Back to all work
— Project 03
Internal AI · RAG

AIVA

Document Intelligence Platform

AIVA is a document-intelligence platform for internal company documents: HR, contracts, financials, and asset records. The hard part was not orchestration; it was making messy PDFs, scans, blueprints, and rotated tables usable. I built a custom Textract parser that turns those files into tagged, searchable content, then a LangGraph RAG pipeline that answers over it without flooding the model context. It now runs as a multi-tenant platform with streaming chat and a live SharePoint crawl.

Role
Solo Founder + Lead Engineer
Period
2024 to present
Status
Production
AWS TextractMicrosoft GraphLangGraphWeaviateFastMCPSvelteKit
— Chapter 01
System shape

How the system fits together.

Click a block to zoom in
AIVA turns messy enterprise documents into searchable knowledge, then answers questions over them. The ingestion side now crawls a full company SharePoint.
Fig. 01 — AIVA architecture
— Chapter 02
Decisions and outcomes

The calls that shaped it.

  1. 01

    Solve the data problem first

    A custom Textract parser classifies every region of every page; tables become their own CSV files, figures are pulled by exact bounding box and read by a vision model (diagrams → Mermaid), and each document lands as one clean, normalized folder. Everything else stands on it.

  2. 02

    Tables stay tables

    Inlining tables burns context and drowns the search, so each stays a file behind a tag the chunker can’t split — and an MCP server turns it into a tool the model queries at runtime to filter, aggregate, and join real rows.

  3. 03

    Routed by cost

    A lookup answers from a template in ~200 ms; a real question runs the full path — three-layer discovery (keyword catalog → summaries → filtered hybrid), RRF fusion, optional Cohere rerank, compose, then verify the answer against its own evidence. Feedback tunes the ranking.

  4. 04

    Built to scale

    From a 5-document prototype to a live company SharePoint crawl: one orchestrator handles local files and SharePoint, skips unchanged content, keeps moved documents on the same identity, and streams crawl progress into the console.

  5. 05

    Run like production, not a demo

    Every sub-project goes spec → adversarial review by a second model → plan → TDD, one commit each — and the docs say plainly what’s proven by tests versus what hasn’t had a real end-to-end run.

— Aside
The interesting work isn't the stack. It's the boundaries.
— Chapter 03
How it runs

What it runs on.

AI / RAG
LangGraphWeaviateCohereOpenAI GPT-4.1FastMCPLangSmith
Data
AWS TextractSQLite FTSMicrosoft Graph · SharePoint
App
Python · FastAPISvelteKitDocker · Redis