Fallom
Fallom is an observability service for developers running AI applications. Its getting-started documentation describes tracing model calls so you can inspect production behavior, debug issues, and understand costs. The product combines call-level records with session analytics, A/B testing, and evaluations.
Inspect individual model calls
Fallom traces tokens, latency, costs, prompts, and completions for each LLM call. That puts the request and response beside the usage and timing information for the same interaction. A developer investigating an application can inspect what the model received and returned, along with the associated token usage and cost.
The introduction shows a TypeScript setup using the @fallom/trace package. You initialize Fallom with an API key, create a session with a configuration key and session ID, then wrap an OpenAI client. Calls through that wrapped client are automatically traced to the dashboard. The documentation also lists a Python package named fallom.
Follow sessions and compare configurations
Session analytics groups calls by user or conversation. This gives a view of related calls beyond an individual request, which matters when a conversation contains several model interactions. The example uses a session ID to identify that group.
Fallom supports A/B testing with production traffic splitting. The introduction describes testing models and prompts, so the experiment can concern either the model choice or the instructions sent to it. Fallom also provides evaluations for LLM outputs at scale.
Connect supported providers
The publisher describes support for OpenAI-compatible APIs and lists OpenAI, Anthropic, Google, and OpenRouter. It also lists Vercel AI SDK support. These are the integration options named in the introduction; the worked example specifically demonstrates wrapping an OpenAI client.
For initial setup, the documentation provides the package names and a short code example. Its focus is getting call traces into the dashboard and organizing them into sessions.