Skip to content
Sophie Yin
← All work

auto-pricing triage

2025 · Shipped · Built at Alt · Patent pending · Python

case study

Problem

A pricing model proposes a price for every item in a quote, but not every price can be trusted, and expert time is the scarce resource. Without a reliable way to tell which prices hold up, experts spent time on items the model had already priced well.

What it does

Every item passes through a sieve of decision-makers. Cheap, safe deterministic rules go first; for the items they can't settle, LLM agents gather evidence that is weighed in tiers. Each item leaves with exactly one route: accept the model's price, send it to an agent pricer whose price comes back through triage for a second pass, or send it to an expert. More items get priced automatically, and experts see the ones that need them.

System

How an item gets priced — and what runs underneath. Click Triage to look inside.

LLM step

What decides the price

Pricing triage flowAn underwriting trigger sends an item to the ML pricer. Triage checks the price with deterministic rules, then evidence, and routes it: accept to a valuation, send to a human expert, or send to the agent pricer, whose price returns to triage for a second pass.ACCEPTHUMANAGENTPRICE RETURNS TO TRIAGE · PASS 2Underwriting triggerML pricerTRIAGEDeterministic rulesEvidence basedHuman expertAgent pricerValuationPRICED BY ML OR AGENTExpert valuationMANUAL REVISIONPricing triage flowAn underwriting trigger sends an item to the ML pricer. Triage checks the price with deterministic rules, then evidence, and routes it: accept to a valuation, send to a human expert, or send to the agent pricer, whose price returns to triage for a second pass.ACCEPTHUMANAGENTBACK TO TRIAGE · PASS 2Underwriting triggerML pricerTRIAGEDeterministic rulesEvidence basedHuman expertAgent pricerValuationPRICED BY ML OR AGENTExpert valuationMANUAL REVISION
Accept · 01/05

An underwriting event arrives and the ML pricer proposes a price.

Architecture

  1. 01The ML pricer proposes a price for each item
  2. 02Deterministic rules: reuse a recent expert price, hard gates, fast auto-price, impact on the quote
  3. 03Evidence: LLM agent enrichment, then signals weighed in tiers
  4. 04One route per item: accept, agent pricer or expert
  5. 05An agent's price returns through triage for a second pass

Stack

  • RulesDeterministic decision nodes in Python
  • EvidenceLLM enrichment agents + tiered signals
  • RuntimeSQS + Lambda, execution workers
  • ServingML pricing API
  • TracingLangfuse, one trace per quote

Outcomes

  • Auto-priced rate up more than 20%
  • Manual pricing workload cut
  • Patent pending — named inventor

Code is private — happy to walk through it.