Large websites scatter useful information across many routes. Generic chatbots do not know what the user is currently seeing, cannot reliably cite the right page, and often require risky or unrealistic access to internal systems.
Safe Site-Aware AI Layer
I designed and built a reusable AI layer that can be adapted to existing websites without direct database access. It combines site-wide knowledge with the user's live current page, answers from accessible content, cites exact source pages, and can take the user to the relevant route.
WORKING PRODUCT EVIDENCE
Real interfaces, not mockups.
The same core is shown running against a university portal and a learning platform with isolated adapters.
The authenticated Oracle APEX portal remains the source of truth and keeps its existing permissions.
The assistant identifies the signed-in user and explains information from the page currently open.
A grounded answer gives practical steps and exposes the relevant page as an actionable source.
The same core runs against a different learning platform, interface, and knowledge structure.
CONTEXT & DELIVERY
The engineering story.
A Chrome Manifest V3 extension and FastAPI service read the visible page at question time, retrieve supporting knowledge from sanitized site content, prioritize live context, and return grounded answers with exact-page navigation. SQLite and ChromaDB keep each website and user workspace isolated.
Product concept, system architecture, site adapters, safe discovery, URL canonicalization, retrieval and grounding, Chrome extension UX, automated tests, and deployment.
The same core architecture was proven on an authenticated Oracle APEX university portal and Taqat Academy. Both integrations work from user-accessible content without collecting credentials or connecting the assistant directly to either website database.
SYSTEM DESIGN
How the product works.
Discover accessible site knowledge
A site adapter maps allowed routes, titles, links, fields, tables, controls, and meaningful content without entering the private backend.
MAPRead the current page
At question time the extension captures what the signed-in user can currently see, so the answer understands the live task in front of them.
DOMSanitize and isolate
Session values and sensitive URL fields are removed, while SQLite metadata and Chroma vectors stay inside a separate site and user workspace.
SAFERetrieve live-first
Current-page context outranks stored copies, then hybrid search adds supporting evidence from the wider site knowledge base.
RAGAnswer and navigate
The response includes a recognizable source title and opens the exact canonical page before a route-level fallback is considered.
SOURCEENGINEERING DECISIONS
What keeps it trustworthy.
No password collection
The user signs into the website normally; the assistant works inside that existing session.
One workspace per installation
Page storage, retrieval, deletion, and vector search share the same workspace boundary.
Sources remain actionable
Answers include exact page navigation instead of an unverifiable block of generated text.
05 / TECHNOLOGY
Tools chosen for the job.
ChromaDBChrome ExtensionDockerFastAPIGeminiJavaScriptRAGSQLite