How Fello observes 3M+ agent traces per day without sampling

Zero

Sampling at 3× lower cost

50%

Time saved on eval review
(~400 hrs/mo)

85%

Lower inference cost on evals

Unified

Agent, infra, and app observability

Fello's AI teammate, Felix, engages real estate customers across voice, SMS, and email, then hands qualified opportunities to human agents.

IndustryB2B SaaS: Real Estate Technology

Company size150–200

Migrated fromLangfuse

Also from Fello

Unified observability, 100% adoption, half the cost

Challenge

When a customer reported something, we couldn't find the trace if it had been sampled out.

Ashish Gupta, Principal Engineer

Running LLM evals on our production traces was costing us a lot.

Aman Gupta, Product Manager

Fello's AI teammate, Felix, makes roughly 100K calls and sends 150K SMS messages and close to a million emails every day. To control storage costs, the team sampled traces, which left gaps in both debugging and evals.

Pain points:

Trace storage costs rose ~20x: So the team started sampling.

Missing traces: When a customer reported a bad call, its trace was sometimes missing due to sampling.

Blind spots in evals: Failures in sampled-out conversations went undetected.

Langfuse search was slow at scale: The team had to file multiple support tickets to get it resolved.

Traces in one tool, metrics in another: Langfuse handled traces but not Fello's custom metrics.

Solution

No sampling at one-third the cost

The cost savings are a no-brainer. Based on my research, nobody else offers this quality at this cost.

Suman Varanasi, Co-Founder & CTO

With Oodle, we can chain evals by dependency. Code-based evals filter traces before we run LLM-based evals.

Aman Gupta, Product Manager

Fello evaluated a few tools. None could store every trace at Fello's volume for a cost that made sense, and none supported hierarchical evals.

What made the difference:

No sampling, lower cost: Oodle costs about a third of Langfuse, so every production trace gets stored.

Hierarchical evaluations: Code and ML-based checks filter traces before LLM-based evals run.

Team responsiveness: Requested features were shipped in days, sometimes ahead of schedule.

Agent to infra in one place: Follow an agent trace from the LLM call to tool calls to the APIs and databases they hit.

Built-in prompt management: Prompt versioning, playground testing, and dataset experiments in one place.

MCP access: Build dashboards and debug traces in plain English.

Impact

We're catching issues before our customers do. Oodle gives full visibility with no sampling across millions of traces.

Ashish Gupta, Principal Engineer

Chaining evals reduced our inference costs by 80–85%.

Aman Gupta, Product Manager

Fello keeps every production trace, so a customer-reported conversation leads to its trace in seconds. Developers, product managers, QA engineers, prompt engineers, and conversation designers now work in Oodle daily.

Complete production context: At one-third the cost, Fello can retain every trace to debug and improve their agent.

80–85% lower eval inference cost: Hierarchical evals reduce the number of expensive LLM-judge calls.

2x faster evaluation review: Review time fell from roughly one hour to ~30 minutes.

Proactive issue detection: Oodle surfaces silent agent failures before customers report them.

Keep every agent trace, with no sampling, at a fraction of the cost.

Request a Demo →