NexH.AI

Landing Page
Main hub
Daily Tatical Action
Weekly Insight
Vision_analysis
Data entry with ocr and csv function

Inspiration

Every day, small business owners in service industries face the same painful challenge: customer churn happens silently. A loyal customer who visited every month suddenly disappears, and by the time anyone notices, they've already found a competitor.

Small and Medium Enterprises (SMEs) are the lifeblood of the global economy, but they are often paralyzed by "Data Gravity"—the sheer weight of manual entry and fragmented records. We were inspired by a simple but powerful realization: Business owners don’t need more data; they need solutions. Our mission was to leverage Gemini 3.0 to transform raw business clutter into a high-definition decision pipeline, turning "what happened" into "what to do now".

The question that sparked NexH.AI was simple:

"In the era of Agentic AI, what if AI could be the manager that never sleeps - one that watches over every customer relationship and tells you exactly who to call today?"

What it does

NexH.AI is a Strategic Business Hub that serves as an impartial AI consultant for service-based industries.

Multimodal Logic Extraction : Goes beyond simple object recognition—uses Gemini's vision capabilities to extract structured business logic from unstructured physical data (receipts, handwritten forms, condition photos), automatically updating SQL databases and triggering industry-specific workflows. Vision isn't the endpoint; it's the bridge from analog chaos to digital action.

Automated Data Mapping : Leverages Gemini’s multimodal intelligence to scan documents and auto-fill complex industry workflows, eliminating manual friction.

Daily Intelligence : Pinpoints the "Top" highest-value actions every day, generating personalized care scripts ready for one-click sharing.

Weekly Strategic Audits : Benchmarks performance against industry standards and identifies cross-state emerging trends.

Core Capabilities

Feature	Description	Why It's Not Simple Vision
Multimodal Business Logic Extraction	Photograph a handwritten service form → AI extracts structured data → Updates `UniversalAsset` table in SQL → Triggers next maintenance alert	Uses vision to populate databases and trigger workflows, not just identify objects
Document-to-Database Pipeline	Snap a receipt/invoice → OCR extracts date, items, amount → Creates timestamped business records → Feeds into churn prediction model	Vision is the data entry automation layer, feeding into complex business intelligence
Condition Analysis → Action Scheduling	Photo of customer condition (skin/vehicle/equipment) → AI assesses severity + timing → Generates follow-up schedule + personalized message → Updates CRM status	Vision triggers multi-step business processes, not just returns labels
Sub-Second Intelligence	Gemini 3.0 Flash completes full analysis (OCR + logic extraction + JSON structuring) in under 10 seconds. For busy shop owners juggling customers, 30-second waits are dealbreakers. This speed makes AI feel like an instant teammate, not a batch-processing tool.	Performance enables real-time business process integration—owners complete customer check-ins with live AI assistance during service, not post-facto offline analysis
Fatigue-Aware Outreach	Prevents over-contacting customers with a specific cooldown algorithm tracked in `AIActionLog` table	AI manages temporal business rules across customer lifecycle

The Intelligence Formula

Our priority customer detection uses a two-layer filtering system. The formula adapts to each industry - below is an example for a Beauty Salon:

$$ \text{Focus List} = \text{Strategy Candidates} - \text{Fatigue Set} - \text{Excluded Status} $$

Where:

Strategy Candidates: Customers matching industry-specific rules (e.g., $ \text{days_absent} > 60 $ for salons)
Fatigue Set: Customers contacted within the cooldown period
Result: Top priority customers per day

Industry Adaptability: An auto workshop might trigger at 90 days with mileage thresholds, while a F&B outlet might focus on loyalty program engagement. The core algorithm remains the same, but parameters are tailored to each vertical

How we built it

Architecture

Flutter web (Front end) > Cloud Run API + Cloud SQL > Gemini 3.0 Flash

Tech Stack

Frontend: Flutter Web with responsive design
Backend: Python FastAPI on Cloud Run (auto-scaling)
AI Engine: Vertex AI with Gemini 3.0 Flash Preview
Database: Cloud SQL PostgreSQL with JSONB for flexible schemas
Auth: Firebase Authentication with custom claims
Storage: Google Cloud Storage for image processing

Key Design Decisions

Multi-tenant Row-Level Security: Every query is automatically scoped by tenant_id extracted from JWT claims - not request body (security first!)

