Vada Decision Intelligence: Project Story

Submission for Qwen AI Build Day · Shinhan Path

Inspiration

The biggest lie in enterprise AI right now is "just connect your data and ask questions."

Every bank, every enterprise we've worked with has the same paradox: terabytes of data, ICT teams running flat-out, and business users still waiting 2 days to 2 months for a single report. Then GenAI arrived, promising to fix this, and 95% of GenAI projects failed to create real impact, because the moment an LLM touches real business data without context, it hallucinates with confidence.

We didn't believe this was a model problem. It's an architecture problem.

The people who actually understand a bank's business are the ones running it: the relationship manager who knows which CASA segment is bleeding, the credit officer who can tell you exactly why NIF moved last quarter. We stopped trying to make the AI understand the business. We set out to build a system that makes it easy for the business to teach the AI, and to prove it by wiring Qwen in as the reasoning layer over a banking-grade decision intelligence stack.

What it does

Vada Decision Intelligence turns BI from a service into a product. A user opens Vee, types a question in natural Vietnamese or English ("Tại sao tháng 3 NIF giảm?"), and gets a verified, traceable answer in under 5 minutes. No SQL. No ticket to ICT. No waiting.

What's actually happening behind that single chat bubble:

The query is routed through a 3-Layer Contextual Intelligence Framework (Industry → Business → User) so Qwen understands not just the data, but the bank's strategy and the asker's role.
It resolves against a Single Source of Truth built from human-verified Units of Knowledge, so the model only reads numbers a human already confirmed, and never invents one.
Every response passes a Who, Does what, Where, Is allowed to know what check, so a branch manager and a CFO asking the same question get correctly scoped answers from the same engine.
When margin on a key segment crosses a threshold, the system traces the cause and recommends an action automatically, before anyone asks. We don't automate questions. We automate thinking.

How we built it

The product rests on four architectural pillars, with Qwen plugged in as the reasoning layer through MCP and A2A:

Vada Ingest. Connectors that pull from Core Banking, ERP, and ad/commerce platforms, transforming raw events into structured records with active enrichment.
Vada Spine. Our proprietary data model. Instead of dumping tables joined by foreign keys, the Spine compresses the entire business into three entities: Transactions, People, Tasks. About 50–60% of every analytical question a bank asks lives in those three. Once the Spine is in place, 70–90% of repeated requests are already answered before anyone asks them.
Vada HEC (Hyper Expeditious Calculation). A calculation engine that returns 95% of queries in under 10 seconds at petabyte scale. Benchmarked at roughly 10× faster than ClickHouse and 50× faster than MySQL on the same hardware. Short-video-shaped brains don't wait 40 seconds; the retrieval matches the speed of human attention.
Vada Auth. A real-time governance protocol that dynamically adjusts access based on the user's working status. Permissions aren't a yes/no; they're a live policy.

The AI layer sits on top:

Context Fabric. When an operator builds a report in Vada, the system decomposes it into Units of Knowledge ("Net Interest Income", "CASA ratio by branch by week"), each one a verified building block tied to a purpose. These units accumulate as a byproduct of daily work. The operator defines what matters; the system structures it.
Path Optimization. When a question comes in, Vada doesn't ask Qwen to scan blindly across thousands of tables. It first computes the optimal path (the exact set of Units of Knowledge needed), then hands Qwen a tightly scoped, pre-resolved context window.
Qwen as the reasoning engine. Because the heavy lifting (validation, calculation, governance) happens before the model is called, Qwen reasons over verified context rather than generating SQL from scratch. The AI assembles the answer from pieces a human already confirmed. No fabrication, because nothing is invented. No black box, because every figure traces back to its Unit of Knowledge.
MCP + A2A integration. This is what made the Qwen swap-in possible in days, not months. The same Vee experience runs on Qwen, OpenAI, or Gemini with no rewrite of our knowledge graph.

Challenges we ran into

Killing hallucinations without killing fluency. Constraining Qwen to "read and advise" only on the validated layer needed tight prompt scaffolding plus a hard validation gate. Too strict and answers became robotic. Too loose and figures drifted. The fix wasn't in the prompt; it was upstream, in making sure Qwen never sees raw tables, only resolved Units of Knowledge.
Mapping messy banking schemas into the Spine. Real core-banking data is rarely clean. Building Ingest enrichment so the three-entity Spine holds across legacy CBS, card systems, and digital channels took more iteration than the model work itself.
Retrieval precision. Retrieve too little and the AI hallucinates. Retrieve too much and it drowns and picks the wrong signal. Path Optimization had to be tuned to be exactly right: not too thin, not too thick.
Latency budget. A single business question can fan out into multiple HEC queries plus an LLM call. Hitting sub-5-minute end-to-end while maintaining 100% accuracy meant pre-computing aggressively at the Spine layer so the LLM call was the only "slow" step.
Bilingual nuance. Vietnamese banking questions mix English KPI names (NIF, CASA, ROA) with Vietnamese phrasing and abbreviations. Qwen handled this gracefully out of the box, which saved us building a translation pre-processor.

