Manual audits are subjective — scores don't reflect reality
The core problem isn't volume. It's that every patroller applies their own standard. There is no single, objective measure of rider compliance. The data below proves it.
Current System — Reported Appearance Score
~95%
Average appearance compliance across all markets in the current system. On paper, this looks excellent.
The numbers say ~95% compliance. The photos tell a different story.
Almost every market shows appearance compliance around 95%. This looks like near-perfect rider quality. But when you look at the actual rider photos from those same checks, the reality is very different.
REAL EXAMPLES — MARKED "PASS" BY PATROLLERS
Rider wearing completely unbranded jacket
No Talabat branding visible. Blue jacket, no uniform underneath. This rider should fail appearance — but was marked as compliant.
Check log: Appearance passes = Yes · Score = 1 (highest) · Every field "Yes"
Rider wearing a non-Talabat black hat
Wearing a non-Talabat black hat — but was marked as compliant.
Check log: Appearance passes = Yes · Score = 4 (highest) · Every field "Yes"
What AI solves
AI sets a single, objective standard. Every rider in every market is evaluated by the same criteria, the same way, every time. There is no subjectivity, no variance between patrollers, and no way to click through "Yes" without a real assessment.
BEFORE & AFTER
Side by side — what changes for the patroller
Left: how it works today (manual input for every item). Right: how it works with SmartPatrol AI (AI pre-fills, patroller reviews).
❌ BEFORE — Manual Input (Current System)
Patroller manually selects condition for EVERY gear item. Each item opens a dropdown with options: Good, Damaged, Missing, Unbranded, Old branding, Dirty. 13+ items, one by one. Subjective — two patrollers can score the same rider differently.
✅ AFTER — AI Pre-Filled (SmartPatrol AI)
AI analyzes the photo and pre-fills every item automatically. Each gear item shows the AI result (Good, Dirty, Missing, Unbranded). Patroller only taps "Disagree" if the AI got something wrong. No manual selection needed — just review and submit.
The key difference
Before: the patroller decides every condition — subjective, inconsistent, gameable. After: the AI decides conditions, the patroller validates — objective, standardized, evidence-backed. Within each market, every rider is scored against the same standardized knowledge base — eliminating patroller-to-patroller variance.
CURRENT SYSTEM — END TO END
Today's patroller journey — examples from the field
The current system is divided into 3 sections (ID Check, Appearance, Behaviour + Vehicle + Hygiene) with 21+ question pages. Patrollers navigate with "Back" and "Next" buttons. If the network drops between pages, all progress is lost. Below are examples from the actual journey.
1. Rider Search
2. Rider Profile
3. ID + Plate
4. Gear (Manual)
5. Condition Select
6. Expired Gear
7. Hygiene
8. Behaviour
9. Comments
10. Photo (at END)
Problem: Multi-page navigation
21 questions across multiple pages. Each page requires "Next". Network drop between pages = lost work. Patrollers have to start over.
Problem: Photo at the END
Confirmation photo is currently the last step. AI needs the photo FIRST to analyze gear. Photo moves to the start.
Problem: 3rd party stamp apps
Some markets use external apps to add timestamps/GPS to photos. Major security risk — photos can come from gallery, not live. See security section for details.
WITH AI — INTERACTIVE PROTOTYPE
Try the new patroller flow
A working prototype of how SmartPatrol AI will look on the patroller's phone. The questions, structure and data flow stay the same as today — only the appearance section becomes AI-automated. Click through on the right to see exactly how it works.
→ THIS IS INTERACTIVE
Use the phone on the right. Follow the steps below. Click through — it actually works with any rider ID you enter.
1
Find the rider
The patroller stops a rider in the field and types the rider ID on the keypad. Same as today, no change.
Tap the keypad to enter any rider ID, then tap "Search →"
2
Review rider profile
The system loads the rider's history, gear list, and any expired items. Same screen as today.
Tap "Start check" to begin the inspection
3
Take the rider photo (NEW: photo first)
The form starts with the photo, not at the end like today. Live camera only — no gallery. Date, time and GPS are stamped natively by the app — no third-party stamp apps needed.
Tap "Take Photo" — see the watermark stamp appear automatically
4
AI analyzes in seconds
As soon as the photo is taken, the AI begins processing. Within ~10 seconds the appearance section is fully pre-filled with PASS/FAIL for every gear item.
Watch the spinner and the form expand below the photo
5
Scroll down — one form, top to bottom
The whole inspection is a single scrollable form. No multi-page navigation, no Back/Next buttons, no risk of losing work to network errors.
Scroll inside the phone to move through the sections
6
Disagree with the AI (the patroller's main job)
In the Appearance section, every gear item has an AI result already filled in. The patroller's only action is to disagree if the AI got something wrong. Click "Disagree" on any item — a bottom sheet opens with the same condition options as today.
Try "Disagree" on any item, pick a different condition, then Save
7
Hygiene — AI + manual hybrid
Visible dirtiness (uniform stains, hair, beard, nails) is detected by the AI from the photo. Smell-based hygiene stays a manual patroller assessment.
