Building a Guardrail System for LLM-based Menu Transcription

doordash · E-commerce · 2025 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Restaurant menu transcription automation

DoorDash wanted to automate restaurant menu-photo transcription into structured menu data without sacrificing the high accuracy required for partner menus and consumer ordering.

Constraints:

manual menu transcription was costly and time-consuming
restaurants use diverse and inconsistent menu structures
poor photo quality, incomplete menus, and confusing OCR order cause LLM transcription errors
the system needed high accuracy before automation could replace human work
new multimodal models were evolving quickly and the pipeline needed to adapt

Architecture

DoorDash first built an OCR plus LLM transcription MVP that extracts menu text and summarizes it into structured data. Because the LLM alone could not meet quality requirements, they added a traditional ML guardrail model that predicts whether each LLM transcription is accurate enough to automate. Photos that pass the threshold flow automatically; failures route to the human transcription process. The guardrail framework later supported multimodal GenAI transcription models as well.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  menu-photo-intake[Menu photo intake] --> ocr-extractor[OCR extractor]
  menu-photo-intake --> multimodal-transcription-model[Multimodal transcription model]
  ocr-extractor --> llm-transcription-model[LLM transcription model]
  ocr-extractor --> guardrail-feature-generator[Guardrail feature generator]
  llm-transcription-model --> guardrail-feature-generator
  multimodal-transcription-model --> guardrail-feature-generator
  guardrail-feature-generator --> guardrail-model[Guardrail model]
  guardrail-model --> automatic-transcription-path[Automatic transcription path]
  guardrail-model --> human-transcription-process[Human transcription process]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Menu photo intake`	Restaurant menu photos submitted for transcription and validation.	`OCR extractor`, `Multimodal transcription model`
`OCR extractor`	Extracts raw text from menu images for the OCR plus LLM pipeline.	`Menu photo intake`, `LLM transcription model`, `Guardrail feature generator`
`LLM transcription model`	Summarizes OCR text into structured item-level menu data.	`OCR extractor`, `Guardrail feature generator`
`Multimodal transcription model`	Alternative GenAI model path that processes menu photos with stronger context understanding but different failure modes.	`Menu photo intake`, `Guardrail feature generator`
`Guardrail feature generator`	Builds features from the menu photo, OCR output, and LLM summarization quality signals.	`OCR extractor`, `LLM transcription model`, `Multimodal transcription model`, `Guardrail model`
`Guardrail model`	Traditional ML classifier that predicts whether transcription accuracy is high enough for automation.	`Guardrail feature generator`, `Automatic transcription path`, `Human transcription process`
`Automatic transcription path`	Publishes structured transcribed menu data for photos that pass the accuracy threshold.	`Guardrail model`
`Human transcription process`	Fallback path for menus whose predicted transcription quality does not meet the threshold.	`Guardrail model`

Data flow

A menu photo is validated and sent through OCR/LLM or multimodal transcription. The system generates features about the photo, OCR output, and LLM summarization. A guardrail model scores whether the transcription meets the accuracy threshold. Passing transcriptions become available automatically; failing menu photos go to the human process.

3 supporting quotes

“The process first uses optical character recognition, or OCR, to extract text from a menu image, which is then passed over to an LLM for item-level information extraction and summarization”article, stated

“Our guardrail framework is based on a machine learning (ML) model that identifies whether an LLM transcription can achieve high accuracy.”article, stated

“For those that don’t pass, the system moves photo menus to the human process. This system marked our first step toward improving efficiency in the manual human processes without sacrificing quality.”article, stated

Why this architecture

DoorDash used a guardrail model because standalone LLM transcription could not meet the required accuracy on messy menus. A lightweight traditional ML classifier was cheaper and lower-latency than repeatedly improving the LLM before production, and it let DoorDash adopt new LLM or multimodal models while preserving a quality gate.

3 supporting quotes

“given our high accuracy standards, we still need a tremendous amount of time and investment to improve the LLMs, postponing the realization of their value.”article, stated

“The key to ensuring an LLM’s accuracy is to build an LLM system with a suitable automatic guardrail process and LLM itself, instead of having LLM being a standalone product.”article, stated

“It enables AI practitioners to swiftly leverage newly released LLMs while mitigating potential risks that may impact the final product quality. In the meantime, the clever use of traditional ML in this system offers advantages in both low latency and cost efficiency.”article, stated

Models used: LLMs · LightGBM · ResNet · Vision Transformers · multimodality models

Scale

The system targets restaurant partners' menu-photo update workflow.
“To ensure accuracy and alignment with their latest offerings, DoorDash’s restaurant partners must actively maintain their menus. This can be challenging, however, for business owners who already are managing demanding daily operations.”article, stated
DoorDash evaluated a large number of menu photos with humans to find error patterns.
“Through human evaluation of a large number of menu photos, a reasonable proportion of menus can be transcribed with various errors”article, stated
The first guardrail model was followed by six months of rapid GenAI model evolution.
“During the six months following the development of our first guardrail model, there was rapid evolution in the generative AI world”article, stated

Alternatives considered (roads not taken)

Manual menu transcription only

Rejected because: Manual work was costly and time-consuming for restaurant menu updates.

