InnerSpark Workforce AI

AI-native workforce operations for SMBs: Gemini Function Calling, biometric attendance, mobile check-ins and deterministic pre-payroll.

Comment

InnerSpark Workforce AI

Inspiration

Small and medium businesses still manage attendance through disconnected biometric terminals, spreadsheets, chat messages and manual supervisor review. Physical clock-ins, remote check-ins, attendance exceptions and payroll preparation become separate workflows. Managers spend time finding records rather than resolving the exceptions that actually require judgment.

We created InnerSpark Workforce AI during the Build with Gemini XPRIZE period to give SMBs an affordable AI-native workforce layer connecting real attendance events, mobile check-ins, employee information and deterministic pre-payroll logic.

Our goal was not to add a chatbot to an HR dashboard. We designed Gemini as an auditable agentic orchestrator that understands workforce requests, selects authorized enterprise functions and operates on real business data while strict server-side controls determine what the AI is allowed to access.

What it does

InnerSpark Workforce AI is a multi-tenant B2B workforce platform combining biometric attendance, secure mobile check-ins, attendance analytics, pre-payroll workflows and a Gemini-powered Workforce Agent.

Employees can record attendance through compatible biometric systems or a mobile workflow that captures a real-time photograph and GPS coordinates. Mobile evidence is stored in private Google Cloud Storage. Authorized administrators can review location evidence, including reverse-geocoded addresses.

Attendance events are processed deterministically into workforce records and exceptions such as late arrivals, early departures and overtime.

The Gemini Workforce Agent uses Function Calling to translate natural-language management requests into authorized business operations. Gemini can select tools for employee lookup, attendance/mobile check-in inspection and deterministic pre-payroll analysis. The application executes the underlying function and returns structured results for Gemini to interpret and explain.

Sensitive calculations remain deterministic application logic rather than free-form LLM arithmetic.

How we built it

The product is built with Next.js and TypeScript and deployed on Google Cloud. Firestore provides the multi-tenant operational data layer and private Google Cloud Storage stores sensitive mobile check-in evidence.

The AI layer uses the official @google/genai SDK with Gemini 2.5 Flash through a server-side /api/agent route and Function Calling.

A core architectural principle is separation between probabilistic reasoning and deterministic business logic. Gemini interprets intent and selects an authorized capability. Attendance and payroll-sensitive calculations execute as deterministic application functions, after which Gemini explains the structured result.

Zero-trust tenant isolation

The application does not trust an arbitrary companyId supplied by the browser. Server-side authentication derives the user's identity, role and tenant membership from the authenticated session before Gemini tools or Firestore queries can execute. This prevents prompts from becoming a shortcut around tenant authorization.

Auditability

Gemini tool activity is represented through structured gemini_function_call logging so model/tool execution can be inspected without exposing secrets or employee PII.

Deterministic operational data

Production workforce calculations no longer depend on random dummy generators. Attendance and novelty logic uses deterministic rules driven by biometric/mobile attendance records, with explicit America/Guayaquil (UTC-5) timezone handling.

Mobile evidence

The mobile check-in flow captures a real-time photograph and GPS latitude/longitude. Evidence is stored privately in Google Cloud Storage, while Nominatim/OpenStreetMap reverse geocoding can turn coordinates into human-readable locations for authorized HR review.

Privacy-safe judging environment

Production PII and hardcoded secrets are excluded from the public repository. An isolated XPRIZE sandbox can be populated with realistic synthetic profiles through seed_innerspark_labs.mjs, allowing evaluators to exercise the application without accessing customer employee data.

Challenges we ran into

The hardest challenge was making the product genuinely AI-native without allowing an LLM to become the source of truth for sensitive HR calculations.

Payroll and attendance are unforgiving domains for hallucinations. We therefore moved numerical and policy-sensitive calculations into deterministic functions while preserving Gemini for intent understanding, tool selection, contextual reasoning and explanation.