Scroll to the Hygiene section to see both parts
8
Submit once — done
Single submit at the end. All data saves at once. The footer shows live count of overrides logged.
Tap "Submit Check" at the bottom
✓
No disruption to existing flow
Same questions as today. Same data captured. Same OMS records. Same scorecards. Same downstream reporting. The only thing that changes is how the appearance section gets filled in — by AI instead of manually.
AI CAPABILITIES
What the AI analyzes
The AI evaluates a single rider photo and assesses each item independently. The list below shows what is automated and what stays as manual patroller input.
AI ANALYZES AUTOMATICALLY
👕
T-Shirt
Branding, condition
👖
Trousers
Branding, color
👟
Shoes
Open vs closed
🧥
Jacket
If worn: branding
🧢
Cap
If worn: branding
🛡️
Safety Gear
Vest, reflective
🔍
Windshield
Presence, damage
🧼
Hygiene
If visible: hair, nails
🎒
Bag
If visible
🦵
Knee/Elbow
If visible
⛑️
Helmet
If visible
STAYS MANUAL (PATROLLER INPUT)
👃
Odor / Smell
🗣️
Politeness
END-TO-END PROCESS
New patroller journey — photo first, one form, AI-assisted
PATROLLER IN FIELD
1
Search rider & verify ID
Same as today — search by rider ID, confirm identity, check for fake ID.
2
Take live photo (MOVED TO START)
Camera only — gallery blocked. Date, time, GPS auto-stamped by the app. No third-party stamp apps needed.
3
AI analyzes in under 10 seconds
Appearance section auto-filled: each gear item shows condition (Good/Dirty/Missing/Unbranded/Old branding). AI validates photo quality — rejects blurry, dark, or non-rider photos.
4
Patroller reviews AI results
Scrolls through single form. Taps "Disagree" on any item where AI is wrong. Everything else accepted automatically.
5
Complete manual sections
Hygiene (odor/smell), politeness, comments — same as today. All on one scrollable form.
6
Submit — one tap
All data saves at once. No multi-page risk. Results + photo sync to OMS.
SUPERVISOR REVIEW
7
Supervisor views all checks in dashboard
Every check, violation, override visible — per patroller, per market.
8
Supervisor agrees or disagrees on flagged cases
Only supervisor-confirmed disagreements count as actual AI inaccuracy. This is the true accuracy metric.
CLEAR SUMMARY
Exactly what changes and what stays the same
What changes
Photo moves to the start — taken right after ID check, not at the end
Appearance section is AI-automated — gear conditions pre-filled by AI, patroller reviews
Hygiene partially AI — visible dirtiness detected by AI, odor remains manual
One scrollable form — replaces multi-page setup, no more data loss from network errors
Camera only, no gallery — live photos enforced, gallery access removed
Market Knowledge Base — the foundation of AI accuracy
Every market must submit a compliance Knowledge Base that defines what "compliant" looks like in that market. The AI is trained on this base — its accuracy is a direct reflection of what each market provides. Submission and Boolanga + Talabat regional approval is required before pilot starts in that market.
Knowledge Base structure — what each market provides per gear item
1. Photos of correct kit — front + back per gear item 2. Violation example photos per item (the more, the better) 3. Strictness level per item (zero tolerance vs. warning) 4. Market-specific exceptions (seasonal gear, regional variants, regulatory stickers) 5. POC sign-off before submission to Boolanga 6. Approval by Talabat regional + Boolanga before AI training starts
Patroller alignment is non-negotiable
If the AI and patrollers judge by different rules, disagreements will be constant — distorting accuracy metrics and frustrating the field. Both must be aligned on the same Knowledge Base before pilot. Patroller briefings are mandatory before any market goes live.
FOR IT & SECURITY
Technical architecture
Runs on a dedicated local GPU. No images sent to third-party AI providers.
📱
Patroller
Browser (PWA)
→
🖥️
Boolanga Server
FastAPI + PostgreSQL
→
🧠
Dedicated GPU
Local AI Engine
🔒
Zero third-party exposure
Images stay in Boolanga infra.
⚡
4-7 sec processing
Dedicated GPU, consistent.
🔄
Auto fallback
Backup if GPU offline.
🛡️
Encrypted E2E
TLS 1.2+, at-rest, audit logs.
DATA, PRIVACY & SECURITY
Built local. No external APIs. No third-party exposure.
SmartPatrol AI runs entirely on Boolanga's own infrastructure. There are no external API calls, no cloud AI providers, no third-party photo processing. Every check stays inside our environment from capture to submission.
⚠️
CRITICAL ISSUE TODAY: Third-party stamp apps are a major security risk
In several markets today, patrollers use external third-party apps to add timestamp and GPS location to rider photos before uploading them. These apps:
• Process Talabat rider photos through unknown external systems
• Allow photos to be selected from the gallery — meaning old or staged photos can be submitted
• Have no audit trail or oversight from Talabat
• Create a clear data exposure path that bypasses Talabat's security policies
SmartPatrol AI eliminates this completely. Live camera capture is enforced. Date, time and GPS are stamped natively by the app. No third-party tools are needed or allowed.
