The Krungthai
Frontier Lab

Building the wholesale bank that thinks.

An internal capability that builds, owns, and ships the main bank’s AI and frontier, wholesale first, starting with the one thing already proven: the corporate credit report.

Banking is being rewired,
and the line is who can build.

Krungthai has a digital arm.
It doesn’t have a wholesale brain.

The wholesale bank is bleeding time.

Your own numbers, from the work we’ve done inside the bank:

The strategy was right.
The execution never happened.

In April 2025 we handed KTB the full blueprint, BASIN, a proprietary PD brain, the automated CA report, VayuLynk. A year later, NRT has shipped none of it. The only thing that exists today is the CA-report discovery we just finished, on our environment.

In three years, building the bank gets almost free.

Agentic coding, Claude Code and its successors, is collapsing the cost of building and maintaining banking software. That one shift rewrites the build-versus-buy math the entire credit stack was bought on.

The bank’s brain, wholesale first.

A permanent internal capability, not a consumer-app shop, but the engine that makes the main bank intelligent and opens the wholesale frontier.

One lab, a portfolio of products.

Everything the Lab builds is a product Krungthai owns. Here is the whole line on one screen; the slides that follow open each one.

The wholesale credit flow, rebuilt for the AGI era.

Drag the year from 2026 to 2030 and watch every box move from manual, to AI-augmented, to Lab-owned IP, Moody’s CAP shrinking to “industry scores only,” the human veto staying gold, the wholesale frontier going tokenized. Click any box for the weeds; flip build-vs-buy; trace the path that routes around KTB IT.

The corporate credit report, drafted by AI.

This is not a hope. SMC’s discovery phase, Initiative C3, already built and tested an engine that drafts the CA report across its key sections, on 14 real corporate cases (CKP, PTTGC, IVL, CPF, ThaiBev and more). Test Point 2 scored 3.52 / 5 and is rising test-point over test-point, with Gemini selected at ~฿350–1,150 per report. The engine works. The Lab’s first act is to make it a product, in the hands of relationship managers and credit analysts.

How ~15 people rebuild a bank’s brain.

Not by hand-coding systems the way IBM-DT tried. By building an agent factory over one thin data spine, and reusing it for every box. Agents do the labour; humans do judgment and evals; the shared spine makes each new box incremental, not greenfield.

BASIN-for-CA, the thin spine, not the lakehouse.

The grand all-NRT lakehouse is the thing that never got built and won’t in a year. 2027 builds only the slice the CA report needs, and nothing more.

Ingest once. Let agents consume.

Be real about sequencing: machine sources first, SET/PSIMS, filings, audited F/S, news, where AI gives a clean win; the human-touched, reconciliation-heavy sources later. And be clear: ingestion is the easy 20%; reconciliation and upkeep is the 80%, and BASIN needs a standing owner, not a project that ends.

The CA report, as a multi-agent system.

The discovery pipeline becomes a production agent graph, the one box already proven at TP2 3.52/5, shipped via INF.

• 1Planner / orchestrator, decomposes the memo into sections and routes the work.
• 2Per-section drafting agents, business, financials, risk, structure, each retrieving over BASIN-for-CA gold products (RAG).
• 3Deterministic calculation agent, every number comes from a tool, never LLM arithmetic. This is half of how you kill the trust problem.
• 4Auditor / critic agent, the “Auditor Highlights”: gap, inconsistency and report-risk detection before a human ever sees it.
• 5Question-suggestion agent + a human review UI where every committee edit is captured as an eval signal.

The PD brain is a multi-year arc, and we’ll say so.

Promising a homegrown risk brain in a year is exactly the Big-4 deck every banker is tired of. Here is the real sequence, and the weak legacy models it replaces.

BSC origination & verification, a ViaLink satellite.

The Business Service Centers do intake, verification and disbursement, paper, Line and email, expanding nationwide. This orbits the brain; it does not eat the Lab’s ~15.

Blockchain, stated specifically, one pilot, not a slogan.

Not “tokenize wholesale.” One champion supply chain on PromptBiz, in the BOT programmable-payment sandbox, doing three concrete things.

The VayuLynk seed, kept in proportion.

VayuLynk, SMC’s 2025 answer to the fragmentation, is the single system that eventually consolidates the credit process and retires Moody’s & legacy. We don’t pretend to build it in 2027. We seed it.

Automated macro & industry research.

A high-value, low-political-risk win that’s perfect for AI: the research KTB needs to underwrite well, but which only ~2 people at the bank can actually write, slowly, at the cost of an economist’s month.

One factory, shared parts.

The boxes aren’t islands. They run as a line, on one spine, with humans holding the vetoes, and the same agent library serves each station.

• →Origination, the RM (and later the deal-maker); BSC intake via ViaLink/Manuel.
• →Spine, BASIN-for-CA ingests every source and serves gold products.
• →Assessment, the CA multi-agent app drafts; Auditor Highlights flag; the committee edits and vetoes.
• →Grading, Moody’s CAP today; the proprietary PD over 2028–30.
• →Booking & monitoring, OneLynk/TOP, seeding VayuLynk; EWS / EAD-creep agents watch the book.
• →Disbursement & trade, ViaLink verification; the tokenized supply-chain pilot in the sandbox.

Every product owns a line on the balance sheet.

This bank is run line by line, each with an owner. The Lab is built the same way: every product points at a P&L or balance-sheet line, with a number a credit committee recognises.

Scale the judgment that built this bank.

The frustration is real: the teams cannot underwrite the way the best banker in the building can. So make that judgment the standard the Lab clones, not the median analyst’s.

Route around the friction, without a fight.

The biggest delivery risk is not the AI, it is the legacy IT environment: slow SDLC, no environment standards, and an understandable defensiveness. The plan is built to de-risk it.

• 1Build off-platform. Prove on the Lab sandbox, ship through Infinitas’s compliant rails (the proven retail lane), consume in TOP/OneLynk. KTB IT is touched only where regulation forces it.
• 2Read, not write. Ask only for governed read access to the specific CA data products, made a charter condition precedent, so it is a directive, not a turf negotiation.
• 3Co-opt, don’t bypass. A respected KTB-IT senior on the steering group, credited in the data standards. Consulted, not replaced.
• 4Time-box and parallel-run. New beside legacy until proven; nothing cut over until it passes. This answers the real fear of breaking BAU.
• 5A pre-agreed escalation SLA to the CEO’s office for any access or environment blocker. Its existence alone speeds the rest up.

A small room of senior builders.

Not a consulting bench. The Lab is a tight team of product-minded engineers who use AI as leverage, led by a builder who has shipped, and who recruits a Krungthai successor as it matures.

Four things we bring to the table.

Capability, not numbers. Beyond the strategy, this is what Siametrics can actually stand up.

Two pillars of a modern Krungthai.

On blockchain: Infinitas’ chain talent serves the wallet. The Lab applies tokenization to wholesale, deposits, settlement, trade, collateral. Same frontier, different layer. No overlap with the consumer virtual bank.

SoFi’s Galileo, not SoFi’s app.
Stripe’s treasury rails, not the checkout.

The Lab takes Krungthai’s national rails and SMC’s blockchain veterans into the wholesale frontier, where few Thai banks can credibly play.