Things I have built — tools, systems, and experiments in AI, cloud, and enterprise technology.
Part 2 of the OnDevice.ai build notes. Seventeen days, 240 catalogued changes, and one idea that reorganised everything — the model is a component, and what you actually build is the shell around it. Six turning points, plus an honest ledger of what still isn't done.
OnDevice.ai is a private, local-first AI workspace — chat, projects, files and deterministic PPTX/XLSX/DOCX generation, running on your own hardware. The architecture, the design bet behind it, and an honest comparison with Claude and ChatGPT.
Why I built 'Forward Deployed Engineer: AI for the Infrastructure Estate' — course seven on AICademy. FDE method, infrastructure substance: embed with a client, assess the estate, then deliver and govern an AI change behind a human gate, across one accumulating 10-week engagement.
AICademy: six hands-on AI courses and 82 modules built for infrastructure support engineers at GSIs like TCS, HCL and Infosys. The philosophy, the stack, the courses, and why it's different from every AI course online today.
A rebuild log for AICademy: the six upgrades I shipped after a code review — scroll restoration, empty-state removal, lazy loading, a week's-path checklist, catalog wayfinding and worked examples — the two findings I declined, and the performance fix I deliberately left half-done.
A visual guide to AICademy's five hands-on paths — Service Desk, SysAdmin, IT Project Manager, RAG Engineering, and Agentic AI Leadership — and how to pick the one that matches the job you already do.
A technical deep-dive into my IIT Madras Agentic AI capstone: a framework-free customer-support resolution agent with deterministic pre-LLM safety, grounded RAG, guarded tool use, memory — and the live date-arithmetic failure that reshaped its architecture.
How I built a retrieval-based (RAG) jewellery-store chatbot with Flowise — a low-code visual pipeline that reads a product catalogue, remembers context, and suggests alternatives when an item is out of stock. My Week 20 mini-project write-up.
Two real AI failures, two domain safeguard designs, and the responsible-AI thinking that connects them. My write-up from a mini project on agent risk, ethics, and governance.
Our first build mirrored the org chart. It was the wrong shape. Here's how five recognised agentic design patterns, scored against the same criteria, led to a hybrid recommendation — and what changes in v3.
How I built LumenAI — a local-first, native macOS RSS reader with full-text search, offline reading, and pluggable AI summaries — from an empty folder to a signed DMG in about an hour, with Claude writing every line.
How a MacBook without a touchscreen led to Glide: an invisible touchscreen powered by computer vision.
The second chapter of Aether: how a local-first team of IT architecture agents grew from a clean idea into a 13-agent, web-first, self-escalating system — and every bug that shaped it along the way.
An idea piece: how Digital Workplace Operations teams are structured today, where agentic AI is heading, and how a five-agent swarm can replace the follow-the-sun model — same SLA, 24×7 coverage, multilingual by default, with a hash-chained audit trail any G500 internal-audit team can verify in one click.
A domain-specific Retrieval-Augmented Generation assistant built with LangChain, OpenAI embeddings and FAISS that answers questions about the GitHub REST API strictly from indexed public documentation. Week 15 graded mini-project of the IITM Pravartak Professional Certificate Programme in Agentic AI and Applications.
How to build a contextual technical-support and architecture agent that is meaningfully better than a general LLM at one product, using nothing but publicly available vendor documentation. A walk-through of a working BigFix Workspace+ POC, the design decisions behind it, and the pattern that generalises to any vendor.
How I built Aether — a local-first, multi-agent AI system that runs 10 specialist IT architecture advisors on a single MacBook M5 Pro, with no cloud, no API costs, and zero data egress.
Fed up with ad-infested download sites, I built a clean local tool that downloads TikTok, Instagram, and X videos without watermarks. Here's how the provider chain, proxy download, and caching work — conceptually.
How I got fed up with macOS native screenshot chaos, missed Greenshot from my Windows days, and built a lightweight menu bar screenshot app using Swift and Claude.
A comic-style walkthrough of how I built curiousbit.netlify.app using Claude, Codex, Hugo, Tailwind, and Netlify — zero frontend experience required.
A small structured review seam beats blind automation. The case study: a two-stage Python tagger that turned 35 garden HEICs into captioned posts, with 97% acceptable output after thirty seconds of human review.
A walk through the two foundational mechanisms behind every modern NLP system — generative language models and similarity-based embeddings — using five hands-on exercises with Hugging Face and Gensim. Week 6 mini-project of the IITM Pravartak Professional Certificate Programme in Agentic AI.
