PlanckSpace

The challenge

PlanckSpace's local agent streams session metadata continuously from every developer machine, and that data has to be ingested, validated, and attributed in near real time, with every session, token, and dollar traced back to the exact developer, repo, team, and model that generated it. The platform has to guarantee architecturally that only metadata is ever ingested, never source code or prompts, with encrypted, auditable processing and no public exposure of pipelines or storage. Sync activity is also bursty, spiking with the working day across time zones, and every cost-saving recommendation has to be re-measured from the customer's own telemetry after a fix so savings are booked from data rather than claimed from an estimate.

What we did

• Built a custom VPC with private subnets, NAT gateways, and restrictive routing, so no processing or storage service is publicly exposed
• Authenticated agent syncs through Amazon Cognito and routed them via API Gateway to a VPC-enabled Lambda that validates and normalizes each session record into an append-only S3 lake
• Used AWS Glue ETL to attribute every session to the developer, repo, team, and model that generated it, holding attributed spend and detector state in DynamoDB
• Built a detection layer that prices context re-reads, premium-model misrouting, cache misses, marathon sessions, and idle seats as dollar-valued line items
• Scheduled EventBridge jobs to re-measure each recommendation from the customer's own telemetry after a fix, booking only the savings that actually landed
• Used Athena to reconcile attributed usage against provider invoices, and QuickSight to turn that into finance-ready dashboards with cost per shipped session
• Set up CloudWatch dashboards, logs, and alerts alongside CloudTrail auditability, backed by least-privilege IAM and end-to-end KMS encryption

Architecture and stack

Outcomes

• Metadata-only ingestion end to end, so customer source code and prompts never touch the platform, with full CloudTrail traceability
• Every session attributed to its developer, repo, team, and model within minutes of a sync
• Savings verified rather than estimated, with automated re-measurement checking every recommendation against the customer's own telemetry
• New workspaces reach their first synced session in under 5 minutes, with cost resolved down to $0.0001 per session across 6 live leak detectors