Carbon bot

Carbon Bot solves a problem every technical sales and support team knows well, your product catalog is locked inside PDFs and data sheets that nobody wants to read. When a customer asks "what's the tensile strength of product CF-200, and does it meet aerospace tolerances?" The answer exists somewhere in your documentation, but finding it quickly, accurately, and without burdening a specialist is the real challenge. Carbon Bot turns your existing product data into a conversational knowledge base. You upload a data sheet once, and from that point forward anyone on your team, or your customers directly, can ask plain-English questions and get precise, sourced answers in seconds, without ever leaving a web browser.

Under the hood, Carbon Bot uses a technique called Retrieval-Augmented Generation (RAG), rather than asking an AI to guess from general knowledge, it retrieves the relevant section of your actual product data first, then uses that as the basis for its answer. The stack was chosen deliberately for cost predictability and operational simplicity. Product data is stored and searched in PostgreSQL using the pg_vector extension, which keeps the entire knowledge base inside a single managed database with no separate vector service to maintain.

Embeddings, the mathematical representations that power semantic search, are generated locally using Ollama, eliminating per-query API charges that add up fast at scale. Inference (the actual answer generation) runs through spot instances, giving access to frontier-quality language models at a fraction of the cost, with usage that stays predictable and easy to budget.

The frontend is deliberately lightweight, vanilla JavaScript with no framework overhead, served by a thin Sinatra layer in Ruby that handles routing and streams responses to the browser in real time.

The heavy lifting of retrieval and prompt construction lives in a dedicated Python backend, where the ML ecosystem is mature. Each layer does exactly one job, which makes the system easy to reason about, cheap to host, and straightforward to extend as your catalog grows.