100% LOCAL
No external APIs
No calls to OpenAI, Anthropic, Google or any external AI provider. The AI model runs on Boolanga's dedicated GPU.
DEDICATED INFRA
Not shared cloud
A dedicated GPU server is reserved for Talabat workloads. Not multi-tenant, not shared with other clients.
EPHEMERAL
In-memory only
Photos are processed in memory and wiped after analysis. No persistent storage of images on the AI server.
ENCRYPTED
End-to-end
TLS 1.2+ in transit, encrypted at rest, full audit logs. Same standards as the existing OMS.
AUDITABLE
Talabat security access
Talabat security team has periodic review access to the infrastructure. External security audits are conducted regularly.
CAMERA-ONLY
Gallery disabled
Live camera capture is enforced at the device level. Patrollers cannot upload from gallery, screenshots, or any saved photos.
A structured rollout with explicit accountability at each phase. Boolanga, market leadership, and patrollers all have defined responsibilities. Weekly touchbases keep performance and accuracy issues addressed in real time.
PHASE 1
2 WEEKS
Development
Boolanga IT builds the new flow: photo-first capture, native GPS/timestamp, single scrollable form, AI integration into the appearance section, supervisor audit queue in OMS, dashboard metrics. No external API dependencies — all on local GPU.
BOOLANGA
PHASE 2
PARALLEL
Knowledge base submission & approval
Each market submits its compliance knowledge base: photos of correct kit (front + back per item), violation examples (unbranded, dirty, old branding, damaged, faded), strictness levels, market-specific exceptions, and POC sign-off. Boolanga + Talabat regional review and approve before AI training begins.
⚠ Critical: AI quality depends entirely on the quality of this knowledge base. No approval = no pilot start in that market.
MARKET POCs
PHASE 3
1 WEEK
Pilot coordination & planning per market
Per-market kick-off: confirm participating patrollers and supervisors, schedule patroller briefings on AI parameters and what stays manual, align on local guidelines, distribute the SmartPatrol AI link, finalize reporting cadence with regional leads.
REGIONAL + BOOLANGA
PHASE 4
~4 WEEKS
Live pilot — weekly touchbases
Patrollers use SmartPatrol AI in the field. Each market runs a weekly performance review jointly with Boolanga: review accuracy data, address inaccuracies flagged by supervisors, retrain AI on edge cases, monitor patroller adoption. Issues identified one week are resolved before the next.
ALL TEAMS
PHASE 5
ONGOING
Supervisor accuracy audit
When a patroller marks an AI result as inaccurate (clicks Disagree), the check enters the supervisor audit queue. Supervisors review and either confirm the AI was wrong (counted as true inaccuracy) or confirm the patroller was wrong (logged as incorrect override). Only supervisor-confirmed inaccuracies count against AI accuracy metrics.
SUPERVISORS
PHASE 6
END OF PILOT
Joint feedback & go/no-go decision
A structured feedback form — designed jointly by Talabat regional leadership and Boolanga — is distributed to all participating patrollers and supervisors. Combined with hard pilot data (accuracy, override rate, check time), this informs the go/no-go decision per market and shapes the production rollout plan.
REGIONAL + BOOLANGA
Pilot data isolation
All pilot data lives in a separate dashboard, ring-fenced from production OMS reporting. If the pilot doesn't meet targets in any market, the data retires cleanly with zero impact on existing compliance reports or scorecards.
SUCCESS CRITERIA
What we're measuring
Accuracy is the #1 metric. Average check time is #2. Volume is not a target — it's a natural outcome of faster checks.
🎯
AI accuracy (supervisor-confirmed)
≥ 90%
💬
Patroller satisfaction (feedback survey)
Positive trend
🔄
Override rate (patroller disagreements)
Track & optimize
🗣️
Feedback sessions with each market team
Week 4
Note on volume
We are not setting a volume increase target. If average check time drops from 10-12 minutes to ~6 minutes, the math speaks for itself — markets can calculate their own expected volume increase based on patroller working hours. We provide the data, teams do the math.
ROLES DURING PILOT
Who does what
T-Patrollers
● Use SmartPatrol AI
● Confirm or override AI
● Complete feedback survey
Compliance / RQS
● Monitor pilot dashboard
● Review accuracy data
● Define standards per market
Regional Logistics
● Coordinate briefings
● Track check time data
● Report regional results
Supervisors
● Review all disagreements
● Confirm true inaccuracies
● Final audit sign-off
Boolanga
● AI training & updates
● Technical support
● Accuracy optimization
SUMMARY
One standard. One system. Every market. Every rider.
SmartPatrol AI replaces subjective assessments with objective, AI-verified compliance. Patrollers validate. Supervisors confirm. The data tells the truth.
