ARTIFICIAL INTELLIGENCE

AI that does the work,
not just the talking

We put AI where the result is measurable — fewer inbound calls, fewer site visits, less data entry — with guardrails and a human approver on anything risky.

01 — CONVERSATIONAL AI

AI Customer
Service Agent

The bot works like a full-time service representative — it understands natural Thai, reads the live customer database, and completes transactions rather than merely replying.

  • Answers general questions: opening hours, contact numbers
  • Looks up customers: name, outstanding balance, active services
  • Accepts payment → verifies the slip → records it → activates service
  • Hands off to a human on request, with a conversation summary

Real conversation example

● ONLINE

Just transferred — internet for unit 405

AI Got it — could you send the transfer slip?

[ slip image ]

AI Slip verified — ฿300, bank and account number match.
✓ Unit 405 activated valid until 31 Aug 2026.

2.4s response · no human involved

STEP 1

Receive message

STEP 2

Understand intent

STEP 3

Verify + activate

STEP 4

Escalate if needed

ARCHITECTURE

Why two tiers of AI

Simple turns go to a small, fast, cheap model; complex ones escalate to a stronger one. That is what keeps cost under control as chat volume grows.

Tier 1 · The receptionistFAST

Greeting, general questions, collecting basic details — fast and inexpensive.

Tier 2 · The analystREASONING

Database lookups, slip verification, activation decisions, multi-step tool use.

Tier 3 · A humanHANDOFF

On customer request or low AI confidence — the bot stops the moment staff join the thread.

02 — AUTONOMOUS NETWORK

NAMS — AI-driven
network configuration

A local-first AI agent for multi-vendor network operations. Two halves: configuration in natural language, and genuine troubleshooting analysis — not just querying a monitoring database.

Supports Juniper · MikroTik · Cisco · TP-Link · Fortigate, escalating in stages — on-device model → cloud model → human — as confidence drops.

Every write to a device passes these five stages

  1. 1Read actual state firstRead-before-write — never assume the device's condition.
  2. 2Dry-runSimulate the outcome and assess the blast radius before anything is applied.
  3. 3Human approvalApproval is accepted from people only — a deliberate defence against prompt injection.
  4. 4Confirmed commit, then verifyIf confirmation doesn't arrive in time, the device reverts itself.
  5. 5A rollback plan, alwaysNo change is allowed that cannot be undone.

Source of truth: the central inventory is always authoritative — changing a device outside the system is forbidden, and drift detection is mandatory.

03 — PRIVATE AI INFRASTRUCTURE

AI in-house,
data stays in-house

Many organisations want AI but cannot send customer records, contracts or source code to a public cloud. We deploy language models on your own hardware and escalate outward only for work the local model judges itself insufficiently confident to handle.

The clearest benefit

Predictable cost — you pay for the machine once instead of paying per word every month. The more you use it, the better the economics get.

Example — six workloads on a single machine

  1. 01Primary agent brainReasoning, tool-calling and Thai-language communication
  2. 02Customer service botAbsorbs most chat volume locally, cutting external API spend
  3. 03Coding pipelineAssists developers without source code leaving the organisation
  4. 04NOC auto-remediationAnalyses monitoring alerts, then proposes or applies a fix within its granted authority
  5. 05Internal knowledge RAGSearch manuals, policies and historical documents in plain language, with citations
  6. 06Thai-language co-pilotDrafting, summarising and correspondence support for general staff

04 — DOCUMENT AI

TaxLens — tax invoice capture

Receipt and input-tax invoice capture using a vision language model, built for accounting teams — removing the repetitive data entry that consumes the team at every month end.

  1. 01Photograph it — a phone camera is enough, no scanner needed
  2. 02AI extracts — tax ID, VAT amount, date, vendor
  3. 03Automated validation — mismatched totals, malformed IDs, duplicates
  4. 04Humans review only flagged items — not every document
  5. 05Export the input-VAT report — .xlsx for the Thai P.P.30 filing

05 — MEETING AI

AegisMeet — on-premise meeting AI

An online meeting platform that runs entirely inside your organisation — no audio and no transcript ever reaches an external cloud. Built for teams who discuss contracts, pricing and customer data.

  • Real-time Thai captions — text appears as people speak
  • Two-way Thai ↔ English translation — meet foreign partners without an interpreter
  • Automatic recording and minutes — the summary is ready when the call ends
  • Searchable history via RAG — ask "what did we agree on pricing?" directly

Designed for organisations whose policy forbids discussing classified material on foreign conferencing services.

06 — GUARDRAILS

AI with brakes

AI that touches production needs hard limits. We never let it make an irreversible decision, and we never accept an approval instruction from anywhere except a person.

Blast-radius limits

A ceiling on how many endpoints one change may affect; beyond it, the work must be split or separately approved.

Change windows

Risky work happens only in agreed windows, never at peak usage.

Forbidden command patterns

Commands that would sever the management path or destroy data permanently are blocked upfront.

Human-only approval

A deliberate defence against instructions smuggled in through data (prompt injection).

No raw data leaves

If an external model is genuinely needed, identifying data is masked or aggregated first — always.

Every decision recorded

You can audit afterwards what the AI did, why it did it, and who approved it.

Wondering which of your workflows AI could take on?

Tell us what consumes most of your team's time and we will assess whether handing it to AI is worth it — including the cases where the honest answer is "not yet."