INTERNSHIP / B2B SAAS / ENERGY EFFICIENCY

EcoRebates: Agentic AI Review System. 25,000+ live rebate programs, each re-checked by hand every 6 to 8 weeks.

~767Reviewer-hours back every cycle

EcoRebates / Agentic AI Review System

Role
TPM Intern, Agentic AI and Automation
Timeline
Jun to Sep 2026, remote. EcoRebates (Castanea Labs, Inc.)
Stack and methods
Claude Managed Agents, Claude API, Claude Code, n8n, Firecrawl, AWS, Google Sheets, Jira
  • 25,000+ live programs triaged
  • ~767 reviewer-hours saved per cycle
  • 14x ROI
  • 2 weeks early on a 12-week roadmap

Review console mock

Interactive / try it

The problem

EcoRebates maintains a catalog of 200,000+ US energy rebate programs. An offshore team manually verifies the 25,000+ active ones on a 6 to 8 week cycle. Every rule change at a utility can silently make the database wrong.

What I built

A human-in-the-loop first-pass triage system. Scraping, extraction, and comparison agents check each live program against its source and route findings to reviewers through a console that cites a source for every finding. The offshore team remains the gate: nothing reaches customers unverified.

Additional results

  • Database gap detector validated at 92% recall with zero false positives.
  • Prompt caching cut input token cost 70%.
  • A dispatcher runs 8 concurrent agent sessions.
  • Met 6 of 8 success measures.
  • Runs at roughly 1 to 2% of manual labor cost.
  • Won pre-launch adoption from the Director of Data and Programs, plus a request to expand to new categories.

Decisions worth asking about

Decision 01

Field-level routing, not confidence scores.

Rejected: Auto-approving anything above a confidence threshold.

A quarter of confidence-scored approvals hid silent mismatches. Routing by field cut auto-approval from 48% to 26%, trading throughput for trust.

Decision 02

Hybrid deterministic and agentic.

Rejected: Agents end to end.

Deterministic steps where rules suffice, agents only where judgment is needed, at roughly 1 to 2% of manual labor cost.

Decision 03

The human stays the gate.

Rejected: Fully automated publishing.

Every finding cites its source and reviewers make the final call, which is what won adoption before launch.

Decision 04

Cache before you scale.

Prompt caching behind a dispatcher cut input token cost 70% while running 8 sessions concurrently.

← All workRun Fit Check