Netflix wiz creates app to slash AI bills, then open sources it
Netflix engineer Tejas Chopra built Project Headroom to trim redundant tokens before prompts hit an LLM; he says it has saved users about $700,000.
Intelligence analysis by GPT-5.4 Mini
Project Headroom sits in the developer workflow and compresses logs, JSON, file trees and other boilerplate before they reach an LLM. Chopra says it can reverse-compress and preserve output while cutting token bills, and Netflix teams already use it.
Imagine a backpack packed with the same worksheet over and over. Headroom is like a helper that removes the extra copies before the backpack gets too heavy.
It does that for AI chats and code tools. It looks at logs, tables, and other messy computer text, then keeps the parts that matter and trims the rest.
That matters because AI charges can go up when too much stuff is sent in. Chopra says trimming the extra words has already saved a lot of money, like paying only for the school supplies that are actually used.
Analysis
What Headroom does
Netflix senior engineer Tejas Chopra built Project Headroom to reduce the amount of redundant text that gets sent into an LLM. The idea is simple: a lot of prompt payloads are not actually useful to the model, but they still count toward token usage and cost.
Chopra says the app can strip boilerplate from logs, MCP tool output, database responses, file trees, code, and other repetitive structures before they reach the model. It runs locally as a proxy and fits into a developer workflow, including command-line use.
Why he built it
Chopra says the trigger was a $287 Claude Sonnet bill from a personal project. On closer inspection, he found a large share of the context was redundant: verbose JSON schemas, repeated metadata, and nested templates rather than meaningful instructions.
He argues that much of this material is "compressible data masquerading as text," and that the real opportunity is not just shorter prompts but smarter handling of data that does not need to be preserved in full.
How it works
Headroom first uses a component called CacheAligner to detect what has changed and send only the new pieces, instead of refreshing large sections of mostly unchanged context. After that, a router sends different content types to specialized compressors. The article says there are separate compressors for code, JSON, and DOM content, plus "squashers" that decide what matters based on statistical analysis.
Chopra also emphasizes reversibility, so compressed material can be expanded again when needed. That distinguishes it from some other token-saving tools and services.
Traction so far
The project is not an official Netflix product, but several internal teams use it and outside projects do too. Chopra said at Open Source Summit that Headroom has saved an estimated $700,000 for users and now covers about 200 billion tokens. It has been open source since January, reached version 0.22, gathered about 2,000 GitHub stars, and been forked more than 120 times.
Key points
- Project Headroom trims redundant tokens before text reaches an LLM.
- Chopra says the tool has saved users an estimated $700,000.
- Netflix teams use it even though it is not an official Netflix project.
- It targets logs, JSON, database output, file trees, and code.
- The project is open source and has drawn GitHub stars and forks.
