This repository contains use case demos developed for Private Cloud AI (PCAI).
The most generic, vertical-agnostic demos, implementing some of the most recurrent use cases are found are the root level of this repo.
These are the following:
| Demo | Short Description | Demo Video |
|---|---|---|
| Basic Agent Langflow | A Langflow setup defining a basic agentic flow to answer questions requiring informations from both local files, using RAG, and data from a SQL database. Relies on MLIS for model deployment. MCP server usage optional. | link |
| Base Code Assistant - Opencode | An explanation on how to setup and use Opencode, an open source AI coding agent, in a VS Code server, leveraging models deployed using MLIS. Includes an optional step to leverage GitHub MCP server. | link |
| Conversation Toolbox | A custom web application connecting to a chat model, an ASR model and Fish Audio S2 pro TTS model, deployed using MLIS, to provide a multilingual AI voice assistant, as well as file transcriptions capabilities. Voice Assistant accepts connections to MCP servers to enrich its capabilities. | link |
| Finetune Tool Calling LLM | Notebooks, using Nemo microservices to fine-tune an LLM to improve its tool-calling capabilities. | link |
| Image Generation - ComfyUI | An explanation of how to simply use ComfyUI, an AI creation engine enabling powerful media creation AI workflows, such as, but not limited to image generation, image editing and video generation. | link |
| Image Segmentation | Python scripts to fine-tune CNNs for segmentation tasks on provided datasets, expected to be executed in a Jupyter notebook, with experiment tracking on MLflow. Also includes a streamlit application to display segmentation results from any checkpoint saved, on any dataset image. | link |
| Multimodal RAG | An advanced retrieval-augmented generation (RAG) flow, supporting multiple files modalities, including text, images, audio and video, coming with its own MCP server to easily reuse this RAG flow elsewhere. Includes steps on how to use it with Open WebUI and Opencode. | link |
| NL to SQL | An Open WebUI setup to allow chatting with SQL data, leveraging tools from an MCP server to interact with data from a Postgres database. Relies on MLIS for model deployment. | link |
| Object Detection - YOLO | A simple streamlit application running object detection inference using a YOLO model on images and videos it takes as input. | link |
| Text Document Analysis | A simple web application in which users can upload text and PDF files, ask or upload a list of questions and get answers for each document in an Excel sheet, after document analysis leveraging an LLM deployed using MLIS. | link |
| Vision Analytics | A Gradio application using a VLM to analyze images, videos and/or streams. Files can be uploaded from the UI, or read from the filesystem. Relies on MLIS for model deployment. | link |
The vertical demos folder contains demos bound to a specific vertical, usually providing their own data.
It contains the following demos:
| Demo | Short Description | Demo Video |
|---|---|---|
| Molecular Aligned Multi-Modal Architecture and Language (Biomed-MAMMAL) | A BentoML inference service for a biomedical foundation model which achieves state-of-the-art results over a variety of tasks across the entire drug discovery pipeline and diverse biomedical domains. | - |
| Blood Vessel Geometry Analysis and Reconstruction | A streamlit application relying on NVIDIA Vista 3D model (deployed using MLIS) to analyze, reconstruct and render vessels in 3D. | link |
| Defence Ops | A web application leveraging a VLM (deployed using MLIS)to analyze videos, with preloaded defence-related ones provided for example. | link |
| Genome Sequencing | Notebooks leveraging NVIDIA Parabricks for genome sequencing. | - |
| Hospital Visit Summary | A streamlit application that can display patient information regarding their previous visits from a database, and summarize it. Requires deploying an LLM using MLIS. | link |
| Lawfirm Co | An application using RAG and video analytics in the context of legal documents analysis. Requires deploying a VLM and embedding model using MLIS. | - |
| License Plate Number Detection | An application using an object detection model (YOLO) and an OCR one to extract license plate numbers from videos. Uses MLIS for model deployment. | - |
| Maintenance Ticket Assistant | An application that can classifies tickets and provide expected resolution steps using a chat model. Also uses OCR to analyze text from network equipment photos for diagnostic purposes. Relies on MLIS for model deployment. | - |
| Predictive Maintenance | A predictive maintenance model trained leveraging Jupyter Notebook, tracked in MLFlow, packaged with BentoML and deployed via MLIS. | link |
| Secure Loan Verification Demo (SLVD) | A governed, human-in-the-loop AI workflow for loan renewal: an agent gathers credit data from five bank systems via a governed MCP server and drafts a decision memo; a policy gate escalates large or risky cases to a human approver, and every step is audited. LLM served via MLIS or LiteLLM; includes a React portal and approval-email flow. | link |
| Traffic Report | A streamlit application that uses a VLM and YOLO to detect vehicles in images/videos and provide an analysis of the scenes. Relies on MLIS for model deployment. | link |
| Water Utility Planner | A chat assistant in charge of predicting which sewer pipes require inspection and why, requiring XGBoost model training with Jupyter Notebooks, tracking with MLflow, packaging with BentoML, deployment with MLIS, using Open WebUI for interaction. | link |
The misc demos folder contains demos that are neither implementing a solution to a common use case, nor bound to a specific vertical. They may or may not partially overlap with primary demos, implement an uncommon use case, or just be custom apps built as pure technical demos.
