1 / 9
Nexus Insight | Microsoft Innovation Challenge - March 2026

Analytical AI
with Measured Results

We transform business questions into actionable executive answers through a secure, self-correcting, and auditable multi-agent architecture.

Enterprise Security
Low Latency
Cloud-Native

The Nexus Insight Identity

More than a tool. A paradigm shift bridging the gap between fluid human intent and rigid enterprise data.

Nexus

The central connection point. We seamlessly link disparate, complex relational schemas directly to the user's natural language.

Insight

The ultimate outcome. We don't just execute SQL and return raw tables. We synthesize data into polished, actionable executive intelligence.

Designed for Impact (Judging Criteria)

25% Performance (Measurable)

Not just a proof of concept. We validated a 62.5% reduction in tokens and 50.5% in P50 latency using a Multi-Agent pipeline with schema caches and fast-tracks.

25% Innovation

Grounded in the MAC-SQL paradigm. We divide Text-to-SQL complexity into specialized roles, creatively overcoming the context limits of monolithic models.

25% Responsible AI

Strict alignment with Microsoft's 6 AI principles: Guaranteed blocking of injections, Human-in-the-Loop workflows, and forensic W3C Traceability.

25% Azure Ecosystem

Production-grade architecture. Seamlessly orchestrates Azure Container Apps, Entra ID, AI Foundry, Azure SQL, Key Vault, Web PubSub, and App Insights via Bicep.

Agent Flow: Efficiency by Design

Theoretically inspired by MAC-SQL research, each stage of our architecture exists to reduce costs, improve accuracy, or protect execution.

Root Problem

Enterprise Text-to-SQL combines high natural language ambiguity with extreme relational schema rigidity. If a monolithic agent attempts to solve everything, errors, latency, and token costs soar.

Design Principle

Decompose into agents with bounded responsibilities + stage telemetry + secure execution to iterate with empirical evidence.

1) Planner + Short-Circuit

Cuts early if it detects a simple chat and jumps straight to Evaluator. Avoids full SQL stages when they add no value.

2) Librarian Cache (TTL 300s)

Reuses catalog and enriched schema (Semantic Pruning) to reduce repeated round-trips to SQL metadata.

3) Critic + Execution

Blocks DDL/DML. If it fails, it feeds the error back to the Coder and retries guided by real execution (max 2).

4) Fast-Track

Reuses results from repeated queries dynamically, achieving a 50% hit-rate in real-world tests.

5) Streaming + Transparency

The backend emits deltas with transparency metadata: timings per stage, SQL generated, and live rows processed.

Technical Defense & Business Value

Zero-Trust Execution

How do we prevent destructive queries (DROP TABLE) or massive data extraction?

Critic Agent
Blocks DDL
Guardrails
TOP 200 / Timeout
RBAC
Read-Only SQL
PII / Confidential Data triggers HITL (Teams)

Ad-Hoc Agility vs. Static BI

Why use this chat instead of existing corporate Power BI dashboards?

Traditional BI Known metrics only
Time-to-Market: WEEKS
VS
Nexus Insight 100% Ad-Hoc Questions
Time-to-Market: 5 SEC.

Decoupled Async Streaming

How do we handle progressive responses (streaming) without saturating the Next.js HTTP server?

Next.js
Auth Only
Web PubSub
WebSockets
AI Engine
Token Stream

Strict WebSocket Segregation

How do we ensure privacy and prevent messages from crossing between user sessions?

1. Entra ID Login
Certifies user identity before any connection
2. Signed URL & Access Token
Temporal, restricted access via PubSub
3. Ephemeral Exclusive Group
Strictly tied to a unique `correlation_id`

Performance: Live & Synthetic Benchmarks

Real Scenario (Live Azure SQL)

Variant A (Monolith)
32 tokens avg
Variant C (Nexus)
12 tokens avg

62.5% Reduction in Computational Cost

Synthetic Scenario (Latency)

P50 Baseline
2760 ms
P50 Nexus Insight
1365 ms

50.5% Reduction in user wait time

Responsible AI: Microsoft's 6 Principles

Reliability & Safety

Our Critic Agent acts as a firewall, validating 100% of queries. Integrated natively with Azure AI Content Safety to prevent prompt jailbreaks and ensure totally safe operations.

Privacy & Accountability

Built on Entra ID for robust identity management. The Human-In-The-Loop workflow guarantees that a real human is strictly accountable for executing high-impact queries.

Transparency & Inclusion

Every token is tracked using W3C Trace Contexts in Log Analytics. Plus, our architecture sets the foundation for true accessibility utilizing seamless voice integrations.

End-to-End Architecture Interactive

Docker deployment in Azure Container Apps, CI/CD, Agent Framework, and Identity. Click on nodes, scroll to zoom, and drag to explore.

CI/CD & IaC
Identity & Security
Azure Container Apps (Docker)
Azure AI Foundry & Agent Framework
Data Layer
Observability & Benchmarking

GitHub Actions

CI/CD pipelines

Bicep (IaC)

Infra as Code

Entra ID

Authentication & RBAC

Azure Key Vault

Secrets & AI Keys

Next.js Frontend

UI Container

Redis Cache

Chat history container

Python FastAPI

Backend API

Azure Web PubSub

Streaming WSS

Planner Agent

Router & Short-circuit

Librarian Agent

SQL schema cache

SQL Coder Agent

Generation in Foundry

Critic Agent

Zero-trust DDL blocker

SQLExecution

DB Reader with Retries

Evaluator Agent

Natural formatter

RealtimeListener

Events status & telemetry

Azure SQL DB

Transactional destination

App Insights

APM, latencies, exceptions

Log Analytics

KQL logs and metrics

Benchmark Engine

Internal A/B analytics

🚀 Future Vision: Cloud-Native Ecosystem

Established Foundation (Phase 1)

Container Apps Azure SQL Agent Framework Zero-Trust Sec

Extended Generation

Integration of Model Context Protocol (MCP) to automatically generate dynamic Power BI Dashboards, extract PDF reports, and build automated PowerPoint presentations.

Comprehensive Inclusion

Native use of Azure AI Speech to democratize data: enabling spoken queries (Speech-to-Text) and reading executive reports out loud (Text-to-Speech).

HITL & Teams Workflows

Approval of critical queries (Human-In-The-Loop) via Adaptive Cards in Teams for authorized users, and autonomous dispatches using Microsoft Graph API.

Detail

Component