Nasi Campur Sample Collection Method
Field Guide & Commercial Framework for Bibora Vision AI Edge Optical Calibration
Comprehensive manual for hardware deployment leads, food & beverage commercial executives, and merchant operators. Combines high-density optical training protocols with the Partner Calibration Incentive Scheme to eliminate cold-start friction in high-variance mixed buffet retail.
Executive Overview & The Two-Phase Pipeline
Nasi Campur (and regional mixed rice variations including Nasi Kandar, Nasi Padang, and Economical Rice) presents one of the most demanding computer vision edge-inference challenges globally. Unlike pre-packaged barcode-driven retail, plate configurations are fluid, unconstrained, and dynamically assembled by individual consumers or counter attendants.
Fluid Gravy Physics ("Kuah Banjir")
Gravies and viscous curries (Kari Kambing, Kuah Hitam, Kuah Dal) diffuse across base starches, creating amorphous non-rigid visual boundaries. Color signatures warp underlying ingredients and alter optical RGB hue-saturation values unpredictably.
Food-on-Food Occlusions
Dense three-dimensional stacking obscures secondary and tertiary food portions. A piece of Ayam Goreng Berempah may physically mask 70% of an underlying Telur Masin or Sayur Kubis portion, necessitating spatial volume deduction.
Manual Scoop Irregularity
Human hand-portioning generates extreme geometry variance. A ladled serving of Sambal Sotong varies by diameter, droplet distribution, sauce-to-solid ratio, and plate placement between different line cooks during peak shifts.
Specular Glare & Warmers
Overhead halogen food heat lamps (2700K–3200K) create localized saturated glare on oily surfaces (fried chicken skin, sambal oils, slick chili pastes), blowing out fine optical edge features on standard CMOS sensors.
The Two-Phase Data Distribution Formula
Controlled Cold-Start & Menu Seeding
Establishes initial optical anchor representations before the store's public rollout. Focuses on pure single-item profiles, pristine contrast backgrounds, standardized rotational perspectives, and multi-ingredient combinations arranged in isolated quadrants.
Live Ground-Truth Production Capture
Captures real customer checkout plates during live operations. Exposes the optical inference neural network to genuine messy piling, chaotic cross-sauce contamination, irregular cutlery shadows, and dynamic cashier interaction dynamics.
Adjust deployment split to assess operational lead time vs initial launch accuracy.
Standard Training Fee Offset Structure
Bibora invoices an enterprise "AI Vision Model Training Fee" (RM1,300 – RM3,500 per store concept) to digitize recipes, build customized convolutional weights, and conduct on-site optical calibration. Merchants participating in structured Phase 1 sampling qualify for immediate, direct invoice rebates up to 20%.
Standard Operating Setup, Hygiene & Legal Boundaries
Physical alignment tolerances and operational hygiene rules must remain uncompromised throughout the lifecycle of data acquisition. Every training image fed to the Bibora pipeline must conform to precise optical parameters.
Optical Calibration Rig Architecture
TOLERANCE: ± 5.0 mmLive telemetry from on-site edge optical diagnostic unit.
Class Coverage Matrix (Minimum Quota per SKU)
| Menu Category | Representative SKUs | Min Instances (P1) | Min Instances (P2) | Total Quota | Model Failure Consequence |
|---|---|---|---|---|---|
| Core Proteins (Tier 1) | Ayam Goreng Berempah, Ayam Percik, Rendang Daging, Ikan Keli Goreng | ≥ 75 | ≥ 75 | ≥ 150 instances/SKU | Critical Undercharge / Revenue Loss |
| Seafood & Premium (Tier 2) | Udang Galah, Sotong Sambal, Ketam Masak Lemak, Pari Bakar | ≥ 60 | ≥ 60 | ≥ 120 instances/SKU | Severe Margin Bleed / Customer Dispute |
| Vegetables & Tofu/Eggs (Tier 3) | Sayur Kubis Goreng, Kangkung Belacan, Telur Dadar, Tauhu Sumbat | ≥ 50 | ≥ 50 | ≥ 100 instances/SKU | Portion Misclassification / Cashier Override |
| Gravies & Sambals (Tier 4) | Kuah Kari Daging, Sambal Belacan, Kuah Dal, Kuah Hitam Kicap | ≥ 40 | ≥ 40 | ≥ 80 instances/SKU | False Positive Sauce Add-on Friction |
| Base Starches (Tier 5) | Nasi Putih, Nasi Kunyit, Nasi Briyani, Nasi Dagang | ≥ 30 | ≥ 30 | ≥ 60 instances/SKU | Tare Baseline Drift / Portion Overcount |
Bibora Technologies Inc., its field hardware deployment contractors, vision engineers, and sub-processors operate under a strict non-contact advisory protocol. The F&B Retail Operator retains exclusive and unassignable liability for all aspects of consumable product handling.