Multi-tenancy created a second challenge. Prompt instructions are not security boundaries, so tenant authorization had to be enforced server-side before any Gemini tool or Firestore operation executes.

Real attendance workflows also exposed issues synthetic demos hide: timezone consistency, biometric event normalization, GPS evidence and incomplete attendance sequences. These constraints shaped the production architecture rather than being hidden behind mock data.

Finally, judges need to evaluate a real commercial system without seeing real employee PII. The isolated synthetic sandbox addresses that conflict.

Accomplishments

We are proud that InnerSpark Workforce AI became a commercial product rather than a hackathon-only demonstration:

  • Gemini Function Calling connected to actual workforce capabilities rather than a standalone chatbot.
  • Deterministic attendance and pre-payroll logic.
  • Server-side tenant isolation before AI tools or database queries execute.
  • Mobile attendance with photographic and GPS evidence.
  • Integration of biometric and mobile attendance records.
  • Structured AI execution logging for auditability.
  • Privacy-safe XPRIZE sandbox using synthetic users rather than production PII.
  • Removal of random/mock workforce calculations from production logic.
  • Deployment on Google Cloud infrastructure.
  • A multi-tenant SaaS architecture designed for real organizations.

The central design principle is simple: Gemini reasons. Deterministic systems calculate. Zero-trust authorization decides what the AI is allowed to see and do.

Real business traction

IA Pro / IAPRO S.A.S. became the first external paying B2B customer for this new Workforce AI offering during the hackathon. PC Doctor invoiced USD 300 for development and implementation plus USD 100 for the first monthly cloud workforce service, or USD 400 before VAT. The corresponding USD 460 invoice including VAT was paid during the hackathon period.

PC Doctor also provides an internal validation environment. FEMAR S.A. is an active commercial opportunity combining physical access-control modernization with a complementary Workforce AI software/service proposal; FEMAR is pipeline and is not counted as revenue.

Business model

The model is B2B. Customers pay an initial implementation/configuration fee followed by recurring monthly service for cloud operation, maintenance, support and access to the workforce platform. Hardware, access-control deployments, integrations and additional modules can be quoted separately.

Customer acquisition during the hackathon has been direct and referral-based, with zero paid marketing spend.

AI-native operations

AI is used both inside the product and in how the business operates. Gemini is the production reasoning/orchestration layer. Google Antigravity and Gemini have accelerated implementation and technical iteration. ChatGPT Business has supported architecture review, orchestration, documentation and submission operations, while Natively AI was used during early frontend prototyping with promotional credits.

Development during the hackathon was founder-led and AI-accelerated, with no external engineering payroll.

Google Cloud

Google Cloud provides the production foundation: Cloud Run for serverless deployment, Firestore for workforce data, private Cloud Storage for mobile evidence and Gemini API for the AI reasoning layer.

Pre-existing resources disclosure

The project uses pre-existing PC Doctor legal/tax and billing infrastructure, prior IT/access-control expertise, referral and customer-introduction relationships, reusable testing hardware and general-purpose company accounts/tooling. These resources accelerated validation and commercialization. InnerSpark Workforce AI itself, including the SaaS product, Gemini agent workflow, packaging, pricing and recurring service offering, was created during the XPRIZE competition period.

Impact: Small Business Services

Our thesis is that SMBs should have access to secure workforce automation and useful AI without enterprise-scale HR software budgets. We measure progress through paying customers, recurring revenue, active organizations and employees, attendance events processed, Gemini Function Calls, mobile check-ins, exceptions detected, time saved in manual review and expansion into additional workforce operations.

InnerSpark Workforce AI connects the physical world of attendance devices and mobile work to Gemini-powered business reasoning while keeping sensitive calculations, authorization and final employment decisions controlled and auditable.

Judge Access / Testing Credentials

Production URL: https://workforce.pcdoctor.ai XPRIZE Judge Account Username: DEVPOST-JUDGE Password: Admin123!

Built With

Share this project:

Updates