AI observability / from request to record

See what your AI does.
Know when it changes.

Trace prompts and responses, catch risky outputs, and keep a record your team can investigate.

Start monitoring free Explore the platform

Python & JavaScript SDKs. No credit card required.

SDK / from request to record

// npm install @drift_rail/sdk
import { DriftRail } from '@drift_rail/sdk'

const client = new DriftRail({
  apiKey: process.env.DRIFTRAIL_API_KEY,
  appId: 'support-assistant'
})

await client.ingest({
  model: 'gpt-6.1-sol',
  provider: 'openai',
  input: { prompt: userMessage },
  output: { text: assistantReply }
})

// A trace your team can investigate.

Follow a request through the platform

One sample, four views. Explore detections, event traces, drift alerts, and the numbers behind them.

Sample workspace / production

The health of your AI, at a glance.

Reporting period

Events captured: 124.5K; +12.0% vs previous period
Health score: 87/ 100; Across monitored applications
SLA compliance: 99.2%; Requests within target latency
Estimated cost: $847; Model usage this period

Event volume

Requests captured across your applications

Previous periodCurrent period

Risk over time

Mean risk score · lower is better

15 points lower

Risk score / 100

Latency p95: 142ms
Error rate: 0.12%
Current risk: 23/ 100
Open incidents: 2 to review

Illustrative data. Change the reporting period to compare the sample.

Every request, accounted for

Your inference stream, in focus.

1,247 inference events captured

Average latency: 142ms; Across captured events
Total tokens: 48.2K; Input + output
Production events: 847; Production environment
Models observed: 5; Across three providers

Sample inference events. Select an event ID to inspect its trace.
Event / inspect	Model	Environment	Latency	Tokens in / out	Decision

Illustrative data. Filter the stream or select an event to inspect its trace.

Try the full platform free

No credit card required

Works with any LLM provider

OpenAI
Anthropic
Google Gemini
AWS Bedrock
Meta Llama
Mistral AI
LangChain
LlamaIndex
Vercel AI SDK
Cohere

From a trace to an answer

A clearer view of production AI.

Connect the request, the risk, and the evidence. Give your team a place to start when something changes.

An illustrative interactionevt_90c support-assistant

PII detected

model: gpt-6.1-sol
input.prompt: "Find the customer's email."
output.text: "Email alex@example.com."
latency: 142 ms

01 / Capture

Start with what happened.

The prompt, response, and model stay together. Follow the interaction with its timing and context intact.

pii.detected: true
risk_score: 87 / 100
guard.action: block

02 / Detect

Make the risk visible.

A contact detail becomes a signal. See what was flagged and the guard decision attached to this interaction.

trace_id: tr_19a
event_id: evt_90c
export: JSON / CSV / PDF

03 / Investigate

Keep a path back to the evidence.

Open the trace, review the event, and export the record. Your team can work from the same evidence.

One interaction, shown from capture through review. Sample content and values.

Built around your application

A small addition.
A much clearer picture.

Send interactions from your existing application with the Python or JavaScript SDK. Keep your model provider and use scoped API keys for each environment.

Read the integration guide

app.ts / npm install @drift_rail/sdk

import { DriftRail } from '@drift_rail/sdk'

const client = new DriftRail({
  apiKey: process.env.DRIFTRAIL_API_KEY,
  appId: 'support-assistant'
})

// Record the interaction after your model responds.
await client.ingest({
  model: modelName,
  provider: providerName,
  input: { prompt: userMessage },
  output: { text: aiResponse }
})

One record. Different questions.

For the people responsible for AI.

01 Engineering

What caused
this response?

Start at the interaction. Inspect the prompt, model, and output together, then follow the trace back to its context.

The request is your starting point.

02 Risk & safety

What needs
a closer look?

Find the signal in the output. Review sensitive data, policy flags, and changes in behavior. Add detections for the risks your team needs to track.

The detection gives you a direction.

03 Governance

Can we show
the evidence?

Bring the record into your review process. Export interactions and reports for incident investigations and compliance work across healthcare, finance, and legal teams.

The record supports the review.

Supporting SOC2, HIPAA, GDPR, or EU AI Act reviews?

Explore compliance reporting

Keep your team in the loop

A signal where
you already work.

Send risk, drift, and volume alerts to your team's channels, with the details needed to open an investigation.

Slack
Microsoft Teams
Discord

Set up team notifications

#ai-operations

Example notification

A change worth investigating.

The risk distribution for your support assistant has shifted from its baseline.

Event: Drift detected
Application: support-assistant
Signal: Risk distribution change

Explore a drift investigation

Start with one application

Make your next AI interaction
one you can understand.

Connect an SDK, follow a trace, and give your team a clearer view of what happens in production.

Start monitoring free Talk to us

No credit card required.

See what your AI does.
Know when it changes.

Follow a request through the platform