NeuronX
EN ไทย
Live demo

Why does every decision still wait for someone senior?

Answers with proof.
Decisions without waiting.

Your people can't act until someone confirms which rule is current. NeuronX answers from your documents, regulations and every amendment, with the clause and the version in force, so teams decide on their own. The AI agents you add next act on the same proof, on your own servers.

Try the live demo See how it works
  • Thai and English documents
  • Runs on your own GPUs
  • Measured in production
Procurement · Example Co. Example data
  1. Scan
  2. Thai OCR
  3. Version check
  4. Knowledge graph
  5. Answer

บริษัทตัวอย่าง จำกัด

ระเบียบว่าด้วยการจัดซื้อจัดจ้าง (ฉบับที่ ๒) พ.ศ. ๒๕๖๘

ข้อ ๑๒ การจัดซื้อโดยวิธีเฉพาะเจาะจง ให้กระทำได้ในวงเงินครั้งละไม่เกิน ๑๐๐,๐๐๐ ๕๐๐,๐๐๐ บาท

มีผลบังคับใช้

Changes shown against the 2021 edition

วงเงินจัดซื้อแบบเฉพาะเจาะจงสูงสุดเท่าไรWhat is the limit for direct procurement?

ไม่เกิน ๕๐๐,๐๐๐ บาทต่อครั้ง ตามระเบียบฯ (ฉบับที่ ๒) ข้อ ๑๒ ซึ่งแก้ไขวงเงินเดิม ๑๐๐,๐๐๐ บาท

Up to 500,000 baht per purchase, under Amendment No. 2, clause 12, which replaced the earlier 100,000 baht limit.

Regulation No. 2, clause 12 · in force 2021 edition · amended, not used

Measured on a live production system

Right document in the set the AI reads before answering (Recall@20)
88.6%
Right document in the top 10 results (Recall@10)
85.9%
AUC-ROC for telling answerable from unanswerable questions
0.941
Real documents across 16 categories
249

A Thai state energy enterprise · production deploymentRead the case study

What it costs you today

Your senior people have become your company's search engine.

Every uncertain question goes up the chain, to the few people who remember which rule is current. They spend their days answering instead of deciding, the company moves only as fast as they are free, and when they retire or resign, what they know leaves with them.

  • Every answer waits for someone

    Fixed in 01 Know

    Staff ask around on chat and wait, and new hires take months to find their way through scanned files and old folders.

  • The rule changed. Not everyone knew.

    Fixed in 02 Trust

    Someone acts on last year's version, and it surfaces in the audit.

  • Too sensitive to share

    Fixed in 03 Control

    Not everyone may see every document, so nothing gets opened up, and AI tools are blocked before they start.

The company agentic platform

Four layers to an agent-ready company.

Each layer stands on the one below it. Start with your documents and finish with agents that act.

LiveBetaComing soon

  1. Know

    Live

    Read every document, in Thai and English.

    • Upload PDFs, scans, Office files, CSV, audio and video into business domains.
    • Thai OCR built for official documents, with a review step before anything is indexed.
    • A knowledge graph of the people, rules and terms inside your documents.
    • Ask questions in chat, and see the documents each answer came from.

    Module · Document Manager

  2. Trust

    Live

    Answer only from the version in force.

    • Every upload becomes a revision of a document, so its history is never lost.
    • Amendments and repeals are tracked as links between documents, the way regulations actually change.
    • Only the version in force reaches the knowledge graph. Nothing is retired on its date alone.
    • Knowledge health shows duplicates, unpublished revisions and drift for each domain.

    Module · Archive Management

  3. Control

    Beta

    Decide who, and which agent, sees what.

    • Model your organisation as units, people, groups and roles, or start from a template.
    • Grant access per domain, and read the reason behind every decision.
    • Guardrails stop people raising their own access and keep an owner on every domain.
    • An append-only audit log that exports to CSV.

    Module · Access Control Manager

  4. Act

    Beta

    Turn knowledge into work.

    • Describe a task in plain language and get a skill you can reuse and audit.
    • Skills pull evidence from your documents, your databases and your APIs.
    • Every published skill carries test cases that run each night.
    Coming next:Document workflows with approvalsWorkflow automationAgent squads

    Module · Skill Builder

Case study

Thai state documents, measured in production.

A Thai state energy enterprise

The enterprise needed staff to find answers in a large archive of scanned Thai documents: contracts, regulations, announcements, orders, reports and acts. NeuronX runs on the organisation's own GPU servers, with an active and a standby system kept in sync.

We tested the live system with questions written from the documents themselves, 51 with an answer in the archive and 20 without one. Every question passed the same filters before any result was seen.

Measuring also found three production bugs, all now fixed: Thai vowels and tone marks dropped by the wrong OCR engine, uploads failing on long Thai filenames, and empty chat answers.

Where this deployment is on the journey
  1. KnowIn production
  2. TrustNext
  3. ControlNext
  4. ActNext
Evaluation · live production system
MetricResultWhat it means
Recall@200.886Right document is in the set the AI reads
Recall@100.859Right document is in the top 10
AUC-ROC0.941Tells answerable from unanswerable questions
Corpus249Documents in 16 categories
Questions51 + 20With an answer + without one

Marker shows the 80% target set by the customer

How we measured

Recall@20 counts a question as found when the right document is among the roughly 20 passages the model reads before it answers. With 51 questions the margin is about ±0.10 at 95% confidence. The full test re-runs in about 20 minutes after any change to the system.

Deployment

Runs where your data lives.

NeuronX is installed inside your organisation, on hardware you control.

Your servers, your GPUs

Models for reasoning, Thai OCR and embeddings run on your own hardware, behind your firewall. Documents, the knowledge graph and audit logs stay in your own databases.

Reasoning
Qwen3.6 35B
OCR
Typhoon OCR 1.5
Embeddings
BGE-M3

Models used in the reference deployment above. All open-weight.

Built to stay up

An active and a standby system with database replication and a floating address, plus a written runbook for failover and failback.

Thai first, English too

OCR, search and text normalisation tuned for Thai, Thai numerals included.

Measured on your documents

A repeatable evaluation runs against your own archive after every change, so regressions show up before your users find them.

Choose your models

Swap in better models through standard OpenAI-compatible endpoints, without rebuilding the platform.

Live demo

Open a working workspace.

The live demo is a real NeuronX workspace. Sign in, open a domain, check its knowledge health, and ask a question.

Open the live demo

Demo sign-in

Ask the NeuronX team for a demo sign-in.