Back office ops · Production

Tiger Data builds a production AI Slack bot with Pydantic AI and Logfire

The problem

Tiger Data's entire company operates on Slack, but after reaching a certain size it became nearly impossible for employees to catch up on ongoing conversations, leaving engineers drowning in channel history with no way to surface relevant context instantly.

First attempt

Tiger Data initially built without distributed tracing, which made it very difficult to debug agentic systems in production. They also tried Jaeger for tracing first but found the developer experience poor.

Workflow diagram · grounded in source
1
Slack mention triggers agent
trigger
“When a Slack mention event comes in”
2
Load MCP servers from config
integration
“the agent loads MCP servers from configuration, creates a Pydantic AI agent with those servers as toolsets”
3
LLM calls tools via MCP
ai_action
“Pydantic AI calls the LLM, handles tool invocations, retries failures”
4
Response returned to Slack
output
“runs the agent with a user prompt, and returns a response”
5
Logfire captures full trace
feedback_loop
“Logfire captures every step”
Reported outcome

Tiger Data built a production AI Slack bot handling thousands of concurrent conversations, with more than half of the company using it daily within 6 weeks, reduced debugging time, and significant development time savings from LLM provider abstraction.

Reported metrics
Concurrent conversations handledthousands of concurrent conversations
Company daily adoption within 6 weeksmore than half of the company uses it daily
Logfire setup timeless than one hour
Debugging timeReduced debugging time
Show all 5 reported metrics
concurrent conversations handledthousands of concurrent conversations
company daily adoption within 6 weeksmore than half of the company uses it daily
Logfire setup timeless than one hour
debugging timeReduced debugging time
development time savedsignificant amount of development time
Reported stack
Pydantic AILogfireSlackOpenTelemetryPostgreSQLpsycopghttpxClaude CodeJinja2JaegerClaude 4.5 SonnetGPT-4oMCPSalesforceGitHubLinear
Source
https://pydantic.dev/articles/tiger-data-ai-slack-bot-pydantic-logfire
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Tiger Data built a production AI Slack bot handling thousands of concurrent conversations, with more than half of the company using it daily within 6 weeks, reduced debugging time, and significant development time sav…

What tools did this team use?

Pydantic AI, Logfire, Slack, OpenTelemetry, PostgreSQL, psycopg, httpx, Claude Code, Jinja2, Jaeger.

What results were reported?

Concurrent conversations handled: thousands of concurrent conversations; Company daily adoption within 6 weeks: more than half of the company uses it daily; Logfire setup time: less than one hour; Debugging time: Reduced debugging time (source-reported, not independently verified).

What failed first in this deployment?

Tiger Data initially built without distributed tracing, which made it very difficult to debug agentic systems in production.

How is this back office ops AI workflow structured?

Slack mention triggers agent → Load MCP servers from config → LLM calls tools via MCP → Response returned to Slack → Logfire captures full trace.

Related back office ops cases