Tom Xiao

AI agents · data platforms

Seven years making numbers right: first for analysts, now for AI agents.

I build AI agents on real financial data, and the data platforms underneath them: governed access, correct numbers, checked output. Open to consulting and full-time roles.

7+ years on data platforms10 articles, one covered on the AWS Big Data Blog

portfolio-assistant · one turnsample data
You
How concentrated is my portfolio?
  1. Holding (sample)ValueWeight
    Equity fund A43,673.0034%
    Bond fund B33,397.0026%
    Single stock C23,121.0018%
    Gold fund D15,414.0012%
    Cash E12,845.0010%
    priced 5/5 · snapshot frozen · retries reuse it
  2. # Decimal, no model involved
    hhi = sum(w*w for w in weights) * 10000
      → 2400
    tier = TIERS.lookup(hhi)
      → "moderate" (1,500 – 2,500)
    defensive = (bond_B + cash_E) / priceable
      → 36%
  3. HHI 2,400tier moderatetop weight 34%defensive 36%4/4 checked
    # draft answer, claim by claim
    HHI 2,400          match
    tier "moderate"   match
    largest weight 34% match
    defensive 36%    match
    4/4 pass · 0 repairs
  4. # input to the model: computed figures only
    { "hhi": 2400, "tier": "moderate",
      "top": ["Equity fund A", 34],
      "defensive_pct": 36 }

Your portfolio is moderately concentrated (HHI 2,400). The largest holding, Equity fund A, is 34% of value, and defensive assets are 36%.

computed in code figures highlighted above come from the code, not the model
done · 4 steps

How I got here

One question kept pushing each step: is this number right, and who is allowed to see it?

  1. 2018–2020 · Stage 1

    Data Analyst

    a consumer platform

    Answered business questions and built the reports people relied on.

    LearnedMost wrong answers come from definitions and upstream data, not from the analysis.

    See the stage 1 work

    so → I moved upstream

  2. 2020–2021 · Stage 2

    Data Engineer

    a consumer platform

    Built a layered warehouse across the business domains, with scheduling generated from SQL dependencies, and worked on a cross-cloud change-data pipeline.

    LearnedScale comes from automation and conventions, not from people watching jobs.

    See the stage 2 work

    so → I went where a number must be exact

  3. 2021–now · Stage 3

    Data Engineer

    a financial platform

    Worked on the shared data platform and warehouse; merged two platforms into one and moved it to Kubernetes; built reconciliation, a live risk-feature service and security ID mapping.

    LearnedCorrectness is a property of the system: data contracts, gates and human approval.

    See the stage 3 work

    so → the users changed: agents began reading the data

  4. Now · Stage 4

    AI data engineer and agent engineer

     

    Gave agents correct, permitted data: a semantic layer, identity from the gateway, SQL that runs only after the user confirms. And built agents that ship: a portfolio assistant for real users, a public multi-agent harness.

    LearnedThe model explains. The code counts.

    See the stage 4 work

Selected work

Each case is shown the way it behaves when it runs. All figures in the interface previews are sample data.

Stage 4AI agent · live

AI portfolio assistant

Users ask about their own assets in plain language. Valuation, concentration and diagnosis tiers are computed in code; the model only explains.

Stage 4AI agent · in production

Agents on your data

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.

Stage 3Data platform

Reconciliation platform

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

Stage 3Data platform · public

Data on EKS

A Hadoop-era stack rebuilt as Helm charts on Kubernetes, secured by one identity store and one policy store. Covered on the AWS Big Data Blog.

All work, with the shorter entries

Keep going

Three more places, depending on what you came for.

Tell me the problem in two sentences

I reply within two working days.

tom.xiao@hotmail.com