Answers from your own documents

Organisations whose knowledge lives in documents, and teams that answer the same questions every week.

Generic chatbots don't know your policies and confidently make things up. Retrieval fixes that only when chunking, retrieval and prompting are tuned to your documents — and every answer points back to its source.

• +Ingestion pipeline for your formats (PDF, DOCX, web pages, exports)
• +Chunking and embedding strategy tuned to your content
• +Vector store setup — Pinecone, ChromaDB or FAISS
• +Answers with source citations, and a clear “I don't know” path
• +An evaluation set built from questions your team actually asks
• +Chat or search UI, or an API for your existing product

• LangChain
• Pinecone
• ChromaDB
• FAISS
• FastAPI
• Next.js

I'll work with your existing stack where it makes sense — these are defaults, not requirements.

BUBT AI Assistant · Open source · 2025

A university chatbot that answers from official documents — and cites them.

1. 01 · In / out

   University PDFsHandbooks, admission guides
2. 02 · Code

   Parse & chunk1,000 chars · 200 overlap
3. 03 · Model

   EmbedGemini text-embedding-004
4. 04 · Data

   FAISS indexCosine similarity · persisted
5. 05 · Model

   Grounded answerTop-5 chunks → Gemini
6. 06 · In / out

   Answer + sourcesCited documents

1. 01

   Intro call

   30 minutes, free. You walk me through the workflow; I ask about your data, your systems and what “done” means.

2. 02

   Written scope

   A short document: what gets built, what's explicitly out, milestones, and a price. No surprises later.

3. 03

   Build in the open

   A demo on a live staging URL every week. You see working software, not status reports.

4. 04

   Handover

   Your repository, deployment, documentation and tests. Support after launch if you want it.

• →Deterministic where it must be right. Money, permissions, routing and anything with consequences is decided by plain code that can be tested and audited. Models read, interpret and draft — they don't get the final say.
• →Guardrails are code, not prompts. A system prompt is a request. Validation, schema checks and post-processing are guarantees. Every model output is checked before a user or a database sees it.
• →It keeps working when the model doesn't. Providers time out, keys expire, outputs drift. I design a safe fallback path for each model call so the product degrades gracefully instead of failing loudly.