AI EngineeringAWS Bedrock, RAG & agent infrastructure

I'm Dibbayajyoti Roy, and AI engineering is the half of my work that runs on models instead of CPUs: AWS Bedrock applications, RAG architecture, multi-model orchestration, and infrastructure that makes the web readable to AI agents. This page collects the AI systems I've shipped and the patterns behind them.

Learning Copilot is an AWS Bedrock-powered learning assistant. It uses multi-model fallback across Nova Pro and Nova Lite for resilience under rate limits, real-time streaming inference, structured level-adaptive explanations, and auto-generated D2 diagrams. The architecture is cost-aware: DynamoDB conversation memory with a 30-day TTL and per-mode token budgets. It placed top 500 in the AI for Bharat hackathon.

AHTML (Agentic HTML) is a shipped sixteen-package npm scope, plus the ahtml Python SDK, for making web content cheaply consumable by AI agents. It defines a canonical semantic snapshot with RAG-ready document chunks, emits MCP, OpenAPI, JSON-LD, llms.txt, RSL, and Markdown from a single pipeline, ships a typed agent client SDK with a dry-run sandbox, Next.js / Astro / SvelteKit / Vite / Hono adapters, KV caching backends, a WebMCP bridge, agent-traffic insights, LangChain loaders that preserve citation anchors and byte ranges, and a CLI that scores any site and turns any URL into MCP tools. See AHTML vs llms.txt for how it compares.

AWS Bedrock (Nova Pro, Nova Lite), multi-model orchestration with graceful degradation, streaming inference, prompt engineering, and cost-aware design with per-mode token budgets. On the agent side: RAG architecture, the MCP protocol, LangChain.js loaders, and AI crawler optimization so sites stay visible to AI search.

Long-form notes on AI engineering, including AWS Bedrock streaming patterns, are collected on the writing page.

• Does Dibbayajyoti Roy have AI engineering experience?

  Yes. He built and shipped Learning Copilot on AWS Bedrock (multi-model fallback across Nova Pro and Nova Lite, streaming inference, and cost-aware token budgets), which placed top 500 in the AI for Bharat hackathon.

• What is multi-model fallback?

  A resilience pattern: when the primary model is rate-limited or unavailable, requests fall back to a secondary model. Learning Copilot falls back Nova Pro → Nova Lite so the assistant stays responsive under load.

• What is AHTML and how does it relate to RAG?

  AHTML (Agentic HTML) is a sixteen-package npm scope plus a Python SDK that emits a canonical, agent-ready snapshot of any website. Its schema ships RAG-ready document chunks with stable IDs and byte ranges, and LangChain loaders in both TypeScript and Python turn any AHTML site into vector-store-ready documents.

• Does he work on AI search and crawler optimization?

  Yes. AHTML emits MCP, OpenAPI, JSON-LD, and llms.txt so AI crawlers and agents can read a site cheaply, work that overlaps directly with AI search optimization and generative-engine visibility.

Keep reading: AHTML · whatbroke · AHTML vs llms.txt · Next.js SEO · projects & experience