I design backend systems for geospatial imagery: ingest and catalog, raster serving, and the production side of computer vision.

The work started in 2019 at the ingest end — turning aerial and satellite products into a catalog — then moved to XYZ/WMTS serving, on-the-fly mosaicking, and a composable image-enhancement pipeline. From 2021 I also built and ran the inference service the platform uses for detection, segmentation, change detection and OCR. Most of that job was production plumbing, not model research: devices, preload, failure modes that only show up on real rasters.

Since 2024 the role has been architectural. I led the move from separately-grown Flask services to one async FastAPI backend with a shared auth, cache, database and job layer, and I own the shared library, data-access conventions and deployment story those services sit on. I was the sole engineer on the first systems, so the early design was mine — and I stayed long enough to open them to a team and later rewrite the decisions that did not hold.

Currently: FastAPI geospatial backends, LLM/VLM pipelines over large rasters, and the occasional problem outside our stack.

MediaMTX — RTSP sources that expose a single MPEG-TS track could be pulled but never served as HLS or WebRTC. I diagnosed the gap, implemented the initial RTSP pull-path demux behind the opt-in rtspDemuxMpegts flag, and published a GStreamer repro lab. The change shipped in v1.21.0 with co-author attribution.

Issue #6138  ·  PR #6181  ·  v1.21.0  ·  rtsp-mpegts-lab

deprem-yardim-uydu-hasar-tespiti — Post-earthquake satellite workflow: tile GeoTIFFs, crop against shapefiles, and emit segmentation masks plus COCO annotations.

Day to day
Python · FastAPI · PostgreSQL / PostGIS · SQLite · GDAL · NumPy · OpenCV · Redis / Tile38 · Docker · Kubernetes

Also
Flask · PyTorch / TensorFlow · Go · Kafka · vLLM · Nginx