Unclassified

IntegrateWise - Universal Cognitive Operating System

Agent Context Document

Version: 1.0
Last Updated: February 2026
Classification: Confidential

1. Executive Summary

IntegrateWise is a Universal Cognitive Operating System — an AI-powered Knowledge Workspace that fundamentally changes how organizations work with their tools, data, and AI.

Core Philosophy: "AI That Thinks in Context, Waits for Approvals"

Tagline: Normalize Once, Render Anywhere | Enhance, Don't Replace

The Differentiator: First application since 1945 to complete the full loop of human approval and post-approval agent execution with audit rails and evidence backing.

2. The Core Loop

LOAD → NORMALIZE → STORE → THINK → REVIEW & APPROVE → ACT → REPEAT

The REVIEW & APPROVE (Govern) stage is a hard gate. Nothing executes without:

Human approval
Audit trail
Evidence backing

This is what makes IntegrateWise fundamentally different from every AI-native tool on the market.

3. Four-Layer Architecture (L0–L3)

Layer	Name	Purpose	Key Components
L0	Onboarding	Signup, context questions, tool connect, first data hydration	4 stages: Entry + AI Insights → AI Loader Demo → Context Selection → Tool Connect + First Hydration
L1	Context-Aware Workspace	15 modules rendering context-specific views of normalized data	Home, Projects, Accounts, Contacts, Meetings, Docs, Tasks, Calendar, Notes, Knowledge Space, Team, Pipeline, Risks, Expansion, Analytics
L2	Universal Cognitive Layer	14 cognitive components for reasoning, action, and governance	SpineUI, ContextUI, KnowledgeUI, Evidence, Signals, Think, Act, HITL, Govern, Adjust, Repeat, AuditUI, AgentConfig, DigitalTwin
L3	Universal Backend	The engine: 8-stage mandatory pipeline, 4 accelerators, 6 data stores	Pipeline stages, accelerators, stores, engines, dual-loop architecture, security model

Critical Rule: Every piece of data must flow through L3 before reaching L1 or L2. There are no shortcuts.

4. L0: Onboarding Flow

Stage	What Happens	Duration
1. Entry + AI Insights	Name + DOB input. AI personality analysis generates working style predictions, strengths, and growth areas. First "wow moment".	0–15 sec
2. AI Loader Demo	Natural language input demo. AI transforms unstructured input into structured actions across 4 tools simultaneously. Second "wow moment".	15–30 sec
3. Context Selection	Two workspace paths: Productivity Hub (OS) for ICs/PMs/Team Leads, or CS Platform for CSMs/AMs/CS Leaders. Same user can switch anytime. Signup (Google OAuth / email) happens here.	30–50 sec
4. Tool Connect + First Hydration	Connect 1–2 tools via OAuth or upload a data dump (CSV/JSON/PDF). Backfill sync runs paginated historical import through the full 8-stage pipeline. Onboarding Accelerator bootstraps the workspace.	50–120 sec

5. L1: Context-Aware Workspace (15 Modules)

The L1 workspace consists of 15 modules, each providing a context-aware projection of normalized data from the Spine.

Module	Purpose	Key Data Sources
Home	Personalized dashboard with signals, pending actions, and key metrics	All stores → aggregated view
Projects	Cross-tool project tracking with tasks, milestones, and ownership	Tasks, Accounts, Contacts
Accounts	Unified account view with health scores, risk signals, and engagement	Spine SSOT + Accelerators
Contacts	Normalized contact directory with engagement history and relationships	Spine + Context Store
Meetings	Calendar integration with AI-generated agendas and action items	Calendar tools + Think Engine
Docs	Document hub with unstructured content indexed and searchable	Context Store (embeddings)
Tasks	Cross-tool task aggregation with priority scoring and due dates	Spine + multiple tool sources
Calendar	Unified calendar view across all connected tools	Calendar tool connectors
Notes	AI-assisted note-taking with auto-linking to entities	Context Store + Spine links
Knowledge Space	Organizational knowledge base built from docs, chats, and interactions	Context Store + AI Chats Store
Team	Team performance, workload distribution, and collaboration metrics	Spine + Activity logs
Pipeline	Deal and opportunity tracking with velocity metrics	CRM connectors + Accelerators
Risks	Proactive risk detection with evidence-backed alerts	Think Engine + Signals component
Expansion	Upsell and growth opportunity identification	Accelerators + Spine analytics
Analytics	Cross-module analytics with growth-aligned reporting	All stores + Domain Accelerators

6. L2: Universal Cognitive Layer (14 Components)

The L2 cognitive layer is the reasoning and governance engine. Its 14 components are universal — they operate across all workspace contexts.

