BackEcho Roberts

Customer Reference Agent

AI-powered search that surfaces customer quotes from Salesforce in under 3 seconds.

Role

Full-Stack Developer

Timeline

2025

Stack

Next.js · TypeScript · Vercel AI SDK · Claude/GPT-4o/Gemini

Problem

Sales reps frequently interrupted account managers mid-pitch to ask for customer quotes. Although Salesforce contained 200+ approved quotes, keyword search failed on synonyms and shorthand, returning either dozens of irrelevant results or none at all.

Constraints

  • Semantic understanding needed—synonyms and sales shorthand ('SF', 'SFMC', 'SMB') had to resolve correctly
  • Zero downtime during live sales calls
  • Highlights had to reference exact source text, not AI-generated summaries

What I Built

An internal search tool that surfaces customer quotes from Salesforce with confidence scores and exact-match highlights.

Key Decisions

Two-pass ranking (relevance → highlighting)

Separate initial ranking from highlight extraction—single-shot conflated keyword matches with actual relevance

Multi-provider failover

Claude → GPT-4o → Gemini → keyword matching. API outages during live sales calls would be unrecoverable

Conservative filter inference for segmentation

Only apply industry/segment filters when explicitly mentioned—wrong filters hide results invisibly

Exact substring extraction

AI tends to paraphrase when extracting highlights ('reduced costs' instead of 'reduced our costs'), breaking the highlight match. The prompt explicitly instructs 'extract the exact text as it appears, character-for-character.' Validation step: if the returned highlight doesn't exist in the source quote, drop it. Incorrect highlights were dropped rather than risk misleading output.

Outcomes

  • Searches dropped from 5+ minutes (manual Salesforce scanning) to under 3 seconds
  • Reps stopped calling account managers mid-pitch (5+ weekly searches)
  • Zero full outages in production; keyword fallback activated twice
  • 24-entry abbreviation map resolves sales shorthand without model tuning