Infozense Insights

One Engine. Any Infrastructure. Sub-Second Answers.

2026-07-27T11:07:17.242Z

StarRocks: open-source analytics engine that queries data where it sits. Zero copy. Sub-second. Deploy anywhere.

Secure Data Sharing — Infozense

2026-07-27T11:07:17.242Z

How we built a platform that lets 500+ analysts query 100TB+ of sensitive data — without ever copying it. Sharing data without giving data.

Thailand's Model Risk Inflection — Infozense Research Note

2026-07-27T11:07:17.242Z

Three regulatory pressures are converging on Thai banks at once: BOT model-governance direction, SET disclosure rules, and EU AI Act spillover. What 2026 asks of the model-risk function — and what to build now.

Model Risk Atlas — Banking, Vol. 1 — Infozense Research Note

2026-07-27T11:07:17.242Z

How model risk maps across Retail, Commercial, Markets, and Digital divisions of a Thai bank — and where Intelligence Engineering puts each model class. The companion atlas to Thailand's Model Risk Inflection.

What 55 Million Order Book Updates Reveal — Infozense

2026-07-27T11:07:17.242Z

Market microstructure findings from 55.4 million order book updates on the Stock Exchange of Thailand. Spread landscape, depth analysis, imbalance signals, auction dynamics.

Where AI Actually Earns Its Keep in Hospital Revenue Cycles | Infozense

2026-07-27T11:07:17.242Z

ICD-10-TM pre-submission audit for Thai hospitals. Rule-based, on-premise, calculator-checkable. First three hospital partners; 60-day measured pilot.

Your Coding Has Six Leaks. They're All One Problem | Infozense

2026-07-27T11:07:17.242Z

Under-coding, denials, audit exposure, drug-claim clawbacks, code lookup, thin documentation — a Thai hospital's reimbursement team fights six problems with six tools. They're one. One governed knowledge base catches all six: pre-bill, cited, checkable in baht.

FPGA Real-Time Market Data — Infozense

2026-07-27T11:07:17.242Z

4,500 live order books on a single FPGA. <1us query latency. From the team that processed 85M tick records across 4 databases — now moving to silicon.

Build the Consultant. Once. — Knowledge Engineering | Infozense

2026-07-27T11:07:17.242Z

The reading is the expensive part. Build the in-house AI consultant on a system you own — every claim cited, every access audited, the AI on top replaceable, the documents and audit log durable. The Knowledge Engineering behind the built-in-house consultant.

Anatomy of an Anomaly — Multi-Signal Detection | Infozense

2026-07-27T11:07:17.238Z

No single signal is enough. When 4 signals fire together, that's a pattern worth investigating. 10.4M trades, 214 alerts, 65ms detection.

When One Stock Steps Out — Tick Data Intelligence

2026-07-27T11:07:17.238Z

SET's Dynamic Price Band auto-pause from raw ITCH: a deterministic 120.000-second per-stock re-auction, 155 events across 22 days, and how execution, risk, and surveillance desks apply it. Real SET tick data.

Where the Manipulation Fingerprint Lives — Tick Data Intelligence

2026-07-27T11:07:17.238Z

Aggregated tests on SET data show null. A tier-stratified event study, after one critical correction, finds Harris's asymmetric-impact signature alive in mid-volume stocks. The high-volume tier absorbs it; the low-volume tier is underpowered. The fingerprint lives in the middle.

Weak Signals, Strong Verdicts — Tick Data Intelligence

2026-07-27T11:07:17.238Z

588 to 2,125 to 5,240 to 195 to 214 alerts. How weak signals become strong verdicts through multi-signal scoring. Real SET ITCH data.

How Much Signal Is in L1? — Tick Data Intelligence

2026-07-27T11:07:17.238Z

Deeper book levels add noise, not signal. On SET data, L1-only beats L1–L10 at every horizon out-of-sample — and at long horizons, deeper-book models go negative. 22 days, 25 stocks, 750 OOS regressions.

The Cost of Moving Size — Tick Data Intelligence

2026-07-27T11:07:17.234Z

Anyone can see the order book. Almost no one knows what it costs to take it. Walk-the-book on 22 days of SET tick data: POV impact curves, time-of-day patterns, refill speed, capacity tiers, and why half-spread is most of the bill.

Why the Spread Doesn't Move — Tick Data Intelligence

2026-07-27T11:07:17.234Z

SET's tick grid pins the spread at 1 tick 98% of the trading day. Market makers can't compete on price — only on quantity. Chapter 4 explains Chapter 3's finding: two observations, one market structure. Real SET ITCH data analysis.

Four Dimensions of Liquidity — Tick Data Intelligence

2026-07-27T11:07:17.234Z

Four dimensions of liquidity — width, depth, immediacy, resiliency — measured from real tick data.

The First 60 Seconds — Tick Data Intelligence

2026-07-27T11:07:17.234Z

How 1,447 instruments find their opening price through 2.9 million equilibrium messages. Chapter 2 of the Tick Data Intelligence series.

The Dataset — Tick Data Intelligence

2026-07-27T11:07:17.234Z

Multiple message types. Millions of records. What your exchange feed contains — and what you're not using. Chapter 1 of the Tick Data Intelligence series.

From Raw Data to Real Intelligence — Infozense

2026-07-27T11:07:17.234Z

85 million market events. 4 database architectures. End-to-end data engineering from raw protocol parsing to production analytics. Real data, measured results.

You Governed the Agent. You Forgot the Data Layer Beneath It. — Infozense

2026-07-27T11:07:17.234Z

Every AI agent project focuses on two things: how the agent works (orchestration) and what keeps it in line (guardrails). Both are the agent itself, the part you can see. Beneath it sits the data layer, where your regulated risk actually lives: what the agent can reach, what may leave your organization, and what you can prove months later. Almost no one is governing it.

You Deleted the Record. Your AI Didn't. — Infozense

2026-07-27T11:07:17.234Z

Someone asks you to delete their data. You delete the record. But months ago you fed that record to your AI, and it can still pull the person's data into an answer from a source that officially no longer exists. Under today's privacy laws, a deleted row is not the same as a forgotten person. A governed data layer makes the deletion reach the AI, and proves it did.

The Missing Piece Between Your Data and Your AI — Infozense

2026-07-27T11:07:17.234Z

Everyone is building AI agents. Almost no one is building the layer beneath them. So each agent connects straight to the data, one at a time — and the governance, the consistency, and the cost break apart across every connection. The missing piece isn't a smarter agent. It's the governed layer they should all share.

The Static Pharmacy — Hospital Drug Supply Intelligence

2026-07-27T11:07:17.230Z

73M THB per year lost to static inventory management in a 500-bed Thai hospital. 12 vital drug stockouts per month. 5% drug expiry rate.

From Scattered Signals to Forecast Intelligence — Infozense

2026-07-27T11:07:17.230Z

A data-driven study on demand forecasting for Southeast Asian supply chains — real data, measured results, honest tradeoffs.