💼 Experience
AI Infrastructure & Full-Stack Engineer — Neuralix.ai
April 2025 – Present · Delhi, India · On-site
🛡️ EKAM AI
An indigenous Defence AI-as-a-Service platform, built under the Ministry of Defence iDEX ADITI 2.0 initiative and launched nationally at the Chanakya Defence Dialogue 2025. It runs inside secure air-gapped sites, so every package, model weight and upgrade has to be carried in offline.
What I own
Model serving Production inference on vLLM and NVIDIA Triton Inference Server across GPU nodes, serving LLMs, SLMs, VLMs, OCR, embedding and reranking models. Tuning continuous batching, KV-cache sizing and request scheduling so a fixed number of GPUs serves more users at lower latency.
Model lifecycle Benchmarking candidate open-source models on latency, cost and task accuracy, validating their outputs, then promoting them from evaluation through staging into production serving.
Air-gapped delivery Offline package mirrors, model-weight distribution, dependency resolution, version upgrades and on-site incident support with no external network path. Standalone desktop distributions and VDI deployments for restricted sites.
Application layer The platform's backend in Python and FastAPI, with React interfaces, PostgreSQL, Redis and NGINX behind AI-powered document and retrieval workflows.
Reliability Telemetry with Prometheus and Grafana, profiling application hot paths, then tuning async execution, caching and database access — cutting average API response latency by 40%.
GPU coordination Primary interface between the AI platform team and the GPU hardware and infrastructure teams, running root-cause analysis across application, CUDA driver, network and hardware layers.
Documentation The SOPs, runbooks and architecture documents used for on-site deployment, troubleshooting and knowledge transfer at client sites.
🧰 Stack
Python · FastAPI · React · vLLM · NVIDIA Triton · CUDA · Docker · Kubernetes · Linux · PostgreSQL · Redis · NGINX · Prometheus · Grafana · Qdrant