Lab Alley
Nine years running commerce that standard e-commerce cannot do.
Selling regulated chemicals online means pricing, shipping and compliance rules no off-the-shelf store supports — and shipping costs nobody could account for.
Home/Case Studies/AI sales & support assistant
An AI assistant that answers from the merchant's own site — or fetches a human.
Generic chatbots either say nothing useful or invent an answer. This one answers from the merchant's own content, and knows when to stop.
AI sales & support assistantClient
14.6s → 1.3srepeat-question response time after the semantic cache
5–15s → 2.5spage content extraction, per page
7fallback stages before a site scan is declared impossible
Sector
SaaS · Retrieval-augmented AI for online retail
Technologies
Python 3.12FastAPISocket.IOPostgreSQL 16pgvectorNext.js 15TypeScriptStripeAWS
Services involved
The situation
Online stores field the same questions endlessly — is this in stock, what does it cost, what is the returns policy, which of these two is right for me. Answering them costs staff hours; not answering them costs sales. Generic assistants do not know the merchant's catalogue or policies, so they are either useless or confidently wrong.
What we found
Every store is built differently, so extracting reliable product data from an arbitrary site — missing sitemaps, pages that only render under JavaScript, servers that block automation — is most of the difficulty, and rendering every page in a headless browser is far too slow to run at signup. On the answering side, putting a full catalogue into a model prompt is slow and expensive, and a confidently wrong answer about price or stock is worse than no answer at all.
What we built
- A website ingestion pipeline that fingerprints the platform, then falls back through robots.txt, declared sitemaps, conventional sitemap paths, the platform catalogue API and finally a bounded crawl — recording each failed rung so a failed scan says what was tried
- Retrieval-augmented answering over a per-merchant vector index, with a swappable vector backend
- A three-tier answer path: a semantic cache of previous answers, then a no-model answer built directly from matched catalogue records, then a full model call
- Merchant tuning: persona and guidance rules, product attributes that become retrieval filters, and escalation rules matched by similarity
- Knowledge-base document upload indexed alongside the crawled site
- Live human handoff with agent presence, a queue, claim and resolve, and the transcript carried into the escalated conversation
- A merchant-run quality harness that replays question sets against the live assistant and scores the results
- A deploy wizard that mints an app id and generates the embed snippet, plus metered subscriptions billed per resolved conversation
Before and after
| What changed | Before | After |
|---|---|---|
| Answering a repeat question | Full retrieval plus a model call every time, around fifteen seconds | Served from the semantic answer cache in about 1.3 seconds, with no model call |
| Getting a site into the assistant | Sitemap-only discovery — a missing sitemap ended the scan and reported the site as impossible | Seven-stage discovery with platform fingerprinting, catalogue APIs and a bounded crawl fallback |
| Billing a conversation | A single conversation could be counted more than once | One row per session with an idempotent billed flag, so a conversation consumes exactly one resolution |
Where it stands now
Live. The merchant dashboard and marketing site are in production, with the API running on AWS and the front end on a managed platform.
More work
Other case studies.
Premium watch brand, US
The platform that sits between Shopify, the warehouse and the marketing stack.
Shopify runs the storefront. It does not run the parts of the business that sit either side of it.
DAR Transport
A dispatch platform, and an honest service list.
Quote, dispatch, track, deliver and invoice in one platform — starting with a signed audit of what the company actually hauls.
Free consultation
Recognise something from this project?
If your situation looks like this one, tell us about it. A senior engineer will tell you how close the match really is and what it would take.
Or email info@eye4tech.com — a senior engineer replies, usually within one business day.
Free · 30 minutes · no obligation
Free 30-minute technical consultation
- A technical opinion on your website, application or infrastructure
- A review of an existing project — what is solid, what is risky, what it would take
- A second opinion on an approach or a quote you have been given
- An honest answer on whether we are the right people, and what to look for if we are not