Back office ops · Production

Tradestack builds an AI-powered quoting assistant for trades businesses using LangGraph Cloud

The problem

Trades businesses face an extensive administrative burden when creating project quotes, a process that typically consumes 3.5 to 10 hours per project and involves analyzing floor plans, reviewing images, estimating effort, and calculating material prices.

First attempt

Building an AI agent system that consistently performed at high quality with diverse user inputs was not straightforward, with explicit failure modes including input ambiguity, variable workflow paths between users, and inconsistent LLM planning and routing.

Workflow diagram · grounded in source
1
User submits via WhatsApp
trigger
“Given the widespread adoption of WhatsApp, especially among non-tech-savvy users, Tradestack chose it as their primary interface”
2
Supervisor node plans task
ai_action
“a hierarchical multi-agent system with a supervisor node that expanded on user queries and created plans based on the task's goals”
3
Sub-graph routing by use case
routing
“Tradestack customized instructions and pathways in their cognitive architecture, selecting sub-graphs depending on specific use cases”
4
Multimodal input processing
ai_action
“This setup maintained input flexibility (voice, text, images, documents) while producing accurate, personalized client quotes”
5
Edge-case human intervention
human_review
“When edge cases arose—such as users requesting materials unavailable in the UK—the system would trigger manual intervention. Tradestack's team could then step in via Slack or directly in LangGraph Studio to adjust the conversation.”
6
Aggregator combines outputs
output
“An aggregator node was added to combine outputs from various intermediate steps, providing a consistent tone of voice in all communications.”
7
LangSmith evaluation loop
feedback_loop
“By setting up node-level and end-to-end evaluations in LangSmith, Tradestack could experiment with different models for the planning node and see which models performed the best”
Reported outcome

Tradestack launched their MVP in 6 weeks to a community of 28,000+ users, secured their first paying customers, improved end-to-end performance from 36% to 85% through rapid iteration, and saved two weeks of internal testing time.

Reported metrics
End-to-end performancefrom 36% to 85%
MVP build and launch time6 weeks
Community size at launch28,000+
Quote creation time before3.5 to 10 hours
Show all 7 reported metrics
end-to-end performancefrom 36% to 85%
MVP build and launch time6 weeks
community size at launch28,000+
quote creation time before3.5 to 10 hours
quote creation time targetunder 15 minutes
internal testing time savedtwo weeks
first paying customersfirst paying customers
Reported stack
LangGraphLangGraph CloudLangGraph StudioLangGraph TemplatesLangSmithLCELSlackgpt-4-0125-previewgpt-4oWhatsApp
Source
https://blog.langchain.dev/customers-tradestack/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Tradestack launched their MVP in 6 weeks to a community of 28,000+ users, secured their first paying customers, improved end-to-end performance from 36% to 85% through rapid iteration, and saved two weeks of internal…

What tools did this team use?

LangGraph, LangGraph Cloud, LangGraph Studio, LangGraph Templates, LangSmith, LCEL, Slack, gpt-4-0125-preview, gpt-4o, WhatsApp.

What results were reported?

End-to-end performance: from 36% to 85%; MVP build and launch time: 6 weeks; Community size at launch: 28,000+; Quote creation time before: 3.5 to 10 hours (source-reported, not independently verified).

What failed first in this deployment?

Building an AI agent system that consistently performed at high quality with diverse user inputs was not straightforward, with explicit failure modes including input ambiguity, variable workflow paths between users, a…

How is this back office ops AI workflow structured?

User submits via WhatsApp → Supervisor node plans task → Sub-graph routing by use case → Multimodal input processing → Edge-case human intervention → Aggregator combines outputs → LangSmith evaluation loop.

Related back office ops cases