CLIENTS
TPBank · TPS · Sleek
INDUSTRY · BANKING & FINTECH
Production AI for the institutions that can’t afford a hallucination.
We ship AI for banks, brokerages, and fintech infrastructure — in regulated environments, on customer VPCs, with deterministic compliance. Every system below is in production today, audited, and serving end customers.
USERS REACHED
1.2M+
DEPLOYMENT
VPC · on-prem · hybrid
CERTIFICATIONS
SBV · MAS · AWS Adv
§ 01 / REALITY CHECK
Why banking AI projects actually stall.
Most vendors will tell you the model is the hard part. After shipping nine AI systems in regulated finance, we can tell you the model is rarely the hard part.
▲ WHAT WE BRING
- 01VPC-native deployment. Inference runs inside your AWS account or on-prem. Your data never leaves your perimeter. We’ve done it for TPBank’s core stack and TPS’s trading platform.
- 02Deterministic compliance gates. LLM-as-judge isn’t enough at 96% — the 4% is where you get fined. We design rule-layer + classifier hybrids that hit 99.97% on production eval.
- 03Vietnamese vernacular models. We self-host fine-tunes (Qwen-VN, Llama-VN) so retail customers get answers that read like a Vietnamese analyst, not a translation.
- 04Audit log as a first-class deliverable. Every agent action is traceable in an SBV-compliant log from day one. We don’t bolt it on at certification time.
- 05Phased rollouts as a default. 5% → 25% → 100%. We’ve never put a banking AI in front of 100% of users on day one. Neither should anyone.
▼ WHAT KILLS PROJECTS
- 01Vendor demos run on toy data. Looks great in a sandbox. Falls over on real customer balances, real KYC edge cases, real Vietnamese names with diacritics.
- 02SaaS APIs that don’t pass procurement. Your CISO will not sign off on customer financial data leaving the country. Most vendors don’t support VPC.
- 03No eval suite, no recertification. A model that was 99% accurate in March is 92% in October. Without a recertification cadence, you ship drift to customers.
- 04Generic offshore teams without finance domain. Compliance is not a deliverable you write last. It’s the architecture.
- 05“Pilot purgatory.” POCs that never get to production because the production system was never the goal. We design backwards from production from week one.
§ 02 / CAPABILITIES
Six things we do repeatedly in banking.
▸ 01
Multi-agent investing assistants
Research, risk, compliance, and execution as separate agents with hard-typed I/O. Live at TPS Securities, 200K+ users.
▸ 02
Core-banking AI overlays
Modernize the customer surface without touching the core ledger. Streaming personalization, in-app advice, fraud signals.
▸ 03
KYC & AML co-pilots
Document parsing, sanctions screening, beneficial-owner unwinding. Human-in-the-loop, with full audit trail.
▸ 04
Internal AI data analysts
Branch managers and credit officers asking natural-language questions of the warehouse. Powered by Presight, our governed analyst.
▸ 05
Compliance gates & eval suites
Hand-tuned rule layers + sentence classifiers. 99.97% pass rate on TPS production. Recertified monthly with the legal team.
▸ 06
Trading & treasury automation
FX desks, treasury workflows, post-trade reconciliation. Deterministic-first, LLM only where the rule layer can’t reach.
§ 03 / PLAYBOOK
How a 9-month engagement actually runs.
Drawn from the TPS Securities engagement. Adjusted to your stack, your regulator, and your risk appetite.
- ▸ 01 Weeks 1–2 Discovery on-site. Two weeks at your office. Shadow analysts, traders, compliance. Map the journey, sign data agreements, agree on scope. DELIV: scope doc, eval plan, data map
- ▸ 02 Months 1–2 Platform v0. Control plane, model integration (Bedrock or in-VPC self-host), first three agents bench-tested against an eval set you co-author. DELIV: platform repo, eval harness
- ▸ 03 Month 3 Compliance Gate v1. Co-designed with your legal team. Rule layer + classifier. Compliance officer sign-off on eval before we proceed. DELIV: gate spec, sign-off doc
- ▸ 04 Months 4–5 Closed beta. Internal staff + opted-in users (typically 500–1,000). Two weekly issue-triage cycles. Tight feedback loop with product. DELIV: beta report, p1 fixes
- ▸ 05 Months 6–7 Phased launch. 5% → 25% → 100% over four weeks. Sev tracking, on-call rotation, latency SLOs. We’re in your incident channel. DELIV: prod runbook, SLO doc
- ▸ 06 Months 8–9 Handoff & scale. Knowledge transfer to your team, new agents added on demand, recertification cadence locked. We stay on retainer or step away — your call. DELIV: ops handbook, retainer SOW
§ 04 / COMPLIANCE
Regulators we’ve built against.
We don’t take a project on without an in-house point of contact who has read the relevant circular. If you’re working with a regulator we don’t list, we tell you up front.
SBV · State Bank of Vietnam
SSC · State Securities Commission VN
MAS · Monetary Authority of Singapore
HKMA · Hong Kong Monetary Authority
PCI-DSS · payment data
SOC 2 Type II · controls
ISO 27001 · infosec
AWS Advanced Partner · GenAI competency
§ 05 / PROOF
Two systems, in production, right now.
TPS Securities — twelve agents, one compliance gate.
A multi-agent investment assistant embedded in the trading flow. Built in 9 months, scaled to 200K+ retail traders without a sev-1 incident.
200K+users
99.97%compliance
1.4sp95 latency
TPBank — AI-overlay on a 20-year-old core ledger.
Streaming personalization and in-app advice for 12M+ retail customers, without touching the underlying core. NDA case study; reference call available.
12M+customers
99.9%uptime
20×perf gain
§ 06 / QUESTIONS
What CTOs and Heads of Digital actually ask us.
Will my customers’ data leave our VPC?
No. For every banking client we’ve worked with, inference runs inside their AWS account or on-prem. We support both Bedrock-in-VPC (Claude, Llama) and self-hosted open-weight models (Qwen, Llama, Mistral). The data plane never crosses your perimeter.
How do you handle hallucinations in a regulated context?
We don’t rely on the model alone. Every output that touches a customer passes through a Compliance Gate — a deterministic rule layer plus a sentence-level classifier. On TPS production we hit 99.97% pass rate. The remaining 0.03% are escalated to human review, never shipped. The legal team owns the rule set; we don’t.
Can you work with our existing core / data warehouse?
Yes. We’ve integrated against Temenos, Oracle FLEXCUBE, in-house Vietnamese cores, Snowflake, BigQuery, and on-prem Oracle. Our principle: AI overlays the customer surface, never the ledger of record.
What do you do that an in-house team can’t?
Two things. First, we’ve built the eval and compliance harness nine times — you’d be building it for the first. Second, we have the Vietnamese-vernacular fine-tunes already running in production, which is a multi-month research cost we’ve already absorbed. After GA, we hand the platform over and step away or stay on retainer — your call.
How do you price?
Fixed-scope discovery (~2 weeks, fixed price). Time & materials for build, with a written budget and a quarterly checkpoint. Optional retainer for post-launch ops. We don’t do success fees on regulated AI — the incentives don’t align with the customer’s safety.
Can we talk to a reference customer?
Yes. After mutual NDA we’ll set up a call with a Head of Digital or CTO from one of our banking clients. We’ll tell you which one once we know your stack — we match the reference to the closest analog.