6DoF pose estimation
Satellite pose estimation
Rendezvous-vision pipelines in the ELSA-M mission context.
Deep learning and vision
Computer Vision For Space, Medical Imaging, And Generative Media
Peter has built computer-vision systems where labels were scarce, domain shift mattered, and geometric or biological validity mattered as much as model choice.
FID 109 → 49
AstroGAN
A sim-to-real translation network that halved the synthetic-to-real FID gap.
89% · 8 cell classes
Oxford histology pipeline
Whole-slide nuclei detection, cell classification, tissue segmentation, and feature extraction.
Mean quality 7.9/10
Generative media pipeline
Photo restoration, ArcFace identity checks, image-to-video generation, and film assembly.
Markush OCSR
Praviar chemical-structure vision
Reads compound and Markush structures straight out of pharmaceutical-patent drawings, held-out validated.
Space vision
At Astroscale, Peter worked where classical vision had underperformed and synthetic-to-real transfer was the central risk.
- Built 6DoF pose-estimation systems from ResNet baselines through FPN, RANSAC-PnP, Vision Transformers, and multi-task detection/pose/keypoint/segmentation designs.
- Designed a custom 3D projection loss that projects the satellite model into image space to align supervision with geometric pose error.
- Measured the sim-to-real gap with per-domain image statistics on SPEED+, then built AstroGAN — a UVCGANv2-based translation network — cutting FID from 109 to 49 by 500 epochs.
Medical imaging research
The Oxford DPhil required computer vision at whole-slide scale: nuclei detection, cell classification, tissue segmentation, and biological interpretation.
- Built a RetinaNet-based nuclei detector and an 8-class cell classifier reaching 89% accuracy on mouse placental tissue.
- Recovered the 8 known cell types at ~0.99 across ARI, AMI, and V-measure using Spectral Clustering on UMAP-reduced embeddings.
- Trained a SimCLR contrastive model on a 3-million-sample unlabelled dataset to probe finer cell substructure.
Chemical and document vision
Praviar extends the vision work into a legal-adjacent product: pharmaceutical patents hide their most important claims in drawings that are never transcribed.
- Built an optical chemical-structure-recognition cascade that reads compound and Markush (generic) structures straight out of patent drawings, validated on a held-out set of pharmaceutical-patent chemical drawings.
- Treats patent intelligence as a multimodal problem, so the system does not walk past the exact diagram that decides freedom to operate.
- The model and provider stack stays product-confidential; the disclosable signal is that vision is a first-class product layer, not a text-only afterthought.
Creative image-to-video engineering
A private family-film project adds a different vision signal: generative media built as an inspectable engineering workflow.
- Built a 50+ module pipeline around colour-managed 16-bit scans, FLUX 2 [pro] restoration, and Kling 3.0 image-to-video generation.
- Used ArcFace identity checks, a VLM judge, and five rounds of prompt iteration to keep generated motion stable, achieving mean quality 7.9/10.
- The local six-stage restoration pipeline was built and tested but bypassed; the film shipped on hosted FLUX restoration.
Continue exploring
Related work
Satellite computer vision
Pose estimation, synthetic imagery, and domain-gap discipline for orbital vision.
OpenAstroGAN sim-to-real essay
Approximately paired translation, style encoding, and custom losses for satellite imagery.
OpenOxford medical AI essay
Whole-slide images, placental histology, supervised phenotyping, and morphology research.
OpenOld-photo animation film
AI-assisted restoration, identity-preserving image-to-video generation, and film assembly.
OpenPraviar patent intelligence
Where the chemical-structure vision cascade sits inside a bounded, deterministic patent-intelligence product.
OpenAI engineer profile
The wider engineering context around the vision work.
Open