i
Measure first.
We refuse to build an agent without an evaluation rubric. If we can't tell when it works, we can't ship it.
Where Agentic AI is Engineered.
Agentic AI engineering from a boutique R&D laboratory in Penang, Malaysia — custom AI agents designed for production with the patience of a workshop and the rigour of a foundry.
Lab №01 Custom Fine-Tuning · Token-Efficient Architecture · Tera Grid Deployment agenticlab.com.my
What is Agentic Lab?
Agentic Lab is a boutique agentic AI engineering (R&D) laboratory based in Penang, Malaysia that builds production-grade custom AI agents through custom model fine-tuning, token-efficient architecture, and deployment to Tera Grid — the production fabric operated by Ai Teragrid. Engagements run 7–13 weeks from consultation to launch. Operated by AITG Sdn Bhd (Reg. 202601016521).
Built with
- Llama
- ·
- Qwen
- ·
- Mistral
- ·
- Gemma
- ·
- vLLM
- ·
- TGI
- ·
- Triton
- ·
- LoRA
- ·
- QLoRA
7–13wkMedian engagement
99.9%Production SLO
≤1.5kTokens / call median
APAC∞Primary region
Four steps. No shortcuts.
Each engagement moves through the same disciplined arc — from the first conversation to the moment your agentic AI answers its first production request.
-
01
Consultation
We meet your team, study the workflow, and define what good looks like in measurable terms. No agent is built before the success criteria are written.
-
02
Architecture
We choose the base model, design the tool surface, and draft a token-efficient architecture. Where useful, we curate a fine-tuning corpus from your proprietary data.
-
03
Sandbox
The agent is assembled inside a private Lab Environment. We run adversarial evaluations against real cases until the rubric from Step 01 is satisfied.
-
04
Tera Grid Deployment
The finished agent ships to Tera Grid — Ai Teragrid's production fabric — with autoscaling inference, observability, and per-request cost telemetry from the very first call.
Three rules. The rest follows.
ii
Tokens are money.
Every prompt is a budget. We design with retrieval, distillation and caching so production economics survive scale.
iii
Ship to the grid.
A demo is not a deployment. We hand off agents that scale, observe themselves, and report their own cost-per-request.
A Lab Environment for proprietary agents.
Every agent we build is staged inside a closed sandbox — a private rehearsal room where prompts, tools, and weights can be perturbed without consequence to your production systems.
Input Stream
queue: 12 / 1024
Evaluation
suite: adversarial-1024
lab> eval --suite adversarial --runs 1024
▸ accuracy ........ 0.927
▸ p95 latency ..... 412 ms
▸ tokens / call ... 1,148
▸ cost / 1k calls . US$ 0.41
status: PASS · promote → tera-grid- Isolated. A dedicated tenant; your data never leaves it.
- Reproducible. Every eval run is hashed and replayable.
- Adversarial. Red-team prompts ship with the suite.
- Observable. Token, latency, and cost telemetry on day one.
Agentic AI engineering. Token-efficient by design.
Agentic AI engineering is the discipline behind every engagement — custom AI agents that plan, call tools, and act, built by an AI agent development team in Malaysia and held to production standards.
A
Custom model fine-tuning
We curate task-specific corpora from your proprietary data and fine-tune compact open-weights models — LoRA, QLoRA, or full-rank where the gradient justifies it — until the agent speaks your domain natively.
- Methods
- SFT · DPO · LoRA · QLoRA
- Bases
- Llama · Qwen · Mistral · Gemma
- Eval
- Held-out + adversarial suites
B
Token-efficient architecture
Long context is a luxury, not a strategy. We design retrieval over context-stuffing, distil large teachers into small students, and cache aggressively — so that an agent which costs a dollar in the demo costs a cent in production.
- Patterns
- RAG · Tool-use · Routing · Caching
- Targets
- ≤ 1.5k tokens / call median
- Telemetry
- Per-request token & cost trace
C
Tera Grid deployment
Finished agents ship to Tera Grid — the production fabric operated by Ai Teragrid. Autoscaling inference, observability, and SLOs come standard. Your team gets a dashboard; we get the pager.
- Runtime
- vLLM · TGI · Triton
- Region
- APAC primary · global edge
- SLO
- 99.9% availability
Agentic AI in the field. What we build, where it works.
