Built for the real worldManufacturing+Mechanical engineering+Quality inspection
01 / Our solutions
Your data has more to tell you.
We’re developing two complementary solutions to make engineering knowledge accessible and industrial vision systems easier to train.
01Discover & retrieve
CAD & blueprint intelligence
Find the right drawing, not just the right filename. An intelligent retrieval platform for technical drawings, part images, and project documentation.
Search your way. Start with natural language, technical specifications, or a drawing or part image.
Understand the details. Vision-language models extract geometry, dimensions, materials, and manufacturing steps from PDF and image-based drawings.
Retrieve with context. Hybrid keyword and vector search combines geometry, material, and process signals, followed by semantic and visual re-ranking.
02Generate & train
Synthetic data for industrial vision
Train for the defects you haven’t seen yet. A synthetic data pipeline that turns 3D CAD assets into annotated images for quality inspection.
Build richer datasets. Generate photorealistic training images with varied lighting, camera angles, and surface conditions.
Make rare defects visible. Simulate scratches, cracks, and assembly misalignments for detection and segmentation models.
Prepare before production. Reduce dependence on collecting and hand-labeling scarce physical defect samples.
02 / Under the hood
Deep technology. Practical purpose.
Our approach connects multimodal understanding, precise retrieval, and cloud-native infrastructure. Designed around the complexity of real engineering data.
[ 01 ]
Multimodal understanding
The planned Claude Vision integration focuses on technical-drawing interpretation and structured geometry, material, and process metadata, with validation before downstream use.
Vision-language models / Feature extraction[ 02 ]
Multi-stage retrieval
BM25 and embeddings across independent feature spaces. Reciprocal rank fusion (RRF), metadata validation, and semantic ranking bring the signals together.
DINOv3 region-based embeddings on cropped drawing segments provide an additional re-ranking signal for geometric similarity.
Visual embeddings / Region-based re-ranking[ 04 ]
Cloud-native by design
Containerized services with FastAPI, React, and Express, alongside event-driven ETL using Azure Functions and Event Hubs.
Microservices / Event-driven pipelines[ 05 ]
Connected engineering data
Designed to integrate with PostgreSQL, Blob Storage, and ERP/PLM environments, using delta synchronization and content hashing to track changes.
Enterprise integration / Incremental sync[ 06 ]
Purposeful automation
Agentic development and workflow orchestration support repeatable processing, maintainable systems, and reliable cloud operation.
Agentic workflows / Orchestration
03 / The Vectrio perspective
Built at the intersection of AI and industry.
Engineering teams already have a wealth of knowledge. Too much of it is locked inside drawings, disconnected files, and hard-to-capture visual data.
Vectrio is an independent industrial AI project in development, not yet an incorporated company. We’re bringing together multimodal models, vector retrieval, and synthetic data to help bridge the gap between engineering information and usable intelligence.
Our focus is practical: make existing knowledge easier to find, and make industrial vision systems easier to develop.
In development. Open to conversations.
Let’s build what’s next
Complex challenge? Start a conversation.
Working with technical drawings, CAD data, or visual quality inspection? We’d like to hear about your use case.