Anto Subash — Software Engineer in Vienna

Work

NODES research group · Laxenburg, Austria · since 2017

Mowing detectionsince 2025
Validating the Copernicus grassland mowing product against Alpine webcam archives. A multi-stage pipeline of CLIP and self-hosted vision-language models filters webcams and frames down to usable grassland views, and a Vision Transformer ensemble serves the trained classifier over HTTP.
Self-hosted inference
The institute's own GPU inference stack: vLLM serving generation and embedding models, pgvector retrieval, and tensor-parallelism and memory tuning to fit them on shared hardware. I also train colleagues on LLMs and agents.
Vision models for field photossince 2026
Self-hosted vision-language models that classify and anonymise crowdsourced and survey field photos, with configurable pipelines per collection and background workers for long runs.
Retrieval-augmented assistantssince 2026
Qdrant vector search, LLM reranking checked by an evaluation harness, citation and confidence gates, prompt-injection filtering, and per-tenant rate limits and token budgets.

TransPerfect2012–2016
Full-stack .NET developer in Chennai on TransCEND, a virtual data room with security and audit controls, and TransPort, the client portal for translation requests and delivery.