Industry-Adaptive Prompts: We use the CO-STAR framework to dynamically inject industry context:

[Context] Beauty Salon specializing in: Acne Treatment, Anti-aging [Objective] Analyze skin condition from image [Style] Professional dermatology terminology

Image Compression Pipeline: All uploaded images are compressed to 800px max dimension before AI processing - reducing costs by 60% while maintaining accuracy.
Flutter for Cross-Platform Consistency: Built with Flutter to ensure seamless experience across all devices (mobile, tablet, desktop). SME owners need to access AI insights anywhere—on a phone between appointments, on a tablet at the counter, or on desktop during end-of-day reviews. Single codebase, universal experience.

Challenges we ran into

1. The "List vs Dict" Config Migration Crisis

Mid-development, we migrated our tenant config schema from dictionary to list format. This broke OCR functionality with:

ERROR: 'list' object has no attribute 'get'

Solution: Created a centralized get_tenant_schema_rules() function that handles both formats gracefully.

2. Customer Fatigue Prevention

Early testing showed the AI was recommending the same "high-risk" customers every single day, leading to annoying over-contact. Solution: Implemented a 30-day fatigue filter using AIActionLog tracking:

fatigued_ids = SELECT target_id FROM ai_action_logs WHERE created_at >= (TODAY - 30 days)

3. Cold Start Performance

Cloud Run cold starts were causing 8-10 second delays on first request. Solution: Configured minimum instances and optimized container image size.

4. The Complexity of Domain Specificity

Our greatest challenge was the ambition to span diverse fields—from beauty salons to automotive detailing. We quickly realized that there is no "one-size-fits-all" solution. Different industries, geographic regions, and customer demographics operate on entirely different logics. To provide truly actionable advice, we had to perform intensive research into each sector, as understanding the subtle nuances of a specific industry is the only way to define meaningful AI requirements.

Accomplishments that we're proud of

Zero-Config Onboarding: New tenants can start using AI features immediately with sensible industry defaults - no complex setup required
Sub-10-Second AI Analysis: From photo upload to structured JSON response in under 10 seconds, including image compression and Gemini processing
Production-Ready Security:
- Firebase token verification on every request
- 64-character cryptographic secrets for internal endpoints
- Rate limiting (30 req/min on AI endpoints)
- Per-tenant daily quotas
Cost-Optimized AI Pipeline: Smart compression pipeline delivers enterprise-grade AI analysis at startup-friendly costs

What we learned

Technical Insights

Gemini 3.0 Flash is remarkably good at structured extraction - With proper prompting, it returns clean JSON even from messy handwritten notes
AI works best as an accelerator, not a replacement - The most effective AI doesn't try to replace human expertise—it amplifies it. By positioning AI as a "smart assistant" rather than an autonomous decision-maker, we saw significantly higher trust and adoption among traditional business owners.
"AI Proposes, Expert Approves" drives adoption - SMEs are far more likely to embrace high-tech solutions when they retain final creative control over their professional voice. Our AI generates personalized message scripts, but the business owner always decides when and how to reach out.

Product Insights

Small business owners don't want dashboards - they want answers - "Who should I call today?" beats fancy analytics every time
AI confidence matters - Showing "85% confidence" alongside recommendations builds trust
Contact fatigue is real, but the threshold varies by industry - Over-contacting customers damages relationships, but the ideal outreach cycle differs dramatically: a beauty salon might follow up monthly, while an auto workshop operates on quarterly or annual service intervals. Smart AI must adapt to each vertical's natural rhythm.

What's next for NexH.AI

Validate — Launch a beta for interested users to gather real-world feedback and refine the product-market fit.

Deepen — Enhance our cross-state trend analysis ("Federated Intelligence") while maintaining strict tenant data isolation.

Expand — Extend to more service verticals, so business owners across industries can focus on delivering quality service while NexH.AI handles strategy, decisions, and customer relationships.