Operator bears sole legal and statutory responsibility for compliance with all Ministry of Health (KKM) Food Hygiene Regulations 2009, local municipal health bylaws, cross-contamination prevention, and Halal integrity certifications.
Bibora staff act solely as digital optical calibration technicians. Bibora assumes zero legal, commercial, medical, or criminal liability for any food-borne illness, spoiled inventory, allergen contamination, or regulatory fines incurred during or after sampling.
Bibora personnel are strictly prohibited by company charter from physically touching, dishing, ladling, plating, or discarding any food item. 100% of food movement must be executed by Operator’s typhoid-vaccinated, certified staff.
Phase 1 — Controlled Cold-Start Methods Breakdown (30%–70%)
Detailed technical standard operating procedures for five distinct cold-start capture methods. Each methodology trades off operational kitchen friction, dataset fidelity, and commercial cost.
Staff Meal Plating Capture
Captures natural shift meals prepared by internal restaurant staff immediately prior to opening hours (10:00 AM - 10:45 AM). Zero consumable food waste since staff eat the plated meals after capture.
Plates placed on designated calibration stage. 4 orthogonal burst photos captured at 0°, 90°, 180°, and 270° plate rotation to randomize shadow casting. 8-12 seconds total per staff member.
Staff plating suffers from heavy selection bias toward cheaper proteins (eggs, bean sprouts, chicken) and low menu coverage for premium dishes (crab, prawns, fish head curry). Quota completion requires 14+ days.
Compartmentalized Metal Tray Segregation
Food items are isolated inside stainless steel 5-compartment military mess trays. Provides perfectly clean bounding boxes with zero cross-item gravy contamination for ground-truth training.
Operator kitchen staff deposits exact 120g portions into compartmental quadrants. Overhead camera triggers multi-exposure HDR capture. Items are returned to warming wells if internal temperature stays above 63°C.
High labor burden on kitchen team. Severe Sim-to-Real domain gap: real customers never eat out of compartmentalized trays, resulting in high edge inference false rejections once live sauces blend.
Transparent Barrier / Membrane Plating
Plates are layered with ultra-thin food-grade cling film before food deposition. Eliminates plate washing and allows rapid swapping of dishes back into service wells without surface cross-contamination.
Membrane must be vacuum-tensioned to eliminate specular micro-creases. Polarizing CPL optical filters required on lens to remove film reflections. 1 capture per minute sustained throughput.
Tensioning failure creates high optical artifacts. Condensation under hot gravies blurs bottom margins. Kitchen staff find membrane management tedious during busy pre-rush hours.
Method 4: Artificial Live Sampling (Partner Loss Compensation Route)
Purpose-Built Staging Session • Zero Food Re-Serving Risk • Up to 20% AI training fee discountControlled calibration session where kitchen staff stage real, overlapping combinations of active dishes. Because food safety protocol dictates that once dishes touch, cross-contaminate sauces, or leave temperature holding they become unconsumable and must be written off as kitchen spoilage, Bibora compensates the restaurant for this food loss via an upfront service fee discount.
Conducted during off peak hours. Restaurant kitchen staff scoop target portions onto standard plates, purposely layering proteins, gravies, and sides to simulate authentic messy customer plates. Multi-angle photos captured. Once photographed, all staged food is immediately disposed of or set aside for internal staff disposal pursuant to health bylaws.
From 20% discount applied directly to the restaurant's total AI Vision Model Training Fee. Acts as direct financial compensation credit offsetting raw ingredient cost and inventory loss resulting from discarded calibration portions. Escalates to 25%–30% for large premium portions or dedicated kitchen prep assistance.
Full downward 90° overhead camera rig, standardized restaurant dining plates, tare scale, and dedicated designated food waste bin for immediate post-shot discarding.
Absolute zero public contamination risk. All staged food is formally designated as spoiled/unconsumable and is strictly prohibited from being returned to the counter display pans or service wells.
Generative AI
Utilizes NeRF (Neural Radiance Fields), Diffusion models, and 3D mesh rendering to synthesize synthetic Nasi Campur plates with randomized gravy coverage, lighting specularities, and portion angles in silico.
3D scans of raw ingredients processed via synthetic pipeline. Photorealistic domain randomization applied (chromatic aberration, blur, exposure shift). Over 50,000 synthetic frames rendered on cloud GPUs.