Notes from the lab floor — the honest version of what agentic AI engineering means for Malaysian manufacturers, retailers, and enterprise teams. Written for buyers, not conference audiences: what an agent actually does all day, and what it costs to run one.
Short answer
Agentic AI engineering is the discipline of designing, fine-tuning, evaluating, and deploying AI agents that plan and act on their own — booking, checking, filing, and following through instead of only chatting. Agentic Lab, a boutique R&D laboratory in Penang, Malaysia, engineers such agents for manufacturing, retail, and enterprise teams, then ships them to production on Tera Grid with cost telemetry from the first call.
| Agentic Lab | In-house build | General consultancy | |
|---|---|---|---|
| Evaluation rubric | Written before any code; agent cannot ship until it passes. | Often informal; quality judged by demo feel. | Slide-deep; rarely instrumented. |
| Fine-tuning depth | LoRA / QLoRA on open-weight models, corpus curated from your data. | Possible, but competes with product roadmap. | Usually prompt-level only. |
| Token economics | Per-request cost telemetry is a deliverable, target ≤1.5k tokens/call. | Discovered at the end-of-month invoice. | Rarely modelled. |
| Deployment target | Production on Tera Grid — autoscaling, observability, SLOs. | Whatever the team can operate. | Handover document. |
| Engagement shape | 7–13 week lab project, NDA + DPA signed first. | Open-ended hiring bet. | Advisory retainers. |
Why Malaysia, why now.
LHDN e-Invoicing
Malaysia's e-invoicing rollout turns every invoice, credit note, and validation error into structured work. Agents that reconcile, chase missing fields, and file exceptions are the first thing most finance teams ask us for.
e-invoicing agents · LHDN
PDPA 2010 discipline
Personal data handling is a design constraint, not an afterthought. Every engagement signs a DPA, keeps customer data inside agreed boundaries, and documents what the agent may read, remember, and send.
PDPA-compliant AI agents
National AI agenda
With NAIO coordinating the national AI roadmap and MDEC pushing enterprise adoption, Malaysian organisations are being asked for AI plans, not AI experiments. We supply the engineering that makes the plans real.
NAIO · MDEC · AI roadmap
Penang engineering depth
Penang's five decades of E&E manufacturing engineered the supply chain; the same discipline now applies to AI agents. Local talent, local data centres, and regional timezones keep engagements fast and accountable.
Penang AI R&D
Open-weight sovereignty
Not every workload belongs on a foreign API. We fine-tune Llama, Qwen, Mistral, and Gemma so the weights, the data, and the inference can stay within Malaysia when the mandate requires it.
sovereign AI Malaysia
Cost visibility
Malaysian enterprises are quoting AI budgets in ringgit now, not innovation-lab dollars. Per-request token telemetry means a department head can see what an agent costs before finance asks.
AI cost per request · MYR
Where agents earn their keep.
Manufacturing
Production-line reporting agents that read shift logs and sensor notes, flag excursions, and draft the quality report before the meeting. Engineers review; the agent does the collation. The result: morning meetings start with the report already on the table.
agentic AI for manufacturing Malaysia
Chain retail
Store-operations agents that reconcile promotions across outlets, chase missing closing reports, and answer "did all branches run the campaign?" from data, not from phone calls. Head office stops chasing; the agents report.
AI agents for chain retail
Large corporations & MNCs
Regional teams drown in reconciliations between systems that were never meant to talk. Agents bridge ERP, email, and spreadsheets with audit trails the compliance team can actually read — in English or Bahasa Malaysia, per region.
enterprise AI agents Malaysia
Banking & financial services
Document-heavy onboarding and KYC reviews produce exactly the repetitive multi-step work agents are built for — with every decision logged and reviewable under the bank's own risk framework.
AI agents for BFSI
Logistics & supply chain
Shipment exceptions, customs paperwork, and POD chasing across vendors — agents that read the emails, update the system, and escalate only the genuine judgement calls to the humans on shift.
supply-chain AI agents
Public sector
Enquiry triage and form-completing assistance for agencies serving citizens in Bahasa Malaysia and English — measured against a rubric, deployed inside data boundaries the agency controls.
govtech AI agents Malaysia
How to choose an agentic AI engineering partner.
01
Ask for the rubric first.