“Previously, we relied on humans to transcribe and update restaurant menus manually, which is costly and time-consuming.”article, stated

Standalone LLM transcription

Rejected because: LLMs lacked menu-structure familiarity and could not reliably handle complicated scenarios at required accuracy.

“we require very high transcription accuracy, which is difficult for an LLM to achieve because of its lack of familiarity with the variety of menu structures”article, stated

Complex neural guardrail architectures

Rejected because: LightGBM outperformed the neural models and ran fastest, partly because labeled data was limited.

“the simplest model — Light Gradient-Boosting Machine, or LightGBM for short — outperforms all models while maintaining the fastest run time.”article, stated

Use multimodal models without guardrails

Rejected because: Multimodal models had better context understanding but higher failure rates on bad-quality photos.

“multimodality models are great at context understanding but more prone to errors when handling bad-quality photos, resulting in overall higher transcription failure rates.”article, stated

Patterns

`guardrails-io-filtering` (safety)

A traditional ML guardrail predicts whether an LLM menu transcription is accurate enough to use automatically.

“The guardrail layer serves as an effective control mechanism of LLMs that enables LLM applications to run at scale with high accuracy.”article, stated

`confidence-thresholding` (routing)

Transcriptions that pass the auditing threshold proceed automatically; others route to human work.

“Transcribed information becomes readily available for the menu photos that pass the auditing threshold for accuracy.”article, stated

`human-in-the-loop-gate` (workflow)

Human transcription remains the fallback for low-confidence menu-photo cases.

“For those that don’t pass, the system moves photo menus to the human process.”article, stated

`fallback-degradation` (reliability)

The pipeline degrades from automated transcription to human transcription when the model predicts quality risk.

“partial automation transcription pipeline to combine human and ML transcriptions”article, stated

`fine-tuning-adaptation` (modeling)

DoorDash identifies fine-tuning custom LLM/multimodal models on manually transcribed data as the next improvement path.

“Because we have an increasing availability of manually transcribed data, we could extend its use to fine-tune custom LLM/multimodal models.”article, stated

Tradeoffs

Add a traditional ML guardrail

Chose LightGBM-based quality prediction before automation over waiting for the LLM to become accurate enough alone.

Rationale: The guardrail let DoorDash realize automation value sooner while maintaining quality.

“we still need a tremendous amount of time and investment to improve the LLMs, postponing the realization of their value.”article, stated

Use partial automation

Chose automatic path only for threshold-passing photos over fully automatic transcription for every menu photo.

Rationale: Routing low-confidence cases to humans improved efficiency without sacrificing quality.

“This system marked our first step toward improving efficiency in the manual human processes without sacrificing quality.”article, stated

Retain the guardrail across model generations

Chose a model-agnostic quality gate over hard-coding the pipeline to one transcription model.

Rationale: The guardrail let DoorDash quickly adopt newly released AI models while balancing their failure modes.

“our guardrail model framework has allowed us to leverage newly released state-of-the-art AI models quickly.”article, stated

Gotchas (what bit them)

LLM transcription errors clustered around messy menu photos.

Root cause: Inconsistent menu structure, incomplete menus, and poor photo quality confuse OCR and LLM summarization.

Resolution: Generate guardrail features from the photo, OCR output, and summarization quality signals.

“the LLM created transcription errors primarily when it encountered three sub-optimal types of menu photos”article, stated

Complex vision models did not beat a simpler guardrail.

Root cause: Limited labeled data made it hard to exploit more complex neural designs.

Resolution: Use LightGBM for the guardrail because it performed best and ran fastest.

“A key reason for its poor performance is that we have limited labeled data, making it difficult to take full advantage of more complex model designs.”article, stated

New multimodal models changed tradeoffs but did not eliminate quality risk.

Root cause: Multimodal models improved context understanding but failed more on bad-quality photos.

Resolution: Keep the guardrail framework so new models can be tested and routed safely.

“Each generation transcription model has unique advantages and shortcomings, but none significantly outperforms the others.”article, stated

Evaluation

DoorDash evaluated menu transcription through human assessment of menu photos, guardrail model accuracy prediction, and comparisons of LightGBM, ResNet, ViT, OCR+LLM, and multimodal model tradeoffs.

Metrics: average transcription accuracy across test menu photos · percentage of transcriptions meeting accuracy requirements · guardrail model runtime · automation ratio while preserving quality

“Table 2 below shows the comparison among different model architectures based on two main metrics: average transcription accuracy across all test menu photos and percentage of transcriptions that met accuracy requirements.”article, stated

“helps the system steadily reach a higher ratio of automation while ensuring quality.”article, stated

Outcomes

DoorDash moved from manual-only transcription to a partial automation pipeline
LightGBM was selected as the best-performing and fastest guardrail model
the guardrail framework let the team test OCR+LLM and multimodal transcription models while preserving quality routing
low-confidence menus still route to humans

The source does not publish the exact automation percentage, but it describes restaurant-partner menu updates and a rising automation ratio under guardrail control.

“The guardrail ML model has proven most viable for achieving these purposes.”article, stated

“we are seeing improvement in our current pipeline, even as we explore additional options for optimizing and improving the performance”article, stated