Severe Sim-to-Real domain gap. Synthetic models misjudge gravy viscosity and authentic Malaysian spice textures, resulting in unacceptable checkout latency (>2.4s). Subject to client engineering compute surcharge.
Phase 2 — Live Ground-Truth Production Capture (30%–70%)
Controlled cold-start establishes the baseline, but production reliability hinges entirely on exposure to in-the-wild customer plates assembled under live service chaos.
Checkout Lane Optical Trigger Integration
ZERO LINE DELAYDuring peak hour, customer flow reaches 4.2 transactions per minute. Bibora’s Phase 2 collection integrates seamlessly without slowing cashiers or requiring diner intervention:
The downward camera triggers automatically when the checkout scale senses plate mass stabilization (delta < 2g over 120ms), capturing a raw 4K frame in <15ms.
Optional micro-switch pedal or POS key tap allows the cashier to log ground-truth overrides silently without pausing transaction ringing.
Edge Privacy & Biometric Exclusion Protocol
ISO/IEC 42001 & PDPA CompliantThe optical engine enforces hardware-level spatial fencing. The camera field-of-view is strictly cropped to the 300mm x 300mm plate perimeter stage:
- check Zero Face & Body Capture: Sensor geometry physically prevents human facial elevation angles.
- check Hand & Card Anonymization: Real-time edge segmentation auto-redacts customer fingers, payment cards, smartphones, and cash notes with black solid masks prior to frame serialization.
- check Local Inference Only: Video streams are never recorded or stored. Only single static cropped image frames containing food items are ingested.
Low-Velocity Menu SKU Mitigation
In live buffet production, popular items (Ayam Goreng) collect 200+ snaps within 3 hours, while specialized items (Ikan Pari Asam Pedas, Peria Katak) may collect fewer than 8 daily instances.
Comparative Evaluation & Commercial Decision Matrix
Direct multi-dimensional audit comparing all data acquisition methodologies across out-of-pocket operational costs, kitchen friction, food safety risk, and machine learning accuracy impact.
| Collection Method | Data Share Potential | Food Spoilage / Waste Risk | Kitchen Operational Friction | Hygiene Risk to Public | Real-World Fidelity | Partner Training Fee Impact |
|---|---|---|---|---|---|---|
| M1: Staff Meal Plating | Phase 1 (30%) | Zero (Staff eats meal) | Low (Pre-rush only) | Zero (Staff consumed) | Moderate (Portion skew) | 0.0% Rebate (Standard) |
| M2: Metal Tray Segregation | Phase 1 (40%) | Low (Returned to warmer) | Very High (Loading labor) | Low (Immediate return) | Low (Domain gap error) | 0.0% Rebate (Standard) |
| M3: Plastic Membrane Plating | Phase 1 (30%) | Low (Food returned) | Moderate (Film handling) | Low-Med (Tear risk) | Moderate (Glare artifacts) | 0.0% Rebate (Standard) |
| starM4: Intentional Food Spoilage Concession | Phase 1 (50%–70%) | High (Intentional Spoilage / Written Off) | Low (Off-peak prep) | Zero (Fully Discarded / Staff Consumed) | High (Authentic Gravies & Stacking) | From 20% to 30% Discount (Loss Compensation Credit) |
| M5: Generative AI Augmentation | Phase 1 (30%) | Zero (In Silico / Synthetic) | Zero (Kitchen untouched) | Zero (No physical food) | Poor (Gravy sim gap) | +15% Engineering Surcharge |
| P2: Live Production Ground-Truth | Phase 2 (30%–70%) | Zero (Customer consumed) | Zero (Sub-second passive checkout) | Zero (Non-contact capture) | Maximum (100% Real world) | Mandatory Core Pipeline |
Method 4 completely eliminates kitchen anxiety over keeping photographed dishes re-servable. By staging complex combinations in a dedicated 2.5-hour morning block, optical calibration is finalized on Day 1. This drastically cuts hardware lock-in time, allowing a single kiosk rig to be deployed, calibrated, and cleared for other partner demos in hours rather than locking up hardware inventory across 2 to 3 weeks.
Partner Calibration Rebate Calculator
Determine your total net cost savings by choosing Method 4 Direct Sample Purchase over traditional multi-week calibration. Rebates are deducted directly from the Bibora AI Vision Training Invoice upon on-site signoff.
Rebate credited to offset restaurant food inventory loss from non-reusable calibration samples.
Data Synthesis, Retraining & Active Learning Loop
The transition from initial deployment to production maturity follows an automated closed-loop machine learning cycle. Edge inferences generate telemetry that directly reinforces model accuracy.