Accomplishments that we're proud of

Sub-5-minute report generation, replacing workflows that previously took 2 days to 2 months.
A working Validation Layer proven in production, including a deployment on 360 billion SCADA records for the National Power System operator, completed in 3 weeks.
LLM portability proven, with Qwen as the on-premise winner. The same Vee experience runs on Qwen, OpenAI, or Gemini, so we're not vendor locked. But the real prize is on-prem: most of our enterprise and public-sector pipeline (banks, utilities, government) requires the model to live inside the customer's data perimeter. Qwen is consistently the strongest open-weight family in the on-premise model competition we've benchmarked, which makes it our default recommendation whenever air-gapped or sovereign deployment is on the table.
80–90% reduction in ad-hoc reporting load on ICT teams in pilot deployments, shifting their role from Report Creator to Guardian of the Single Source of Truth.
A live commerce vertical (Vada Commerce) already serving sellers like Anchi Fashion: 8% reduction in ad costs across 200+ ad accounts, with 75% 6-month retention.
An MOU with Viettel for co-selling, and partnerships with the National Institute of Nutrition and UNICEF on a health analytics rollout.

What we learned

Context beats parameters. A mid-sized model with the right business context outperforms a frontier model fed raw tables. Qwen with Vada's Context Fabric beats GPT-class models with naïve text-to-SQL on enterprise data, every time.
Trust is an architectural property, not a prompt. Hallucination control belongs in the data layer (validation, boundaries, single source of truth), not in the system message.
The bottleneck isn't the model. It's retrieval. Once the Spine and HEC were in place, plugging Qwen in was the easy part.
The most valuable context lives in people's heads. A platform that captures it as a byproduct of daily work compounds. One that requires a separate "knowledge entry" step dies on contact with reality.
Self-serve only works with governance. Democratizing data without granular permissions creates a different disaster than the one you started with.

What's next for Vada

We're driving three product streams in parallel, each with a concrete next milestone.

1. Shinhan Bank Vietnam pilot (Shinhan Path commitment)

Our immediate next step, and the reason we joined the Shinhan Path of Qwen AI Build Day, is to take the system from demo to production inside a tier-1 Vietnamese bank.

Phase 1 (Weeks 1–2): Proof of Concept. Isolate 1–2 of the most time-consuming MIS tasks at Shinhan Bank Vietnam (e.g., daily NIF/CASA branch breakdowns, segment profitability). Deploy Vada Ingest and Vada Spine against the relevant data sources. Establish the initial Validation Layer with the bank's MIS team.
Phase 2 (Weeks 3–4): Vee + Qwen on the floor. Launch the conversational interface for a pilot user group in the Strategic Division. Finalize the ICT human-in-the-loop approval mechanism. Set security boundaries via Vada Auth, mapped to Shinhan's actual org hierarchy.
Phase 3 (Month 2 onward): Enterprise Rollout. Full rollout, SLA evaluation (95% of queries under 10 seconds), and extension of self-serve analytics to all business units. Co-publish the deployment as a reference architecture for Qwen-powered banking intelligence in Southeast Asia.

2. Vada Enterprise

The Shinhan pilot is the lead engagement, but the same playbook applies to every large enterprise sitting on legacy data systems.

Productize the on-prem Qwen stack. Package Vada Ingest + Spine + HEC + Auth with a pre-tuned Qwen deployment as a single air-gapped reference architecture for banks, utilities, and government, removing weeks of integration from every new pilot.
Convert active pipeline. Close the Q2 2026 contract with the National Power System and Market Operator (POC already complete on 360B SCADA records), and execute on the Viettel MOU for B2B2B distribution into enterprise and government data projects.
Open the Knowledge Marketplace. Let domain experts publish reusable Units of Knowledge so every new enterprise inherits a head start on industry-verified metrics, turning encoded business wisdom into a compounding asset.

3. Vada Commerce

Same engine, applied to the urgent "sell more, lose more" pain hitting Vietnam's ~530,000 e-commerce sellers.

Public Vietnam launch. Move from current pilots to a full freemium rollout targeting 1,000–5,000 SME sellers, leveraging infrastructure cost per SME of under $1/month.
Distribution through partners. Activate the partner network (e-commerce agencies, service firms, KOC affiliates) to drive low-CAC adoption, with Vee as the always-on AI analyst for sellers who can't afford one.
SEA expansion. Localize and expand into Indonesia, Thailand, and the Philippines via Gobi Partners' regional network. Two-year SOM target: 14,000 paying customers across Vietnam and initial SEA markets.

Across all three: Decision Intelligence → Decision Automation

Today Vee answers questions. Next, it executes the chain of thought continuously (daily, weekly, real-time), turning a company's best analysts into scheduled intelligence that runs before anyone wakes up.

Vada: Versatile, Validated & Valuable Data. A product from Humanise JSC. Contact: [email protected] · +84 94 6789 102