How I’m Using JEV in Production (Plan)
An architectural proposal for leveraging fast, cheap AI API calls in order to structure data into JSON in under a second.
The slowest part of scanning with thepom.app
Is the time it takes for LLM even at 800 TKPs to turn OCR noise into a clean JSON that can be retrieved from the db.
JEV’s decision making insane speeds and cost are suited for toJSON implementation.
Jev does not output JSON, but used algorithmically with parallel requests. I expect to reduce my latency from 4-6 seconds down to 1 second. Maybe even less.
One label, straight off the video frames
INGREDIENTS: WATER, GLYCERIN,
SODIUM LAURYL
SULFATE, NET WT 1.7 OZToday
Same label + 644-token rulebook
gemini-2.5-flash-lite
One opaque call.
The whole list typed back out.
Re-typed, token by token
Planned · JEV
Same label, split on commas
jev · one request
- INGREDIENTS:
- 0.02 Drop
- WATER
- 0.99 Keep
- GLYCERIN
- 0.98 Keep
- SODIUM LAURYL SULFATE
- 0.97 Keep
- NET WT 1.7 OZ
- 0.01 Drop
Keep the selected rows, in code
Both paths produce the identical result
{"ingredients": ["water", "glycerin", "sodium-lauryl-sulfate"]}Time
- Today
- 4-6 s
- Planned
- 70-500 ms
8× faster, at worst
Cost / 1,000 scans
- Today
- $0.21
- Planned
- $0.045
4.7× cheaper
I will report on these findings and the algorithm we decide to go with, potentially open source a Jev toJson open source repo to support those seeking to explore how this new paradigm and model can be applied to solve older problems more efficiently.
BTW, there’s evidence that harmful ingredients are added to everyday products, pom is here to help you decide based on scientific research which products to pick.
Follow me and connect to learn more! I love making friends on linkedin so shoot me a message and let’s call and talk about applied AI and leveraging AI!