Microsoft Build 2025 is here – offering a front-row seat to the latest in developer-first innovation, GenAI tooling, and real-world applications built for scale. If you’re working with unstructured data or RAG workflows, stop by the Microsoft for Startups booth. Our live demos are designed to help you accelerate AI development by orchestrating unstructured, multimodal pipelines using Azure Models-as-a-Service.
In this post, we’ll walk through the two technical demos we’re showcasing live:
How to refine Microsoft’s Kosmos-2 VLM using active Learning pipeline, and
How to convert unstructured enterprise PDFs into chunked, embedded, and semantically structured data for RAG-based agents, using Azure MaaS and Dataloop’s orchestration platform.
Whether you’re fine-tuning a foundation model for your domain or preparing documents for agent-driven retrieval, the challenges are the same: manual effort, fragmented workflows, and scaling challenges. These demos are built to show you how to simplify that complexity – using a visual interface or programmatically via APIs, depending on your team’s needs.
Demo 1 | Continuously Improve Kosmos-2 VLM Performance with Dataloop Active Learning Pipeline
In this demo, you’ll learn how Dataloop’s Active Learning Pipeline automates iterative model refinement with minimal manual data preparation.
Microsoft’s Kosmos-2 is a powerful vision-language model – but real-world performance often requires adaptation to task-specific or domain-specific data. Traditionally, this involves repeated cycles of data collection, review, labeling, training, and evaluation – each time-consuming and resource-heavy.
Dataloop’s Active Learning Pipeline replaces this manual overhead with a fully automated, iterative training loop. The pipeline continuously identifies the most informative data samples, triggers model predictions, routes ambiguous results for expert validation, retrains the model, and compares its performance – automatically. The best-performing version is promoted, and the cycle repeats – refining the model with every round.
Key Capabilities
- Plug-and-play support for Kosmos-2, YOLOv8, and other vision models.
- Automates the entire loop: data selection → validation → training → evaluation → deployment
- Prioritizes edge cases to reduce unnecessary manual input.
- Built for visual execution and collaboration across ML, data, and QA teams.
Why It’s a Game-Changer
- Reduces manual data prep by up to 85%
- Accelerates iteration speed through feedback-driven automation
- Ensures Kosmos-2 continuously adapts to new data with minimal effort
- Replaces fragmented tooling with a repeatable, self-improving pipeline
Real-World Use Cases
- Retail: Improve product matching with visual-textual feedback loops
- Construction: Detect missing safety gear with model versions retrained on hard-to-predict scenarios
Agriculture: Refine plant disease detection models based on incremental domain data
Demo 2 | Convert Complex PDFs into Semantic-Ready Data for RAG Agents Using Dataloop & Azure MaaS
In this demo, you’ll learn how Dataloop’s multimodal pipeline extracts, chunks, and embeds PDF content to power retrieval-augmented workflows.
Enterprise knowledge is often locked inside dense, unstructured PDFs – compliance reports, clinical trial docs, internal manuals. These aren’t usable by LLMs or RAG agents unless they’re cleaned, structured, and embedded for retrieval.
Dataloop’s PDF-to-RAG Pipeline, powered by Azure MaaS, does exactly that.
This pipeline ingests PDF files and applies a sequence of Azure-hosted models — like Donut and Pix2Struct — to extract structured text, images, and tables. Then, it chunks the content and generates vector embeddings, creating a dataset that’s immediately usable by RAG-based agents like LangChain, Azure OpenAI, or Gemini.
Key Capabilities:
Uses multiple OCR and layout models to capture complex document structure
Extracts and merges multimodal elements: text, tables, and images
Custom chunking + embedding strategy for downstream retrieval
Zero-code configuration — run visually, adapt per use case
Why It Matters:
Replaces weeks of manual preprocessing with a single automated pipeline
Enables faster access to critical information locked in complex documents
Real-World Use Cases:
Healthcare: Process protocols or research papers for AI-powered clinical assistants
Finance: Turn regulatory PDFs into searchable vector databases
Legal: Structure contracts for semantic retrieval and question answering
These workflows are built for teams working with unstructured, multimodal data – automating AI processes from model refinement to document structuring for GenAI and agentic systems. Powered by Azure MaaS and orchestrated through Dataloop, they’re fast to deploy, easy to adapt, and ready for production.
