Writing

How my projects work, what I measured, and what I would do differently.

• RAG · FAISS · llama.cpp

  Offline RAG on four budget PCs: FAISS, llama.cpp and Qwen2.5 on a LAN

  How I built offline retrieval-augmented generation (RAG) on four Intel i3 PCs: upload PDFs, ask questions and get cited answers with FAISS, MiniLM and a llama.cpp Qwen2.5 node.

• NLP · Text classification · Domain shift

  Why UK-trained spam filters fail on Indian scam SMS

  ScamShield measures how models trained on UK spam data lose 33–52% accuracy on Indian scam SMS, and how a de-obfuscation step and Indian training data recover it.

• Air quality · XGBoost · Flask

  Building an AI-powered air quality dashboard for Chennai, with forecasts and healthy routing

  How Pollumap V3, an AI-powered air quality project for Chennai, combines live pollution, weather and satellite fire data with XGBoost 24-hour forecasts and an A* healthy-routing engine.

• Local LLM · Ollama · Discord bot

  A local-first AI content pipeline for a Discord community

  How FRONTIER discovers hackathons, research and AI news from 10+ sources, scores them with a local LLM and publishes to Discord only after human approval.

• Next.js · Cloudflare Workers · Hono

  Building an attendance ERP that answers "how many classes can I miss?"

  How Attendance ERP goes beyond tracking: it predicts attendance after future absences, calculates safe bunks, plans recovery and detects mid-class bunking.

• Air quality · Machine learning · Forecasting

  AI-powered air quality projects: forecasting, sensor fusion and mapping

  An overview of four AI-powered air quality projects: Prophet and XGBoost forecasting with mortality reports, a Chennai dashboard, multi-source hyperlocal mapping and single-sensor fusion.