Workflow · Production

Challenges of building an LLM product: a Chrome extension for English writing improvement

The problem

Building LLM products for production requires navigating unreliable API uptime and latency, unpredictable prompt engineering, hallucinations, a lack of evaluation metrics for outputs, and API endpoint deprecation that can force costly prompt rebuilds.

First attempt

The team built a full set of crafted, fine-tuned, few-shot prompts for OpenAI's DaVinci-002 model, which began working well, but OpenAI then deprecated that endpoint — described as not an easy transition.

Workflow diagram · grounded in source
1
Non-native speaker triggers extension
trigger
“a Chrome extension aimed at improving English writing skills for non-native speakers”
2
Few-shot prompts sent to LLM API
ai_action
“We created a whole set of crafted, finetuned, few-shot prompts for that API”
3
Natural language output returned
output
“the model's output, which is in natural language”
Reported outcome

The post recommends best practices — finetuning over API prompts for complex cases, few-shot and chain-of-thought prompting, vector databases for large data, and versioned prompts — but reports no quantified outcome for the Chrome extension itself.

Reported metrics
Prompt performance after fine-tuningstarted working really well
Reported stack
OpenAI APIDaVinci-002
Source
https://mlops.community/blog/it-worked-when-i-prompted-it-or-the-challenges-of-building-an-llm-product
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The post recommends best practices — finetuning over API prompts for complex cases, few-shot and chain-of-thought prompting, vector databases for large data, and versioned prompts — but reports no quantified outcome f…

What tools did this team use?

OpenAI API, DaVinci-002.

What results were reported?

Prompt performance after fine-tuning: started working really well (source-reported, not independently verified).

What failed first in this deployment?

The team built a full set of crafted, fine-tuned, few-shot prompts for OpenAI's DaVinci-002 model, which began working well, but OpenAI then deprecated that endpoint — described as not an easy transition.

How is this workflow AI workflow structured?

Non-native speaker triggers extension → Few-shot prompts sent to LLM API → Natural language output returned.