Component	Type	Purpose
SpineUI	Data Surface	Canonical data browser — view and navigate normalized SSOT records
ContextUI	Data Surface	Unstructured context viewer — documents, embeddings, related content
KnowledgeUI	Data Surface	Knowledge graph explorer — organizational knowledge from all three planes
Evidence	Intelligence	Evidence compiler — assembles proof backing every signal and recommendation
Signals	Intelligence	Signal detector — identifies patterns, anomalies, and changes across data
Think	Reasoning	Situation assessor — evaluates signals against goals and produces proposed actions
Act	Execution	Action executor — carries out approved actions with tool write-back
HITL	Governance	Human-in-the-loop approval gate — the hard stop before any AI action executes
Govern	Governance	Policy and compliance enforcement — ensures actions meet organizational rules
Adjust	Learning	Feedback loop — captures approval/rejection patterns to improve future proposals
Repeat	Learning	Continuous cycle driver — re-evaluates situations after actions complete
AuditUI	Compliance	Audit trail viewer — immutable log of all actions, approvals, and data changes
AgentConfig	Administration	Agent configuration — define, enable, tune, and monitor AI agents
DigitalTwin	Simulation	Workspace simulation — test proposed changes before committing to production

7. L3: Universal Backend Engine

7.1 Three Data Ingestion Planes

Plane	Data Type	Ingestion Path	Storage
Plane 1: Structured (Tool Data)	CRM deals, support tickets, billing events, calendar entries	Webhooks/Polling → Connector Service → 8-Stage Pipeline	Spine (Neon PostgreSQL)
Plane 2: Unstructured (Documents)	PDFs, CSVs, spreadsheets, emails, images	Upload/Watch → R2 raw storage → Knowledge Service → Chunk + Embed	Context Store (pgvector + R2)
Plane 3: AI Chats (MCP Sessions)	Conversations from Claude, ChatGPT, other AI providers	MCP Connectors → Session Summarization → Compress at Source	AI Chats Store

7.2 Eight-Stage Mandatory Pipeline

Every piece of data — regardless of source, type, or urgency — must pass through all 8 stages. There are no shortcuts.

Stage	Name	Processing	Failure Mode
1	Analyzer	Source detection, object type identification, metadata extraction, schema version check. Stamps payload with pipeline_id for traceability.	Unknown source or malformed payload → DLQ
2	Classifier	Category mapping to internal taxonomy, CTX enum assignment, priority scoring (critical/normal/batch), accelerator routing flags.	Unmappable object type → DLQ
3	Filter	Tenant-level and user-level data scope enforcement. Checks data against scope definitions set during L0 onboarding. PII detection flags.	Scope rule error → pass-through with warning
4	Refiner	Field-level transformations: date normalization (ISO 8601 UTC), currency conversion, text cleaning (HTML strip), entity resolution across tools.	Schema mapping failure → DLQ
5	Extractor	Relationship extraction (account→contacts, deal→line items), computed fields (deal velocity, sentiment), change delta calculation, embedding generation for documents.	Extraction failure → DLQ with partial results
6	Validator	Data integrity checks, deduplication (exact + fuzzy), business rule validation, cross-reference verification against existing Spine records.	Integrity violation → DLQ
7	Sanity Scan	AI-powered anomaly detection, outlier identification, test data detection. Uses rules + ML to quarantine suspicious data before it enters stores.	Anomaly detected → Quarantine for human review
8	Sectorizer	Final routing to appropriate stores (Spine, Context, AI Chats). Performs the dual-write. Emits events to Think Engine and triggers Accelerator recalculation.	Write failure → retry queue with backoff

