The Problem
Document assistants are only useful when users can trust where an answer came from. Blue is built around private document workflows: upload a document, extract and chunk the text, retrieve relevant context, and generate a grounded answer inside a user-scoped workspace.
The repository is private, so this page does not claim a public code link or production deployment.
What I Built
I built the full-stack assistant around private document ingestion, per-user retrieval, and conversational answers grounded in uploaded material. The repository remains private, but the implementation shape is verified from source and described without publishing prompts, keys, private documents, deployment endpoints, or account details.
Project Overview
Blue is an independent private AI assistant project centered on document-grounded answers, user isolation, and local-first deployment options.
My Contribution
- Built document upload, text extraction, chunking, embedding generation, and FAISS vector retrieval for user-scoped knowledge.
- Implemented authentication, authorization, PostgreSQL persistence, per-user document isolation, saved chat history, and conversational memory.
- Added source-grounded answer generation through a provider gateway that can use configured API providers and falls back to local Ollama/LLaMA 3 when external keys are unavailable or exhausted.
System Architecture
Private RAG workflow
A private document moves through parsing, chunking, retrieval, and answer generation inside a scoped assistant workflow.
The diagram is high level and contains no private prompts, documents, keys, or repository details.
Private document workspace
A safe sample interface showing document status, grounded answer state, and source references.
3 sample documents
Scoped retrieval
Source-linked answers
Visual uses safe sample data created for this portfolio page.
Technical Decisions
Separate vector data, uploaded documents, database records, and API access by authenticated user scope.
- Retrieval must never mix one user's content with another user's documents.
- Per-user isolation adds storage management and cleanup complexity.
- Every retrieval request is constrained to the authenticated user's document workspace.
How I Work
Engineering Challenge
- A document assistant has to answer from private user material without mixing documents across accounts.
- I separated user vector stores, text directories, database records, chat sessions, notes, feedback, and authenticated API access by user scope.
- Uploads are extracted, chunked, embedded, retrieved from FAISS, and passed into an answer prompt with recent conversation context.
- The assistant can preserve answer grounding and conversational continuity while keeping the repository and private documents unpublished.
Quality & Testing
Quality Evidence
Product engineering evidence
- Private source includes upload, retrieval, chat, notes, feedback, usage, and admin-management paths; no retrieval-quality metrics are public yet.
- The private project includes GitHub Actions and Docker deployment documentation; run details and endpoints are not published.
- The verified architecture includes upload, extraction, chunking, embeddings, FAISS, PostgreSQL, auth, per-user isolation, API-provider routing, Ollama fallback, and conversational memory.
Public page quality
- The sanitized visual includes accessible labels and text equivalents.
- The safe sample visual is responsive inside the shared case-study layout.
- No private repository URL, deployment endpoint, prompt, key, document, or generated vector asset is bundled.
Results & Impact
Blue demonstrates a built AI product path: private document ingestion, retrieval, source grounding, persistence, authentication, user isolation, API-provider routing, local Ollama fallback, and Dockerized setup.
What I Learned
AI case studies need both implementation detail and restraint. The strongest story is not that the model is broadly intelligent; it is that the system controls what context is retrieved, who can access it, and how answers are grounded.