Eldros

Eldros observes your AI agent’s conversations and judges them automatically — hallucinations, missed resolutions, tone and policy violations are scored and flagged on real production traffic, not just in pre-launch testing.

What you send, what you get

You send (automatically)	What Eldros does	What you see
Transcripts — each user message + agent reply	Auto-judge every production conversation	Per-conversation scores + flagged issues
Session grouping — your conversation id	Assemble turns into one conversation	Conversation-level verdicts, not just per-turn
Traces — LLM/tool calls, latency, tokens	Enrich the judgment (the why behind a flag)	Drill-down on flagged conversations

Transcripts are the primary signal — they’re what the judges score. Traces are secondary: they explain why a conversation scored the way it did, and they’re captured with zero changes to your model code.

How capture works

One init() call wires up OpenTelemetry and auto-instruments your LLM library. Then wrap each exchange in a turn:

Capturing conversations

Wrap each exchange in turn(). Works in any framework, independent of your wire protocol.

LLM tracing

Every model call becomes a trace span automatically — install the matching extra and it’s on.

Sessions & turns

How conversations, turns, and traces relate — the three-layer model behind every verdict.

Quickstart

Getting started

Capture

Integrations

Concepts

What you send, what you get

How capture works

Capturing conversations

LLM tracing

Sessions & turns

​What you send, what you get

​How capture works

Capturing conversations

LLM tracing

Sessions & turns

What you send, what you get

How capture works