I build reliable AI systems that ship to production.

Capacity planner for heterogeneous GPU fleets — computes the real training ceiling per node instead of summing pooled VRAM. On 100 consumer GPUs, only 8 can hold a 7B model.

Secure sandbox executing untrusted AI-generated Python in isolated Docker containers — AST scanner + YAML policy engine, four risk levels, full audit log.

Slack-integrated product-design platform used by 100+ designers and managers across 200+ projects at Smartivity Labs — OAuth, per-project channels, reminders, analytics.

Government transparency platform covering every Central department and all 28 states + 8 UTs — budgets, expenditure vs budget, CAG findings, live PIB feed.

CLI that scans your hardware and recommends local LLMs that actually fit — VRAM-aware comparison with Ollama chat built in.

Honest benchmark of a 421M-parameter LLM on 4GB consumer hardware — CUDA-synchronised timing, p50/p95, and measured numbers kept separate from published ones.

4.2M-parameter multilingual (EN/HI/PA) GPT trained from scratch — perplexity 3.32 in 2.38 hours on a GTX 1650 4 GB.

MLOps concept-drift detection with KS tests — async FastAPI logging, Plotly dashboard and Discord alerts, under 500 ms inference.