7.3 Four Accelerator Systems

Accelerator	Purpose	Key Components
Extraction Accelerator	Preserve tool-native structure during ingestion. Captures formulas, automations, and relationships that would be lost in normalization.	Ingress Controller, Tool Puller, Object Graph Builder, Formula Capture, Automation Capture, External ID Resolver
Normalization Accelerator	Map extracted data to the SSOT canonical schema. Handles schema detection, field transformation, and relationship preservation.	Schema Detector, Schema Registry, Category Mapper, Field Transformer, Relationship Preserver, SSOT Builder
Onboarding Accelerator	L0-specific compute. Analyzes initial data to bootstrap the workspace with presets, templates, and scope definitions.	Context Analyzer, Preset/Template Selector, Data Scope Definition
Domain Accelerators (×6)	Post-normalization domain-specific intelligence. Each produces scored, goal-linked metrics.	Customer Health Score, Churn Prediction, Revenue Forecaster, Pipeline Velocity, Support Health, Marketing Attribution

7.4 Six Data Stores

Store	Technology	Purpose	Access Pattern
Spine (SSOT)	Neon PostgreSQL	Canonical normalized records. The single source of truth for all structured data.	Read/write via Pipeline + Act Engine
Context Store	pgvector on Neon + R2	Embeddings, document chunks, and metadata for unstructured content. Enables semantic search.	Write via Pipeline; read via Knowledge components
AI Chats Store	Neon PostgreSQL	Summarized AI chat sessions. Never stores full transcripts — only compressed summaries.	Write via MCP Connectors; read via KnowledgeUI
Audit Store	Neon (append-only)	Immutable log of every action, approval, data change, and system event.	Write-only from all services; read via AuditUI
Hot Memory	Cloudflare Durable Objects	Real-time state for active sessions, collaboration, and in-flight computations.	Read/write by Workers; auto-evicts to cold storage
Dead Letter Queue (DLQ)	Cloudflare Queue + Neon	Failed pipeline items with full context for debugging and replay.	Write from any pipeline stage; read by ops tooling

8. AI Agent System

Only One Cognitive Twin

The system operates with a single Cognitive Twin per tenant/workspace, not multiple specialized agents. This twin:

Maintains continuity across all interactions
Learns from all approved/rejected actions
Operates across all three data planes
Enforces governance policies uniformly

Key Principles:

Approval-only execution: Every agent proposal passes through the HITL gate
The Think Engine generates situations and proposed actions
The Govern layer enforces policy compliance
The HITL component presents proposals to the human approver
Only after explicit approval does the Act Engine execute
The Adjust component captures the approval or rejection to improve future proposals

9. Dual-Loop Architecture

IntegrateWise operates two reinforcing feedback loops that converge to create a self-improving system.

9.1 Loop A: Context-to-Truth (Human-Governed)

Context → Brainstorming → Action → Tool Write → Ingest → Normalize → Truth → View → Repeat

Captures human intelligence from governed channels (Slack, WhatsApp, Telegram, Custom GPTs) and converts it into executable actions.

Critical Rule: Brainstorming NEVER becomes truth directly. Truth is only created when an action writes to a tool, and the ingestion pipeline confirms and normalizes it.

9.2 Loop B: Tool-to-Truth (System-Governed)

Loader → 8-Stage Pipeline → Store (Spine) → Think Engine → Act Engine → Tool Write-Back → Repeat

Processes structured tool data through the continuous intelligence cycle.

9.3 Convergence

Loop A feeds Loop B through tools: human decisions result in tool updates, which Loop B ingests and normalizes
Loop A triggers change; Loop B confirms truth
Both converge into the Think Engine's views
Both repeat continuously
The convergence point is the Spine, where all validated truth resides regardless of origin

10. Infrastructure Stack

Layer	Service	Purpose
Compute	Cloudflare Workers	API Gateway, pipeline stages, connector logic, engine services. Workers Paid plan; 50ms CPU limit; auto-scaling.
Queue	Cloudflare Queues	Pipeline stage-to-stage messaging, DLQ, event bus. Dedicated queues per stage; batch size 10; max retries 3.
Database	Neon PostgreSQL	Spine (SSOT), Context Store, AI Chats Store, Audit Store. Serverless compute; auto-suspend; PITR enabled.
Vector Search	pgvector on Neon	Embedding storage and similarity search for Context Store. HNSW index; cosine distance.
Object Storage	Cloudflare R2	Raw files, document blobs, archived sessions, evidence attachments. S3-compatible; no egress fees.
Cache	Upstash Redis	Rate limiting, session tokens, temporary computation results. Serverless; per-request pricing.
Durable State	Cloudflare Durable Objects	Hot Memory, real-time collaboration state, connection sessions. Per-object isolation; transactional.
Cron / Scheduling	Cloudflare Cron Triggers	Scheduled sync polling, accelerator recalculation, cleanup jobs. Minute-level granularity.
DNS / CDN	Cloudflare	Domain routing, SSL termination, DDoS protection, edge caching. Full proxy mode; WAF rules.
AI Inference	OpenRouter	AI calls for Sanity Scan, Think Engine, session summarization. Multi-model routing; fallback chains.
Auth	Supabase Auth	User authentication, JWT issuance, social login, MFA. RLS integration with Neon PostgreSQL.
Secrets Management	Doppler	Centralized secrets management across all environments. Injects env vars at runtime; no .env files.
Workflow Automation	n8n (self-hosted)	Internal operational workflows, email triage, notification routing. Hosted at n8n.integratewise.online.
Frontend Hosting	Cloudflare Pages	Static site hosting for Vite React app. Automatic deployments from main branch.

