · Qwen2.5 · Transformers · Jinja2 · Pydantic

I care about more than simply making a model produce an answer.

Some engineering decisions I actively consider:

If retrieved context is not sufficiently relevant, there is little value in asking the LLM to guess.

A retrieval-quality check can therefore reduce hallucinations, latency, and unnecessary model usage.

Embeddings and retrieval should remain independently testable from the generation layer.

This makes it easier to evaluate where an answer went wrong.

Pydantic models provide a predictable contract between the LLM and the rest of the application.

LLM APIs can fail.

Fallback providers can improve availability while keeping the user informed about what happened.

Private documents should not automatically become part of a shared knowledge base.

Rather than adding technologies for their own sake, I prefer introducing new components when they solve a measurable engineering problem.

LangChain · ChromaDB · Gemini · OpenRouter · Hugging Face Transformers · Pydantic

PyTorch · TensorFlow · Keras · OpenCV · scikit-learn · MONAI

FastAPI · Docker · Docker Compose · Streamlit · Gradio · pytest · Git · uv · SQL

Vexonova is an open-source AI engineering organization focused on building practical systems across LLM, RAG, AI Agents, Computer Vision, and MLOps.

Co-Founder & CTO

I contribute to the organization's technical direction, architecture, engineering practices, and development of AI systems.

LLM Team Lead

I lead the LLM-focused team working on LLM applications, RAG pipelines, AI agents, and production-oriented AI workflows.

My responsibilities include:

• Technical direction and architecture
• LLM/RAG system design
• Technical coordination across projects
• Code and repository review
• Supporting engineering teams
• Leading LLM project development
• Turning AI concepts into working software

My focus is not only on building models, but on building the systems around them that make AI useful, testable, and deployable.

If you are interested in AI engineering, LLM/RAG systems, collaboration, research, or building practical AI products, feel free to reach out.

• GitHub: HosseinHeydari2004
• LinkedIn: Hossein Heydari
• Kaggle: mrhosseinheydari
• Email: hosseinheydari982020@gmail.com
• Telegram: @Hossein_h830

Build deeply. Lead responsibly. Ship useful AI.