Mapspring — The proving ground
Real product constraints make the learning credible.
Multi-tenant data, APIs, identity, release architecture, public and private states, security boundaries, observability, and recovery all have to work together.
Product · Proof in Practice
Operational productMapspring
A hosted platform for managing and publishing physical-location experiences.
Mapspring manages branches, ATMs, offices, and other physical assets. Within AI Innovation Studio, it provides a realistic product environment where AI architecture, integration, engineering, security, governance, and operating practices can be applied and validated.
Switchkit — The intelligent workflow
A bounded path from source evidence to private locator.
Switchkit analyzes an authorized source, normalizes useful fields, preserves traceability, asks a person to review the result, and creates a private locator before publication is considered.
Target Switchkit lifecycle
Set Up → Analyze → Import & Review → Build → Approve → Publish
The workflow combines automation with explicit state, human review, and a separate publication boundary.
- 01
Set Up
- 02
Analyze
- 03
Import & Review
- 04
Build
- 05
Approve
- 06
Publish
Controlled validation
From uncertainty to governed execution
A bounded AI agent resolved an integration that deterministic discovery could not. Deterministic processing then resumed and produced a reconciled private application outcome.
- 01
Deterministic Discovery
- 02
Bounded AI Reasoning
- 03
Deterministic Processing
- 04
Human Governance
- 05
Private Application Outcome
Controlled synthetic E2E
8 source → 8 prepared → 8 reviewed
1 private locator → 8 locations → 8 visible markers
Accepted real-source evaluation
21 source → 21 normalized → 21 private locations
- 21/21 materialized
- 0 missing
- 0 unexpected
- 0 duplicates
- 0 further model calls
- 0 geocoding calls
Applied-AI architecture lesson
Deterministic first. Model escalation when needed.
Deterministic discovery initially stopped at an unfamiliar integration. A bounded AI agent identified the useful source; deterministic retrieval, normalization, governed private execution, and reconciliation then completed all 21 locations without another model call.
The emerging hypothesis is that validated AI discoveries can become reusable deterministic capability. That learning loop is promising, but it requires replication before it can be treated as proven.
Visual evidence
Three views of the review-before-publish experience
All public evidence uses the fictional Mountain Peak Credit Union fixture.
Production status
Production-capable by architecture, production-disabled by policy.
Mapspring production infrastructure exists. Switchkit is validated end to end in a controlled environment. Production execution architecture is being hardened for bounded enablement; production Switchkit execution remains disabled and fail-closed. No production customer migration has occurred.
Product roadmap → Intelligent platform
From hosted platform to governed agentic capability
The roadmap distinguishes what has been validated from what remains bounded or planned.
- 01Validated
Foundation
Secure platform and API foundations
- 02Validated
Integration
Bounded access to data and systems
- 03Validated
Intelligent Workflows
Deterministic work with model escalation
- 04Demonstrating
Governed Agents
Tools, authority, and human control
- 05Hardening
Enterprise Scale
Progressive, governed production enablement
Reusable enterprise pattern
The architecture is portable across organizations and industries.
Working proving ground
Mapspring and Switchkit exercise and validate portions of the pattern.