All case studies
Active DevelopmentComputer Vision · Multimodal2025 — present
HelixVision
Vision pipeline that turns unstructured imagery into typed scene graphs downstream agents can query.
Role · Lead engineer
Overview
A vision system that turns unstructured imagery into structured, queryable scene descriptions for downstream agents.
Problem
Most vision stacks emit raw detections; downstream systems need typed, relational scene graphs they can reason over.
Solution
A composable inference pipeline that fuses detection, segmentation, and VLM-derived relations into a typed scene representation.
Tech stack
- Python
- PyTorch
- ONNX
- FastAPI
Engineering notesWhat's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
Engineering notes
What's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
Design goals
- Typed scene graphs, not loose blobs of detections.
- Composable inference — each stage swappable behind a stable schema.
Philosophy
The output schema is the product. Get the representation right and downstream code becomes trivial.
Planned architecture
Staged pipeline (detection → segmentation → relation extraction) with ONNX inference behind a FastAPI service; results serialized as a typed scene graph.
Current stage
Interface design and single-stage prototypes.
Links
Repository and demo links will appear here once the project is ready for public review.