Tom Xiao

Work

What I have built, grouped by the stage of the story

Four cases have their own page. The shorter entries open in place. Every figure inside an interface preview is sample data.

Stage 4 · nowStage 3 · a financial platform

Stage 4Agents, and the data they run on

AI data engineer and agent engineer. Now.

AI agent · live

AI portfolio assistant

The model explains. The code counts.

Users ask about their own assets in plain language. Valuation, concentration and diagnosis tiers are computed in code, and every claim in the answer is checked.

AI agent · in production

Agents on your data

Text-to-SQL fails on definitions, not on the model.

A semantic layer so an agent finds the right table and reads values correctly. Identity comes from the gateway, and SQL runs only after the user confirms.

The task card is the only source of authority. It is capped at 12 KB and two rounds, and a hook at the end of every turn audits the diff. Each task borrows a working tree and returns it when it is done.

agent-harness-examples on GitHub

Each episode gives one new model one real task and counts the mistakes. One agent plans and accepts the work, another produces it, and a person confirms before anything is published.

The channel on YouTube

Stage 3Numbers that must be exact

Data engineer at a financial platform. 2021 to now.

Data platform

Reconciliation platform

Count the rows before you trust the totals.

External account balances against the internal ledger: one business day, counts before amounts, and real differences go to a person for approval.

Data platform · public

Data on EKS

One identity store and one policy store for every engine.

A Hadoop-era stack rebuilt as Helm charts on Kubernetes. Public repository, and covered on the AWS Big Data Blog.

Risk checks need fresh features at the moment someone deposits, withdraws or logs in. On a cache miss the queries dropped from 6 to 2 on one path and from 3 to 1 on another, and feature sync went from 30 minutes to 1 minute. When something fails, the service degrades first instead of blocking.

Internal work, so there is no public link.

Several price providers named the same security differently, so their prices could not be joined. I built a mapping between their identifiers. On a sample of 500 securities, the hit rate went from 1.0% to 99.0%.

Internal work, so there is no public link.

The AWS Quick Start CloudFormation extensions were moved to the provided.al2 Lambda runtime so they keep deploying.

aws-cfn-extensions on GitHub

Reports change quietly: a filter edited, a metric rounded differently, a page added. The extension snapshots a dashboard report and shows exactly what changed between two versions. It runs locally and only reads.

Landing page with install steps

Stages 1 and 2 were analysis and warehouse work at a consumer platform. That story is on the About page.