11. Security Model

Security is enforced at every layer, from network edge to database row. Zero-trust principles: every request is authenticated, every query is scoped, every action is audited.

Layer	Mechanism	Implementation
Network Edge	DDoS protection, WAF, rate limiting	Cloudflare proxy with WAF rules; per-IP and per-tenant rate limits
Transport	TLS 1.3 everywhere	Cloudflare SSL termination; encrypted connections to Neon and R2
Authentication	JWT + HMAC + OAuth	Supabase Auth for users; HMAC for service-to-service; OAuth for tool connections
Authorization	RBAC + RLS	Role-based access control at API layer; Row-Level Security at database layer
Data Isolation	Tenant isolation via RLS	Every database query automatically scoped to tenant_id via PostgreSQL RLS policies
Encryption at Rest	AES-256	Neon PostgreSQL encryption; R2 server-side encryption
Secret Management	Doppler	Centralized secrets vault; environment-specific configs; automatic rotation; audit logging
Audit	Immutable audit log	Every data access, modification, and action logged to append-only Audit Store
PII Handling	Field-level detection	Pipeline Filter stage flags PII fields; configurable masking and retention policies
Compliance	GDPR, DPDP Act 2023	Right to access, erasure, portability. Cascade delete with audit trail.

11.1 Data Commitments

No data selling — data is never sold, rented, or shared with third parties
No AI training — customer data is never used to train AI models
Full portability — export all data at any time
Complete deletion — request deletion and everything is removed, with audit confirmation
Transparent processing — data is only processed as the customer directs

12. Queue and Event Architecture

Queue Name	Producer	Consumer	Purpose
connector.inbound	Webhook handlers / Cron pollers	Stage 1: Analyzer	Raw payloads from connectors
pipeline.analyzed	Stage 1: Analyzer	Stage 2: Classifier	Annotated payloads
pipeline.classified	Stage 2: Classifier	Stage 3: Filter	Classified payloads
pipeline.filtered	Stage 3: Filter	Stage 4: Refiner	Scope-approved payloads
pipeline.refined	Stage 4: Refiner	Stage 5: Extractor	Normalized payloads
pipeline.extracted	Stage 5: Extractor	Stage 6: Validator	Enriched payloads
pipeline.validated	Stage 6: Validator	Stage 7: Sanity Scan	Validated payloads
pipeline.scanned	Stage 7: Sanity Scan	Stage 8: Sectorizer	Quality-approved payloads
pipeline.dlq	Any stage on failure	DLQ Manager	Failed/quarantined items
engine.think	Sectorizer (post-write)	Think Engine	Data change events for reasoning
engine.act	HITL approval	Act Engine	Approved actions for execution
accelerator.trigger	Sectorizer (post-write)	Accelerator Router	Accelerator recalculation triggers
context.events	Slack / WhatsApp / Telegram	Context Processor	Governed channel messages
mcp.sessions	MCP Connectors	Session Processor	AI chat session captures

13. Growth-Aligned Data Model

IntegrateWise does not just track data — it tracks growth. Every record, metric, and signal is linked to organizational goals.