A serious lab writes how the agent will be judged — accuracy, escalation rate, cost per request — before quoting. If evaluation is a phase-five deliverable, walk away.
02
Weigh token economics, not demos.
A dazzling demo that burns 40k tokens per task is a liability at scale. Ask every candidate what the median call costs and how it is measured in production.
03
Check where the data lives.
Fine-tuning on proprietary data requires an NDA, a DPA, and a clear answer on residency. Malaysian PDPA obligations and any sector regulator's rules must be in writing.
04
Demand a deployment story.
"We hand you a notebook" is not deployment. The partner should name the runtime, the observability, and who pages whom at 2am when the agent misbehaves.
05
Start with one workflow.
The strongest engagements begin with a single painful workflow and a 7–13 week arc, not a company-wide AI transformation programme on day one.
Request a Lab Consultation.
Tell us where the agent should live in your workflow. A lab lead replies within one business day — never an autoresponder.
Frequently asked.
What is agentic AI?
Agentic AI describes AI systems that plan, make decisions, and act autonomously — calling tools, querying data, and completing multi-step tasks rather than only replying in chat. Agentic AI engineering is the discipline of building such agents to production standard, which is what Agentic Lab does from Penang, Malaysia.
Agentic AI vs chatbots — what is the difference?
A chatbot answers questions in a conversation window. An agentic AI system plans multi-step work, calls tools and APIs, verifies its own results, and acts inside your workflow. Agentic Lab engineers the second kind: custom AI agents that are measured, token-efficient, and deployed to production on Tera Grid.
What is Agentic Lab?
A boutique agentic AI engineering (R&D) laboratory operated by AITG Sdn Bhd that builds production-grade custom AI agents through custom fine-tuning, token-efficient architecture, and Tera Grid deployment.
Where is the lab based?
Headquartered in Penang, Malaysia. Engagements run remotely with on-site sessions across Southeast Asia and worldwide.
What does "token-efficient architecture" mean?
An agent design discipline that minimises the tokens consumed per task — compact prompts, retrieval over context-stuffing, distilled smaller models, aggressive caching — so production agents stay fast and economical at scale.
What is Tera Grid?
Tera Grid is the production fabric operated by Ai Teragrid where finished agents are deployed. It provides autoscaling inference, observability, and per-request cost telemetry.
How long does an engagement take?
From Consultation to Tera Grid Deployment, typical engagements run 7–13 weeks depending on data readiness and evaluation rigor.
How much does custom AI agent development cost in Malaysia?
Custom AI agent development in Malaysia is scoped per project. Agentic Lab quotes after a Lab Consultation, based on data readiness, fine-tuning depth, evaluation rigour, and deployment scale. Typical engagements run 7–13 weeks and target enterprise budgets — request a consultation for a project-specific quote.
How do I request a Lab Consultation?
Submit the consultation form on agenticlab.com.my or email [email protected]. A lab lead replies within one business day, never an autoresponder.
Do you sign NDAs and DPAs?
Yes. We sign mutual NDAs before the first deep-dive and execute a Data Processing Agreement before any client data enters the sandbox.
Do you build AI agents for manufacturing and retail companies in Malaysia?
Yes. Manufacturing and chain retail are core Agentic Lab territories: production-line reporting agents, quality-excursion triage, demand-aware inventory agents, and store-operations assistants. Each engagement starts from your workflow, not a template, and ends in deployment on Tera Grid.
Can you fine-tune models on our proprietary data in Malaysia?
Yes. Under a signed NDA and Data Processing Agreement, we curate a fine-tuning corpus from your data (LoRA / QLoRA on open-weight models) so the knowledge stays yours. PDPA 2010 obligations are handled in the DPA, and sovereign deployment options inside Malaysia are available on request.
What is the difference between an AI agent and RPA?
RPA follows a fixed script: it clicks the same buttons in the same order and breaks when a screen changes. An AI agent reads the situation, decides, and routes around exceptions. Agentic Lab typically deploys agents alongside existing RPA so the robot handles the happy path and the agent handles everything else.
Do you work with clients outside Malaysia?
Yes. Engagements run remotely worldwide, with a focus on APAC — Malaysia first, then Singapore, Indonesia, Thailand, Vietnam and the Philippines. On-site sessions are available across the region.