Skip to content
Mohammad.
All systems
Case Studyoperational2023 – 2024

Applied AI / ML Projects

Vision, RAG chatbots & generative AI

AI Engineer

Request flow
visionYOLO · OpenCV
function-callingRAG agents

01 · Context

Before LLM products were a job description, this was the apprenticeship: a run of applied-AI builds at IR Solutions, each taking one capability — retrieval, detection, generation, similarity — from a research notebook to something a product could actually call. Different domains (healthcare, automotive, finance), same discipline: ship it as an API or it doesn't count.

02 · Architecture

Each build is a small Python service around a task-specific model. The hospital assistant runs RAG over a medical knowledge base with OpenAI Function Calling wired to appointment scheduling. The car-image pipeline chains YOLO detection with OpenCV segmentation, masking and contour post-processing to strip backgrounds automatically. Stable Diffusion sits behind parameterized REST endpoints for text-to-image generation. The face-similarity system embeds faces with FaceNet/ResNet and indexes them in Pinecone for fast top-N retrieval, and the financial work pairs a Streamlit metrics dashboard with a generative chatbot that does web retrieval and multi-turn memory.

03 · Decisions

01One capability per proof

Every project isolated a single AI capability instead of chasing a do-everything system — which made each one finishable, testable, and honest about what it could do.

02APIs as the finish line

A model in a notebook is a demo; a REST endpoint is a product feature. Everything here ended as a callable API — the YOLO pipeline, Stable Diffusion generation, the RAG chatbot — so integration was someone's afternoon, not someone's quarter.

03Function calling for real-world actions

The hospital chatbot doesn't just answer from the knowledge base — OpenAI Function Calling lets it book appointments, turning retrieval into an assistant that completes tasks.

04Vector indexes for similarity at speed

Face embeddings went into Pinecone rather than a hand-rolled nearest-neighbor scan, so top-N matching stayed fast as the index grew.

04 · Outcomes

  • A hospital RAG chatbot with function-calling appointment scheduling over a medical knowledge base.
  • An automated car-image background-removal pipeline built on YOLO and OpenCV.
  • Parameterized text-to-image REST APIs over Stable Diffusion, plus a photo-mosaic application.
  • A FaceNet + Pinecone face-similarity retrieval system and an interactive financial-metrics dashboard with a generative-AI chatbot.

Stack

PythonPyTorchOpenCVYOLOStable DiffusionLangChainPinecone

Want the full walkthrough?

The repository is private — happy to walk through the code and decisions on a call.

Get in touch