End-to-End Pipeline Iteration Architecture
Method 4 Cold-Start canonical samples
Live store deployment & cashier assist mode
Unconstrained real customer plates during rush
Automated threshold trigger: Any edge plate instance producing model classification confidence < 70% or triggering a manual cashier item override is tagged as high-priority and routed to cloud human-in-the-loop relabeling queue within 60 seconds.
The dataset synthesis pipeline uses algorithmic mosaic stitching and copy-paste augmentation to oversample under-represented low-velocity SKUs (e.g., Sambal Petai) against runaway core SKUs (e.g., Nasi Putih), preventing neural overfitting.
Retrained weights compiled to ONNX/TensorRT runtimes and dispatched via OTA firmware delta update at 03:00 AM daily. Production accuracy locks in at >98.4% Top-1 reliability with sub-850ms checkout speeds.
Quick Reference Checklists, Legal Indemnity & Contract Clauses
Operational deployment checklists and the exact binding legal clauses executed between Bibora Technologies Inc. and F&B retail merchant partners.
Pre-Deployment On-Site Operational Checklist
Pre-Flight SOP v2.414.8.1 Spoilage Recognition & Concession Scope. Bibora and Client acknowledge that achieving production-grade optical accuracy for high-variance mixed rice menus requires staging authentic, overlapping dish configurations and heavy gravy pairings. Due to food hygiene standards, cross-sauce blending, and ambient temperature limits, dishes plated during the Phase 1 Method 4 Staging Session cannot be returned to service and must be written off as unconsumable inventory. In consideration of Client absorbing this raw ingredient spoilage, Bibora shall grant an upfront rebate from twenty percent (20.0%) up to thirty percent (30.0%) against the standard one-time AI Vision Model Training Fee specified in Schedule A.
14.8.2 Tiered Spoilage Compensation Scale. The discount rate applied to the AI Vision Model Training Fee invoice is determined by sampling complexity:
- Baseline Spoilage Route (20.0% Discount): Applied where Client prepares 15–20 standard menu permutations and absorbs the food write-off without additional dedicated personnel.
- Premium Ingredient Coverage (25.0% Discount): Applied where calibration includes high-cost Tier 1 and Tier 2 proteins (e.g., Ketam Masak Lemak, Udang Galah, Sotong Sambal, whole fried fish) written off during staging.
- Dedicated Kitchen Plating Assistance (30.0% Discount): Applied where Client provides one (1) dedicated cook to execute directed gravy layering, portion variances, and dish adjustments alongside Bibora’s optical technician for a minimum continuous window of two (2) hours.
14.8.3 Settlement & Absolute Ban on Re-Serving. Concession discounts shall be credited directly upon Phase 1 sign-off. As a strict condition precedent to receiving the rebate, Client covenants that all staged items shall be permanently discarded or consumed internally by authorized staff pursuant to health regulations, and under no circumstances re-introduced to consumer buffets.
19.3.1 Sole & Exclusive Merchant Responsibility. The Client acknowledges and agrees that it retains full, sole, and unassignable legal and operational responsibility for the procurement, preparation, holding temperature, hygiene, allergen disclosure, and distribution of all food items utilized during the data collection process. Client covenants that all activities strictly adhere to the Food Hygiene Regulations 2009 (KKM), local municipal public health bylaws, and any applicable religious dietary certifications (including Halal certification standards).
19.3.2 Non-Contact Advisory Status of Bibora. Bibora Technologies Inc., its employees, agents, officers, and contractors act exclusively as technical optical and data engineering specialists. Bibora personnel are contractually barred from physically handling, plating, ladling, or disposing of consumable items. Under no circumstances shall Bibora be deemed a food handler, food business operator, or caterer under applicable law.
19.3.3 Comprehensive Indemnification & Waiver. The Client hereby unconditionally agrees to defend, indemnify, and hold harmless Bibora Technologies Inc., its affiliates, directors, officers, employees, and technological partners from and against any and all claims, liabilities, losses, damages, penalties, fines, costs, and expenses (including full legal fees on a solicitor-client basis) arising directly or indirectly out of:
- (a) Any food-borne illness, allergic reaction, contamination, bodily injury, or death allegedly resulting from food prepared, held, or sampled on Client premises;
- (b) Any citation, regulatory sanction, temporary closure, or fine imposed by municipal or health authorities relating to store hygiene, waste disposal, or cross-contamination;
- (c) Any breach of Halal integrity protocols occurring on the Client's food preparation or service lines before, during, or after optical data acquisition sessions.
THIS INDEMNIFICATION OBLIGATION SURVIVES THE TERMINATION OR EXPIRATION OF THE MASTER SERVICES AGREEMENT AND BINDS CLIENT SUCCESSORS AND ASSIGNS IN PERPETUITY.