Architected an on-prem HPC stack (96-core Threadripper
7995WX, 48GB A6000, 256GB RAM) serving 1,500+ daily
inference requests at 99.2% uptime.
Configured a first-of-its-kind-in-India 4× Mac Studio
cluster (512GB unified memory, 16TB storage) with a
Thunderbolt RDMA / InfiniBand-style fabric for MLX/MPS
inference and training; leading $300K NVIDIA Blackwell
bring-up.
Real-time computer vision on NVIDIA Jetson carried on
DJI drones: 30fps edge inference over RTMP.
LLM semantic translator for Schneider Electric, with IEC
61131-3 to IEC 61499 automation for legacy PLC
codebases; cut migration effort by about 70%.
Technical document ingestion for machine and
electronics engineering support, evolved from RAG +
BM25 + keyword clusters to trigram indices with custom
pi.dev search agents; VLM + LLM stack across 2,000+
docs at 94% accuracy, roughly 3× faster than the prior
approach.
Application security analysis platform for Schneider
Electric: source and compiled-binary evaluation in
computer-use sandboxes with OS state monitoring (registry
/ DLL deltas), pattern-matched to a correlated CVE / CWE
/ MITRE index; pentest lead time from three weeks to
three days.
Shipped three production AI systems; two peer-reviewed
publications in progress; team write-ups on blog.ecitis.org.