Home / Work / OCBC AI prototypes

Built solo

Three AI prototypes for OCBC

I built these in a week on OCBC's public product data, pulled through its API, to reimagine how AI could work in a regulated bank: grounded, explainable, and never acting on its own.

  • Status Prototypes
  • Built October 2026, solo
  • Stack Python, Streamlit, scikit-learn, Claude

Unofficial prototypes, not affiliated with or endorsed by OCBC. Product facts come from OCBC's public product data, accessed through its API. All customer data is synthetic.

1. Product Q&A for staff

How an answer is made
A staff questionRetrieve product sectionsAnswer with the source citedOr refuse, if the data doesn't cover it
GroundedAnswers only from retrieved sections of product documents.
Honest about gapsRefuses when the data doesn't cover the question, and warns when it is stale.
MeasuredAn evaluation tab compares keyword search with a hybrid approach and lists the wrong answers.
An answer citing the exact product section and source file, with the date the data was pulled
Answered from product data, with the section and source cited and a freshness warning.
A refusal: the annual fee for the 365 Credit Card isn't in the product data
Asked about a card the data doesn't cover, it says so instead of guessing.

2. Campaign targeting by uplift

From customers to compliant copy
CustomersUplift modelRanked list with reasonsDraft offer copyCompliance checks
Uplift, not propensityRanks customers by the extra chance of converting that the offer itself causes.
Reasons shownWhy each customer was chosen, in plain words.
Checked copyClaude drafts from cited product facts, and plain-code compliance checks run on every draft.
Incremental conversions from targeting by propensity, uplift and random, measured on a holdout
On synthetic data, targeting by uplift won about four times the extra conversions of targeting by propensity, measured on a holdout group.
Offer copy for the largest uplift segment, with compliance checks passed and every fact cited
Offer copy for one segment: every number matched to a cited product fact, and every check passed.

3. A copilot for relationship managers

The manager always decides
A manager's clientsRank by who needs attentionSummarise in codeDraft talking pointsFact-check every pointManager reviews and decides
Who to callRanks clients by who needs attention, with the reasons.
Cited pointsEvery talking point cites a source fact with its date.
Never acts aloneIt never contacts a client. The manager reviews, edits and decides.
A relationship manager's 100 clients ranked by likelihood a conversation leads to action, with the reasons
Who to call, with the reason for each ranking in plain words.
Talking points drafted by Claude, each with its source facts and dates, passing a fact-check
Draft talking points: each one cites its sources, and the fact-check passed 11 of 11.

Rules I held across all three

Show whyEvery score or ranking explains itself.
Simple firstExplainable methods before complex models.
Fail safelyEvery AI call has a timeout and a non-AI fallback.
Flag assumptionsAssumptions about banking data and regulation are marked where they are made.

A walkthrough is available on request. The thinking behind how each one splits the work between machine learning, a language model and plain code is in ML decides, the LLM writes, code checks.