Dimension	Traditional Systems	IntegrateWise
Data Model	Static records: "We have 100 accounts with $850k ARR"	Active, goal-linked: "100 accounts at 85% of $1M goal, $150k gap, Acme at MEDIUM RISK"
View	Reactive — "What happened?"	Proactive, risk-aware — "What should I do to hit my goals?"
Action	Manual — human interprets and decides	Suggested, evidence-based — AI proposes, human approves
Learning	None — same static view regardless of outcomes	Continuous — every approval/rejection improves future proposals
Alignment	Data disconnected from goals	Every metric is a percentage of a goal, with risk scoring and gap analysis

14. Unique Architectural Combination

IntegrateWise combines eight architectural elements that no single competitor has brought together:

Element	Exists Elsewhere?	IntegrateWise
L0–L3 Cognitive Architecture	No	✓ Full 4-layer cognitive system
MCP-Native (AI Tool Discovery)	Anthropic only	✓ Integrated for AI session capture
Dual-Write (Truth + Context)	No	✓ Spine + Context Store parallel writes
Preserved Work Layer	Traditional apps only	✓ Enhance, don't replace existing workflows
Approval-Only Agents	No	✓ Hard HITL gate before every action
Edge Correction (Self-Learning)	Research papers only	✓ Adjust + Repeat components in production
Growth-Aligned Schema	BI tools (static)	✓ Active, goal-linked at database level
Continuous Learning Loop	No	✓ Dual-loop with convergence at Spine

No competitor has more than 2–3 of these elements. IntegrateWise has all eight, working together.

15. Development Guidelines

15.1 Code Organization

packages/
├── api/                    # L3 API services (Cloudflare Workers)
├── connectors/            # Tool integrations (Salesforce, HubSpot, etc.)
├── mcp-connector/         # MCP protocol implementation
├── web/                   # L1 Workspace (Next.js)
├── cognitive/             # L2 components (React)
├── types/                 # Shared TypeScript definitions
├── config/                # Environment and configuration
└── lib/                   # Shared utilities

services/
├── loader/                # 8-stage pipeline implementation
├── normalizer/            # Schema transformation (NA0-NA5)
├── spine/                 # SSOT data service
├── context/               # Unstructured data service
├── think/                 # Reasoning engine
├── act/                   # Execution engine
├── govern/                # Policy enforcement
├── audit/                 # Audit logging service
├── accelerators/          # Domain intelligence engines
├── iq-hub/                # AI conversation management
├── knowledge/             # Vector search and embeddings
└── gateway/               # API gateway and routing

15.2 Key Principles for Agents

No shortcuts through the pipeline — All data must flow through all 8 stages
HITL is mandatory — No AI action executes without human approval
Dual-write to Spine + Context — Every normalized record writes to both stores
Evidence for every claim — All signals and recommendations must have provenance
Tenant isolation is absolute — Every query must include tenant_id scoping
Audit everything — Every action, approval, and data change is logged
Normalize once, render anywhere — Canonical schemas are the source of truth
Growth-aligned by default — All metrics link to goals

15.3 Testing Requirements

Unit tests for all pipeline stages
Integration tests for connector OAuth flows
End-to-end tests for approval workflows
Load tests for pipeline throughput (target: 10k events/minute)
Security tests for RLS policies
Chaos tests for DLQ recovery

15.4 Secrets Management (Doppler)

NO LOCAL .ENV FILES. All secrets are managed through Doppler.

Required Doppler Secrets:

VITE_SUPABASE_URL          # Supabase project URL
VITE_SUPABASE_ANON_KEY     # Supabase anonymous key
SUPABASE_SERVICE_KEY       # Service role key (backend only)
OPENROUTER_API_KEY         # AI inference API key
N8N_WEBHOOK_SECRET         # n8n workflow auth

Development Workflow:

# Login to Doppler
doppler login

# Setup project
doppler setup

# Run dev server with Doppler
doppler run -- npm run dev

# Build with Doppler
doppler run -- npm run build

Environment Structure:

dev — Local development
staging — Pre-production testing
prod — Production environment

Security Rules:

Never commit secrets to git
Never share service keys in Slack/email
Rotate keys quarterly via Doppler
Use separate projects for dev/staging/prod

16. Contact & Resources

Company: IntegrateWise LLP
Founded: 2025
Location: Bengaluru, India
Founder: Nirmal Prince J

Key Documents:

docs/ARCHITECTURE_OVERVIEW.md — High-level system design
docs/CANONICAL_ARCH_SPEC.md — L0-L3 canonical model
docs/DATABASE_SCHEMA.md — Data store specifications
docs/SERVICE_MESH_ARCHITECTURE.md — Backend service organization

This document is the authoritative source of truth for the IntegrateWise system architecture. All implementation decisions should align with the principles and specifications outlined here.