Work · IDeaS Revenue Solutions (a SAS company) · 2025–26

Enterprise automation at IDeaS

IDeaS builds revenue-management software for hotels. Around the product sit teams doing a lot of careful, manual, cross-system work. My job has been to turn that work into scheduled, tested automation, and to put LLMs where they genuinely help.

Roles: AI & Automation Developer (Nov 2025 – present); Business Intelligence & Automation Intern (Jul – Oct 2025); Software Developer Intern (Jan – Jun 2025)
Stack: Python, pandas, Playwright, Power Automate, Streamlit, Altair, Java, Spring Boot, Angular, SQL, Docker
Systems: G3 revenue-management platform, Salesforce, Excel/Word/PDF, Microsoft AI Builder
Note: Described at the level of what was built. No internal code, data, names or screenshots.

Forecast-review reports with LLM-written commentary

01
Data
Forecasts and projections per property
02
Calculations
Forecast vs. projection analysis
03
Charts & tables
Rendered per section
04
Commentary
LLM-written narrative (Microsoft AI Builder)
05
Merge
PDF, Word and Excel deliverables

A Python pipeline that assembles client forecast reviews. It is split into modular services (data, calculations, charts, document merge) so a new report section is an addition, not a rewrite, and it handles both full-history properties and those with limited history, where the model runs on a synthetic baseline. It is containerised with Docker and went through SIT, UAT and hypercare.

The LLM writes narrative around numbers the pipeline has already computed. The failure mode worth worrying about is a wrong number in that narrative, which is exactly what led me to build Proofline.

Configuration checks across three systems, in a browser

01
Salesforce
Client account numbers, integration types
02
Product UI
The same settings in the G3 platform
03
Database
What is actually stored
04
Check
Playwright run in Dockerised test environments

Setting up a client correctly means the same facts have to agree in CRM, in the product, and underneath it. I automated that end-to-end check with Playwright, running in Dockerised test environments, so it runs the same way every time instead of depending on someone clicking through three systems.

Analytics and mailbox automation

Pricing analytics in Streamlit

A Streamlit app (pandas, Altair) that takes pricing exports and automates best-available-rate trend analysis, property and chain extraction, date filtering and pricing-decision validation. It is structured as configuration, validation, calculation and chart services, so analyses can be added without touching the UI.

A shared customer-care mailbox that tracks itself

Scheduled Power Automate flows parse incoming client and system emails (chain codes, timestamps) into Excel and reconcile replies from Sent Items, so the team sees what is waiting and what has been answered without maintaining a tracker by hand. It removed ~3 hours of manual tracking.

Modernising a legacy dashboard without a big-bang switch

As a Software Developer Intern I migrated the product's “At-a-Glance” control dashboard from JSP and Struts to Angular and Spring Boot: Angular components on top of REST APIs backed by optimised SQL and Java Stream processing. The APIs moved from single-day data to custom date ranges, which meant reworking and tuning the underlying queries.

The new UI shipped behind a database-driven feature toggle, added through a SQL migration, so clients could be moved from the legacy JSP views to Angular in a controlled rollout, and moved back if something went wrong.

PATTERN

What these have in common

Start from the workflow: who does it, how often, what breaks, what “done” looks like.
Build the smallest reliable automation first; add an LLM only where language is the actual work.
Test it where it will run (containers, SIT, UAT) and roll it out so it can be rolled back.

Next case studyProoflineVerifying the numbers in LLM-written reports.