About Harborview Software
Harborview Software records and analyzes in-person leasing tours for multifamily operators. Thousands of tours a month flow through its platform, and its customers coach, staff, and reward leasing teams based on what it reports — so a wrong claim in a coaching report is not a cosmetic bug, it is a personnel decision made on bad evidence.
The challenge
Operators were sitting on hours of recorded tours nobody had time to hear. Managers wanted every tour scored against their own playbook; associates wanted feedback that quoted what they actually said.
Early LLM prototypes failed in the worst possible way: they scored fluently and occasionally invented quotes that were never said. On poor audio, the transcription layer itself fabricated words. And when four people speak on a walking tour, off-the-shelf diarization merged speakers — attributing one person's words to another. In a product that judges people's work, every one of those errors is disqualifying.
The solution
SurgeX Labs built the pipeline around one rule: no claim without evidence. Speech-to-text runs with per-word confidence — low-confidence spans are bracketed, never asserted. A voiceprint layer matches transcript words to speakers and abstains when attribution is genuinely ambiguous, because a wrong speaker label is worse than none.
Scoring runs against a rubric engine versioned per customer, and every coaching bullet must anchor to a real transcript span — a citation-verification layer rejects anything the model cannot point to. Compliance-sensitive judgments on low-confidence audio are held for human review instead of auto-failing staff. Consent is enforced before any analysis begins.
Every change to prompts or models replays against evaluation sets before it ships. The system went from prototype to multi-tenant production with a two-engineer pod, released in guarded increments behind flags.
The evidence-gated coaching pipeline
How it's built
- ASR with per-word confidence; low-confidence spans bracketed, never asserted
- Voiceprint-assisted speaker attribution for multi-speaker audio
- Cite-verify layer: coaching bullets must anchor to transcript evidence
- Rubric engine versioned per customer; eval set replayed on every change
The results
* Client name changed. Engagement details anonymized under NDA; detailed numbers and reference calls available on request.
[Pull quote pending client approval — e.g. "Coaching only works if the associate believes the quote. Zero fabricated citations is the whole product."]
— [Name], [Title], Harborview Software