Ivan Novikov

This account is verified. Learn more

@d0znpp

AI Search | @xerj_org | SSRF bible author; Bug Hunter (Google/Facebook/Twitter/Yandex/Tesla); Masters in Physics, MSU; Founder at @Wallarm

xerj.org Joined April 2010

1.4K Following 7.3K Followers 11K Posts

Latest posts 5

Pinned post

Ivan Novikov @d0znpp 2h

I've built a new AI model that is not allowed to generate text. Instead, it produces a verifiable program that answers your prompt using the original reference.

XERJ Reader-1 is a neurosymbolic model for working with data. It points to the exact words in your document, writes a small program over them, and exact code computes the answer:

• 20-90x faster
• 10-1,000x cheaper (output tokens free)
• every answer is a program you can replay

The shortest path to automating data work.

2:05

Ivan Novikov @d0znpp 3d

XERJ Query-1 gives you the program behind the answer

ask a question about facts you’ve stored. the model writes a short set of operations, and XERJ runs them to get the result

you get a fast answer and a way to check how it got there

→ look up the relevant records
→ compare values or count matching items
→ execute the program
→ inspect the steps, with source records attached to lookups

model handles the question. the engine handles the execution. you can follow a lookup back to the record it used and check the value yourself

we built this into XERJ, running locally on four CPU cores

try it yourself by giving your agent this prompt

> Install XERJ with the Query-1 model using github.com/xerj-org/xerj, write three facts about this project, and ask a question about them. Show me the program and the evidence it returns

full XERJ Query-1 breakdown below ↓

0:30

Ivan Novikov Verified account @d0znpp 6d

Your AI shouldn't have to read thousands of records to answer one question

We built XERJ Query-1 to turn questions into executable programs instead of generated answers

A breakdown of how it works and why it costs hundreds of times less than RAG in tests x.com/i/article/2118…
View quoted post on X

Ivan Novikov @d0znpp 4d

we built XERJ Query-1, an AI model that turns questions into programs on four CPU cores

ask about facts you've stored. Query-1 writes a short program, and XERJ runs it to get answer

you get answers faster, with the source records right there to check

here's what happens inside

→ Query-1 reads your question and the available records
→ writes steps to look up values, compare them or count
→ XERJ executes those steps
→ lookups return the records they used, and you can inspect the execution trace

you can see which facts it used and how it reached the result, without digging through the documents yourself

try it yourself by giving your agent this prompt:

> Install XERJ with the Query-1 model using xerj.org/llms.txt, write three facts about this project, and ask a question about them. Show me the program and the evidence it returns

full XERJ Query-1 breakdown ↓

0:40

Ivan Novikov Verified account @d0znpp 6d

Your AI shouldn't have to read thousands of records to answer one question

We built XERJ Query-1 to turn questions into executable programs instead of generated answers

A breakdown of how it works and why it costs hundreds of times less than RAG in tests x.com/i/article/2118…
View quoted post on X

Ivan Novikov @d0znpp 4d

give Sonnet XERJ Query-1 and the cheaper model starts setting the pace

both get the task of recreating the IGLOO website with its interactive 3D scenes, with Query-1 helping Sonnet get answers from stored project facts while it works

comparison ends at $21.84 and 27m 36s for Sonnet. Opus reaches the same result at $59.37 and 34m 30s.

63.2% less spending and almost 7 minutes saved.

→ Sonnet asks a question about the project
→ Query-1 turns it into a short program over stored facts
→ the engine returns an answer with its sources
→ Sonnet uses that information to keep building

scene ownership, camera progress, cleanup handlers. those details can send an agent through another round of fixes when it guesses wrong

giving Sonnet those answers helps it spend more of the task implementing

ask Claude Code or Codex to set up XERJ Query-1, store a few facts about your project and show you the answers with evidence

full breakdown below ↓

0:34

Ivan Novikov Verified account @d0znpp 6d

Your AI shouldn't have to read thousands of records to answer one question

We built XERJ Query-1 to turn questions into executable programs instead of generated answers

A breakdown of how it works and why it costs hundreds of times less than RAG in tests x.com/i/article/2118…
View quoted post on X

Ivan Novikov @d0znpp 5d

a million questions for $42 in CPU compute. XERJ Query-1 is making AI search bills look ridiculous

Opus 5.5 RAG setup comes out at $20,950 per million questions in the article’s comparison. almost 500× more

→ XERJ Query-1 turns each question into a short program
→ runs it against your stored facts
→ returns the answer with the records and steps behind it
→ runs locally on 4 CPU cores, with your data on your machine

ask Claude Code or Codex to install XERJ Query-1 and try a question about your project. have it show you the answer and the evidence

full breakdown below ↓

0:24

Ivan Novikov Verified account @d0znpp 6d

Your AI shouldn't have to read thousands of records to answer one question

We built XERJ Query-1 to turn questions into executable programs instead of generated answers

A breakdown of how it works and why it costs hundreds of times less than RAG in tests x.com/i/article/2118…
View quoted post on X