PLUS2026 Singapore Symposium

1 Oct 2026 · 11:05 am

Prompt Fight

AI Brosvs.Insurance Pros

One D&O underwriting worksheet, drafted by an AI agent running on hardware I own. A human still signs it.

Jensen Loke

CEO, Success IT

Today's case study

Sheng Siong
Group

SGX-listed · Supermarket retail

CoverDirectors & Officers

LimitS$10m

Claims historyNone

Current coverNew policy

SourcesPublic filings only

01Your speaker

02

JL

Jensen Loke · Singapore

CEO, product builder and community organiser

I design products, workflows and teams that make complex services work better.

Across consumer and corporate digital products, the goal has stayed the same: use technology to solve real business problems and improve outcomes.

Consumer productsDBS digital banking

Corporate platformsTemasek · Ant Financial · Amazon

NowCEO, Success IT

Community1,000+ Agentic Builders

02Where the agent runs

03

From desk to model

The agent runs
on my own hardware.

Jensen's home local AI hardware setup — **The actual home lab**Compact compute, a Mac mini, networking, and honest cable management.

The simple architecture

MacAgent app
and harness

TailscalePrivate encrypted access

LiteLLMStable API
dgx-current

DGX SparkQwen 3.8 27B
local inference

Private by defaultThe model backend stays on loopback; remote access stays inside the tailnet.

One stable endpointThe agent does not need to know which inference engine is underneath.

No per-token billResearch runs cost electricity and time, not API spend.

Swap the modelAny OpenAI-compatible model, local or cloud, plugs into the same agent.

03How the agent works

04

ANDREW · A Narrative D&O Risk Evaluation Worksheet

Research in parallel.
Price in code.

Code

Gather

Pick the listed company. Market data and candidate annual-report PDFs are fetched once and shared.

→

Model · parallel

Research teams

Market & securitiesFinancialsGovernance & auditShareholdingAdverse news & regulatory

→

Model

Synthesise

A lead underwriter merges the findings into one proposal. Every answer carries a value, a rationale and a source.

→

Human

Review

The underwriter re-sets modifiers and toggles exclusions before anything is priced.

→

Code

Price & refer

The worksheet's rating tables and about 38 referral triggers. Premium, CLEAR or REFER, and the formula shown.

Numbers in codeThe model never does pricing math. The engine reproduces the worksheet's worked example to the cent.

Words in the modelResearch and narrative come with citations. "None identified" is a finding; a blank is not.

Judgement with the humanThe agent refers when it's unsure instead of bluffing a number. An underwriter signs off.

04Live demo

05

Live · Sheng Siong Group

Let's underwrite it.

The brief

InsuredSheng Siong Group Ltd · SGX-listed

CoverDirectors & Officers liability

LimitS$10m

ClaimsNone · new policy, no current cover

InputsAnnual reports, SGX filings, public news

What to watch

1

The research lanesEach team reading filings, with sources as it goes.

2

The cited proposalRisk narrative, modifiers and exclusions, each with its reason.

3

The human reviewWhere I push back on the agent's calls.

4

The price and verdictPremium, CLEAR or REFER, and the formula behind it.

05Scan to connect

06

Thank you

Take the agent home.

QR code for Jensen Loke on X, x.com/jensenloke

X · Twitter

@jensenloke x.com/jensenloke

Jensen Loke linkedin.com/in/jensenloke

QR code for the ANDREW open-source repository on GitHub

Open source · GitHub

ANDREW github.com/jensenloke/andrew-do-worksheet