Teh Tarik Digital · Solutions · Semiconductor & Electronics

Malaysia's Fab Floor Has a Workforce Problem, Not a Model Problem

Malaysia's semiconductor and E&E sector is expanding faster than its workforce can follow — a 10x gap between engineer supply and industry demand, per SEMICON Southeast Asia 2026, with 12,000–15,000 unfilled technical roles expected by year-end. At the same time, revised Employment Pass salary floors (effective 1 June 2026) have raised the cost of technical headcount. QC and inspection lines are where this pressure shows up first: hard to staff, harder to staff consistently, and the first place a labor shortage becomes a yield problem.

Teh Tarik Digital builds computer vision and workflow automation that takes pressure off that line — not a hardware appliance you bolt on, but systems that integrate into what you already run.

10x
engineer supply-demand gap
15K
unfilled roles by year-end 2026
85%+
fault prediction accuracy, 7-day horizon
40%
reduction in incident response time

Who This Is For

// FIT

Not a generic "innovation" pitch — this is scoped for plant and ops managers evaluating QC automation against a shrinking technical labor pool. If your evaluation criteria is "does this sound impressive in a demo," a lot of vendors clear that bar. If it's "does this hold up on a real line with real defect variance," the field narrows fast — and that's the bar we're building to.

Front-end and back-end semiconductor manufacturers — wafer fab, OSAT, ATE, EMS.
Precision engineering and tooling suppliers in the Penang–Kulim corridor.
Plant and ops managers evaluating QC automation against a real labor constraint, not a trend.

Core Capabilities

// CAPABILITIES

Four capabilities, each one built to integrate rather than replace — the goal is never a rip-and-replace project. A vendor selling you new hardware is selling you a longer, riskier deployment than one building on cameras and systems you already have running. We picked the harder path on purpose, because it's the one that actually ships on a real production timeline.

01

Defect Detection

YOLO-based real-time visual QC, calibrated to your defect classes — not a generic off-the-shelf model trained on defects that don't match what your line actually produces.

02

Shift & Yield Reporting

Auto-generated from CV output the moment inspection happens, reducing manual QC logging — the report is a byproduct of the system working, not a separate task someone has to remember at shift-end.

03

Systems Integration

Plugs into existing camera/line hardware and ERP/MES systems; Teh Tarik Digital is the integrator layer, not another vendor portal — the deployment risk lives in the integration, and that's where most of our actual work happens.

04

PDPA-Aligned by Design

Where people-facing monitoring is involved — PPE compliance, safety zones — built with facial recognition disabled by default, written-authorization gated per s.40, so the compliance question is answered before a camera goes live, not after.

Benefits & Results

// PROOF

A workforce gap doesn't fix itself with better hiring — it needs the line to run without depending on headcount you can't find. The numbers below aren't projections; they're measured on SmartPole OS, a platform already running at scales from a 50-pole deployment to a 2,000-pole network, which is the actual test of whether an architecture holds up or just looks good in a pilot.

What Changes

  • Consistent defect detectionHolds a standard shift after shift, unlike manual QC under fatigue or turnover.
  • No headcount scrambleAbsorbs QC demand without proportional hiring cost.
  • Faster shift/yield reportingAuto-generated from CV output.
  • No rip-and-replaceIntegrates into existing line hardware and MES/ERP.

What We've Measured

  • SmartPole OS — 85%+ / 40%85%+ fault prediction accuracy on a 7-day horizon using existing telemetry — no additional instrumentation required — and cuts incident response time by 40% via automated alert routing.
  • Scales without re-architectureThe same platform runs from a 50-pole deployment to a 2,000-pole network on identical architecture — proof the underlying approach holds at very different scales.

How It Works

// PROCESS

Four stages, scoped to your actual defect taxonomy — not a generic deployment checklist. Skipping the assess phase is the single most common way a computer vision project fails: a model trained on someone else's defect classes will miss the ones that actually matter on your line.

01

Assess

Line audit, defect taxonomy, existing camera/hardware inventory — nothing gets proposed until we've actually seen what's running on your floor.

02

Build

CV model trained on your actual defect classes, not generic datasets — coverage that matches your specific production line, not a one-size-fits-all defect library.

03

Integrate

Connects to line hardware and MES/ERP without replacing what's already there — this is where a project actually succeeds or fails, and where we spend most of our time.

04

Operate

Live QC dashboard, auto shift reports, escalation on flagged defects — with the same team that scoped the audit staying on for ongoing support.

// FROM THE LAB

Real Builds, Real Numbers

The same architectural discipline behind our computer-vision QC builds — documented from a different system.

MULTI-AGENT ARCHITECTURE

6 Months of Iteration on a Multi-Agent Architecture

The state-schema-first discipline behind IRIS v2's supervisor-specialist architecture — the same approach we apply to every build, including QC pipelines. Full case study →

FAQ

// 9 QUESTIONS

The questions plant and ops managers ask before they'll commit to a line audit. Most of these come down to one underlying concern: will this actually work on my line, or does it just work in a controlled demo environment.

Computer vision vs. manual QC inspection?
Manual QC scales with headcount you may not be able to find — Malaysia's semiconductor sector has a 10x gap between engineer supply and demand. Computer vision covers the line consistently regardless of labor availability. See our Lab breakdown.
Do you replace our existing MES/ERP systems?
No — we integrate into your existing MES/ERP and line hardware rather than requiring a replacement. See how we structure that integration layer.
What defect types can this detect?
The model is trained on your actual defect classes during the assess/build phase, not a generic off-the-shelf dataset — so coverage matches your specific production line.
Is this one model doing everything, or several working together?
Several, coordinated — defect detection, shift reporting, and escalation routing are handled by separate specialised agents. See our multi-agent orchestration breakdown.
How does escalation work when a defect is flagged — does a person still review it?
Yes — the system flags and routes, a person confirms and acts. Every escalation is logged with the decision attached. See our human-in-the-loop design breakdown.
What's the difference between this and n8n or other workflow-automation tools we've evaluated?
General-purpose workflow tools weren't built for real-time visual inspection at production-line speed. See our comparison of n8n, Kestra, and LangGraph.
Which AI models power the defect detection — is this locked to one vendor?
The CV layer is YOLO-based and model-agnostic at the orchestration level. See our comparison of Claude, GPT, and Gemini.
Is worker-facing monitoring (PPE compliance, safety zones) PDPA-compliant?
Yes — facial recognition disabled by default, written-authorization gated per s.40. See our PDPA data-privacy breakdown.
Does this extend beyond QC — warehouse, logistics, dispatch?
Yes — the same computer vision and workflow-orchestration architecture applies to warehousing and logistics operations. See our logistics automation breakdown.

Tell us your defect classes and line setup — we'll scope what a real audit looks like.