LG Electronics Production AI
Real-time performance monitoring
A production platform that watches LG.com's commerce performance around the clock, investigates important changes automatically, and turns them into evidence-backed cards that business teams can act on.
Product owner and lead engineer. Took the platform from pitch to production, covering architecture, data pipelines, forecasting and detection, the investigation engine, alerting, the web app, releases and operations.
In production · hourly and daily detection
- forecast-based detection
- Hourly
- typical investigation
- ~4 min
The problem
Pricing and coupon errors on lg.com/us were surfacing after customers had already made purchases. More broadly, a sudden change in sales, conversion, traffic or add-to-cart rate could go unnoticed, and explaining it meant manual analysis across several data sources. The business needed an earlier signal and a faster, trustworthy explanation.
What I built
Pitched the platform to LG North America's CEO and business-unit vice presidents, then designed, built and now run it end to end: data pipelines, forecasting, detection, the investigation engine, alerting, and the web app business teams use.
Chronos-2, a pretrained time-series foundation model running on its own GPU service, forecasts an expected range for each KPI across the whole site, each business unit and each product category. Hourly and daily runs flag values outside that range, then statistical and materiality checks tuned per KPI filter out noise before anything reaches people.
Important anomalies on the four headline KPIs (revenue, conversion rate, shopping sessions and add-to-cart rate) trigger an automatic investigation. A playbook runs data checks in parallel, a language model may add a few checks from a fixed menu, and the report it writes must cite evidence for every figure. A second model reviews the report, and code checks the numbers before anything is published.
A prioritisation layer decides what pages people now and what is only recorded. Alerts go to Slack and email once per incident, an Actions board tracks who owns each issue, and a chat assistant with read-only tools answers follow-up questions. The app is available in English and Korean.
Alongside the forecasts, five-minute rule-based monitors flag heavily discounted carts before checkout and track payment failures.
How I evaluated it
Every number shown in the product is recomputed with independently written queries before a release, and the counting logic is tested against randomised reads. A replay harness reruns the full pipeline on real past anomalies as a dry run, with every write blocked, so old and new reports can be compared side by side. Detection rules were tuned the same way: a proposed persistence rule was rejected after replay showed it would have hidden real coupon incidents. The test that decides whether a promotion caused a move was calibrated on placebo data, where it blamed a promotion by chance in under 1% of cases.
Built to operate
Scheduled runs are serialised and recover on their own: missed days are caught up, late hours are reconciled, and failed investigations are retried. Each release lands first with no traffic and passes automated checks, including a live call to every configured model, before promotion, with a recorded rollback target. Alerts go through a durable outbox, and each person is alerted once per incident rather than once per KPI.
The outcome
The platform runs in production for LG.com business and analytics teams. Automated investigations typically finish in about four minutes and produce a card that explains what happened, why it matters, the probable cause, and what to check, in place of hours of manual analyst investigation.