It contains the following demos:
| Demo | Short Description | Demo Video |
|---|---|---|
| AI Vulnerability Scanner | An AI-powered scanner that crawls a bundled OWASP Juice Shop app and uses an LLM (deployed via MLIS) to analyze each page like a pentester, surfacing ranked, remediated findings in a live dashboard. | link |
| Agentic Meetings Simulations | A custom web application that simulates company meetings using agentic AI workflows. Relies on MLIS for model deployment. | link |
| Offline Meeting Transcription | A transcription pipeline converting raw audio recordings into speaker-attributed transcripts and structured meeting minutes, using Whisper (for ASR, deployed on MLIS) and Pyannote (for speaker diarization) connected to Open WebUI. | link |
| Onboarding Buddy | A mock application to help organizations streamline onboarding for new hires, with task management by admin users and an AI Q&A assistant for the new hires. Uses MLIS for model deployment. | link |
| Realtime Live Voice Translation | A custom web application that captures the user's voice and provides transcription and translation in real time. Relies on Whisper ASR model and a generic LLM deployed on MLIS. | - |
| Voice Agent XTTS | A custom Gradio application that connects to a chat model, Whisper for STT and XTTS-v2 for TTS, all deployed on MLIS, to provide a conversational assitant, able to discuss with the user in many different languages. Also includes a "chat with SQL data" scenario. | link |
The archived demos folder contains outdated demos, usually rendered obsolete by newer demos. While these demos may still run fine on newer PCAI instances, we do no longer support them. They are provided for reference only.
It contains the following demos:
| Demo | Short Description | Demo Video |
|---|---|---|
| AI Support Assistant | A mock support application relying on Open WebUI built-in RAG capabilities and Airflow. Relies on Ollama for model deployment. | - |
| Coding Assistant | A setup using MLIS for model deployment, Open WebUI to define a custom pipeline using that model, and the VScode extension Continue.dev using that pipeline to act as code assistant. | link |
| Live Stream Frame Analytics | A Gradio application for analyzing multiple, real-time video streams using a Vision Language Model (VLM) deployed on MLIS. | - |
| Media Database SQL RAG | Helm chart to deploy Vanna AI, a tool that relies on an LLM to convert natural language questions into SQL queries, enabling chatting with SQL data. MLIS can be used to deploy the LLM. | - |
| Voice Agent MagpieTTS | An older demo based on a custom Gradio application that connects to a chat model, parakeet-ctc-1.1b-asr for STT and magpie-tts-multilingual for TTS, all deployed on MLIS, to provide a conversational assitant. | link |
| Voice Agent Open WebUI | An Open WebUI setup using a chat model, Whisper for STT and Chatterbox for TTS, deployed on MLIS, to allow voice-to-voice chatting with the chat model, in many different languages. Includes instructions for chatting with SQL data as well. | link |
The following demos will be updated:
- Finetune Tool Calling LLM
New demos are being considered:
- Model Monitoring
- New agentic demo
We welcome demo contributions, see CONTRIBUTING for more details.