Evaluation, safety & observability

Trace, monitor and debug LLM apps in production

Langfuse/LangSmith/OpenTelemetry traces, token & cost dashboards, drift and failure alerts.

~8 focused hours·intermediate

Tools: Langfuse, LangSmith, OpenTelemetry, Grafana, structured logging

What employers mean

You should be able to…

  1. Instrument an LLM app with traces/spans so every prompt, tool call, and model response is captured with timing
  2. Track token usage and cost per request, per user, and per feature, rolled up into a dashboard
  3. Set up alerts for latency spikes, error-rate increases, or cost anomalies in production
  4. Correlate a user-reported bad response back to its exact trace (prompt, retrieved context, tool calls) for debugging
  5. Detect drift: a model or prompt update silently changing output quality or latency distribution
  6. Self-host or configure a tracing backend (Langfuse) vs use a managed one (LangSmith) based on data-residency needs
  7. Export traces via OpenTelemetry so LLM observability integrates with existing infra monitoring

Needs first: Deploy an AI service to the cloud

Learn — free, link-checked

The few resources that matter

Read · beginner · 25 min · langfuse.com

Overview

Open-source LLM observability platform overview, the most-referenced self-hostable tracing tool in Indian GenAI JDs. — Langfuse
Read · intermediate · 20 min · langfuse.com

Token & Cost Tracking

Shows exactly how per-call token/cost is captured and rolled up into dashboards, the basis for any cost-optimization work. — Langfuse
Read · intermediate · 30 min · docs.smith.langchain.com

LangSmith Observability

Alternative managed tracing stack tightly integrated with LangChain/LangGraph, with cost/latency dashboards out of the box. — LangChain
Read · intermediate · 40 min · langfuse.com

LLM Observability & Application Tracing

Hands-on guide to instrumenting an LLM app with traces/spans/generations for debugging multi-step agent runs. — Langfuse
Read · intermediate · 45 min · opentelemetry.io

Python | OpenTelemetry

Vendor-neutral tracing/metrics SDK underlying most LLM observability tools, worth knowing directly for custom dashboards. — OpenTelemetry
Course · intermediate · 90 min · deeplearning.ai

Evaluating and Debugging Generative AI

Shows how to version prompts/configs and trace experiments so eval regressions are debuggable, not just scored. — DeepLearning.AI (with Weights & Biases)
Practice

Observability dashboard for a multi-tenant RAG support bot

Take an existing (or simple mock) RAG support bot serving multiple clients, instrument every request with Langfuse tracing (prompt, retrieved chunks, model call, tool calls), and build a dashboard showing token cost and p95 latency per tenant per day. Add an alert rule that fires when p95 latency or error rate crosses a threshold, and demonstrate tracing a deliberately bad response back to its root cause.

Done when
  • Every request produces a trace with nested spans for retrieval, prompt construction, and model call
  • A dashboard (Langfuse UI or exported to Grafana) showing token cost and p95 latency broken down by tenant
  • An alert rule configured and demonstrated firing on a simulated latency spike or error burst
  • A written walkthrough of tracing one bad response from user report to root cause using the trace
Prove it

Evidence a recruiter can check

  • Public GitHub repo with the instrumentation code and a screenshot of the live dashboard
  • A short write-up/video of the root-cause debugging walkthrough using a real trace
  • Screenshot or export of the alert firing on a simulated incident
Interview

Questions you'll get asked

  1. How would you debug a single bad response from a production RAG chatbot down to its root cause?
  2. What would you put on a cost/latency dashboard for an LLM feature, and who looks at it?
  3. How do you detect that a vendor's silent model update degraded your app's quality?
  4. Explain the difference between a trace, a span, and a generation in LLM observability tooling.
  5. How would you set up alerting for a spike in token cost without false-paging on normal traffic growth?
  6. Why would you choose Langfuse (self-hosted) over LangSmith (managed), or vice versa, for an Indian fintech client?
  7. How does OpenTelemetry fit into an LLM observability stack that already uses Langfuse or LangSmith?
See where you stand for AI Engineer