MetARAG

A GPU-accelerated document-intelligence platform for CCC Intelligent Solutions that grounds every answer in the exact source paragraph.

Role: Led a team of 5+ engineers
When: Aug 2025 - Dec 2025
Status: Delivered
Context: Client project for CCC Intelligent Solutions
Stack: Python
LangChain
Embeddings
GPU inference
RAG

The problem

CCC Intelligent Solutions needed reliable answers out of hundreds of PDFs, and an answer nobody can check isn't much use. Every answer had to point back to the paragraph it came from.

I led a team of 5+ engineers on it and designed the pipeline end to end, from parsing through generation.

How it works

1
Parsing
Extracts text and structure from the client's PDFs.
2
Chunking
Splits documents into retrievable pieces sized for the questions people actually ask.
3
Metadata enrichment
Attaches metadata to each chunk so retrieval can filter and rank on more than raw text similarity.
4
Embeddings and retrieval
Embeds the chunks and retrieves the best candidates for each question.
5
Source-grounded answers
Generates an answer and cites the exact paragraph it came from, so every claim can be checked.
6
Batched GPU inference
Runs the model on batches spanning many documents at once, instead of calling it document by document.

Decisions and tradeoffs

Batch inference across documents instead of per document. This one change cut document-processing time by about two thirds and inference latency by about a third on a pipeline carrying 100+ GB.
Citations at paragraph level. We tuned chunking, retrieval and the citation logic together so each answer resolves to its exact source paragraph.

Results

Result	What it measures	Source
93%	Retrieval precision across hundreds of client PDFs	project evaluation, tuning retrieval and citation logic
about two thirds	Less document-processing time on the 100+ GB pipeline, from batching GPU inference across documents	project measurements
about one third	Lower inference latency from the same batching change	project measurements

93%
Retrieval precision across hundreds of client PDFs
project evaluation, tuning retrieval and citation logic
about two thirds
Less document-processing time on the 100+ GB pipeline, from batching GPU inference across documents
project measurements
about one third
Lower inference latency from the same batching change
project measurements

A note on the code

This was client work, so the code is private and there's no repo to link. For public code that shows the same retrieval and evaluation skills, see ECI Pipeline and rag-redteam.

Next project

Chain-of-Thought on CLEVR

A controlled test of whether chain-of-thought supervision helps a small vision-language model reason about scenes, using BLIP-2 fine-tuned with LoRA.