All work

Osaka Group

Applied AI & Automation

I built Travel Avenue with a five-person team, the storefront, app and AI assistant that moved a forty-year-old retailer to selling direct.

Dates
Jun 2025 — now · 1 yr 4 mos
Role
AI Software Engineer
Site
travelavenue.ae
Project
Travel Avenue

01

The assistant

Shipped the bilingual shopping assistant to answer from the store's own catalogue and policies, handling the product questions that would otherwise reach support and cutting first response from hours to under 10 seconds.

Guardrails keep it inside its remit. It answers questions and refuses anything that touches money or an order, it holds up against prompt injection, and personal data never reaches the logs. A golden set catches accuracy and scope regressions before a customer meets one.

Every conversation is costed, and that cost decides which model answers it. Routine questions go to a small model and the hard ones escalate to the primary. The saving comes from the price gap between tiers; the difficulty is calibration, because the threshold that routes correctly this month drifts as the traffic changes.

02

Personalisation

Every visitor gets a page ranked for them. A weighted score combines what they have bought before, what they have looked at in this session, and what is actually in stock, so nothing surfaces that cannot be sold. Category quotas stop a single collection taking over the page, and a popularity fallback covers anyone the system has not seen before.

03

Automation

Designed the automation platform on n8n that reconciles inventory, rolls seasonal campaigns and pricing across web and app, and publishes articles on schedule, replacing over 20 hours a week of manual work, with a human in the loop on every write.

Created the generative media capability behind the campaign artwork, so seasonal creative is produced on schedule instead of commissioned per campaign.

Shipped the Instagram DM agent. It reads incoming messages, holds the thread's context, and answers directly where it can, so an enquiry gets a reply without anyone watching the inbox.

04

Platform and data

Replaced a per-seat SaaS CRM with a self-hosted deployment on company-owned AWS, cutting licensing cost by around 85% a year and shaping the data model around how the business actually works: B2B leads, customer queries and orders.

Gave AI agents read-only access to the company's Google Merchant Center, Search Console and Drive through a tool server, so an agent can investigate a listing or SEO problem on its own.

05

How it ships

Worked through each iteration in Agile cycles with the e-commerce, sales and warehouse teams who use the tools daily, so most of what changed came from them rather than from a backlog, shipping every codebase through a reviewed pull-request pipeline with CI/CD.