Can a decoder-only LM be trained end-to-end on a single M4 Mac without sacrificing fundamentals? Turns out, yes — stable val-loss 2.39, perplexity ~10.9.

A Flask + pandas simulator that redistributes hourly crowd flow across six operational groups — the seed for a heuristic-first agentic ops planner.

Twin Flask microservices — one local, one on GCP — deliberately stripped to their essentials so autoscaling, firewalls, and health checks are the only moving parts.

Two VMs, one private network, zero internet — proving that service-to-service contracts hold up under the strictest isolation.

Classical regression against a DMart-style retail dataset to decompose each marketing channel's true contribution to sales — before the generative AI hype.

Before diffusion models made stylisation trivial, a disciplined bilateral-filter + adaptive-threshold pipeline still produces Pixar-adjacent frames — fast, explainable, and GPU-free.

A resilient BeautifulSoup + pandas pipeline for structured Fortune-500 India data — the raw material for downstream MMM, retrieval, and benchmarking experiments.