Data entry ops · Production

DoorDash builds LLM guardrail system to automate restaurant menu transcription from photos

The problem

DoorDash previously relied on humans to manually transcribe restaurant menus from photos, a process described as costly and time-consuming. LLMs alone could not achieve the required high accuracy due to diverse menu structures, incomplete menus, and low-quality photos.

First attempt

LLMs used as standalone transcription tools produced errors due to inconsistent menu structures, incomplete menus, and low photo quality. Intensive efforts to improve LLM accuracy still required too much time and investment to meet production standards.

Workflow diagram · grounded in source
1
Menu photo submitted
trigger
“streamlining efficient updates through submitted menu photos”
2
OCR text extraction
ai_action
“uses optical character recognition, or OCR, to extract text from a menu image”
3
LLM item extraction and structuring
ai_action
“passed over to an LLM for item-level information extraction and summarization, creating a structured data format”
4
Guardrail model quality check
validation
“guardrail model must learn how each menu photo interacts with both OCR and LLM summarization”
5
Route by accuracy threshold
routing
“Transcribed information becomes readily available for the menu photos that pass the auditing threshold for accuracy. For those that don't pass, the system moves photo menus to the human process.”
6
Human transcription fallback
human_review
“the system moves photo menus to the human process”
Reported outcome

DoorDash deployed a partial automation pipeline combining LLM transcription with an ML guardrail model that routes high-confidence transcriptions to production automatically and low-confidence ones to human review, improving efficiency without sacrificing quality and enabling rapid adoption of new AI models.

Reported metrics
Manual transcription effortcostly and time-consuming
Automation ratiohigher ratio of automation while ensuring quality
Reported stack
OCRLightGBMResNetDiTCNN
Source
https://careersatdoordash.com/blog/doordash-llm-transcribe-menu/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DoorDash deployed a partial automation pipeline combining LLM transcription with an ML guardrail model that routes high-confidence transcriptions to production automatically and low-confidence ones to human review, im…

What tools did this team use?

OCR, LightGBM, ResNet, DiT, CNN.

What results were reported?

Manual transcription effort: costly and time-consuming; Automation ratio: higher ratio of automation while ensuring quality (source-reported, not independently verified).

What failed first in this deployment?

LLMs used as standalone transcription tools produced errors due to inconsistent menu structures, incomplete menus, and low photo quality.

How is this data entry ops AI workflow structured?

Menu photo submitted → OCR text extraction → LLM item extraction and structuring → Guardrail model quality check → Route by accuracy threshold → Human transcription fallback.