Satnam Singh← all work
2025Support Automation · FoundryDemo available

Startup Application Review Agent

An AI-assisted application review system built on Azure AI Foundry, Microsoft Fabric, and Azure AI Search for MFS and AFS application checks, duplicate detection, and approve / reject recommendations.

Azure AI FoundryMicrosoft FabricAzure AI SearchLogic Apps
headline impact
Demo
automation for MFS & AFS startup applications · grounded in Fabric + vector knowledge
my role
Role
Developer · agentic architecture, application review logic, backend integration, data flow, and agent grounding
status
Status
Demo available · agentic review workflow
demo
Demo available
Demo walkthrough available
the problem

What was broken

Program and support teams reviewed startup applications across fragmented systems — investor codes, OneVet fields, duplicate applications, LinkedIn profiles, startup countries, and ticket resolution history. The process depended on manual checks across multiple sources, making the review slow, inconsistent, and hard to audit.

my approach

How I built it

I built an agentic review system with Fabric as the governed data layer and Azure AI Foundry as the agent layer. The review workflow receives application status, ticket recommendation, and ticket details from the backend agent process. External sources such as DFM, Founder Hub CRM, and collaboration files land in a Fabric Lakehouse via Shortcut, Dataflow, and Pipeline on a Medallion architecture with backup and domain knowledge infusion. Two specialised Foundry agents — MFS Application Agent and AFS Application Agent — apply their own checks for investor code, OneVet fields, duplicates, LinkedIn, startup country, and tech-based classification, then return an approve / reject recommendation. Vectorised ticket and domain knowledge in Azure AI Search grounds every agent call, refreshed on a daily cadence.

why this way

The reasoning

Splitting MFS and AFS into dedicated agents keeps each prompt focused on the rules that actually apply, with its own evals and retry boundary. Fabric + Medallion architecture gives a single governed data layer, while AI Search vector indexes give the agents grounded retrieval over tickets and domain knowledge instead of hallucinated rules.

architecture

How the pieces fit

Review Operations
Application status workflowTicket recommendation workflowTicket issue details
State & Processing
Application review stateTicket resolution recordsAgent output logging
Data Layer (Fabric)
External: DFM, Founder Hub CRM, collaboration filesIngestion: Shortcut, Dataflow, PipelineLakehouse · Medallion · BackupDomain Knowledge Infusion
Agentic Layer (Azure AI Foundry)
Application Review AgentMFS Application AgentAFS Application Agent
Retrieval (Azure AI Search)
Vectorised Domain KnowledgeStructured Ticket DataEmbeddings · Semantic / Vector Search
what I built

Key components

  • 01Agentic review logic for application status and ticket recommendations
  • 02Application review state and ticket resolution record handling
  • 03Fabric Lakehouse with Medallion architecture, Dataflows, Pipelines, and backups
  • 04MFS agent: investor code, OneVet fields, duplicate checks, approve / reject
  • 05AFS agent: LinkedIn profile, startup country, tech-based classification, duplicates, approve / reject
  • 06Vectorised domain knowledge + structured ticket data indexed in Azure AI Search
  • 07Daily ingestion / refresh cadence from source systems into Fabric and vector knowledge
  • 08Workflow integration from agent recommendations back to the review process
what I used

Tech stack

Agentic AI
Azure AI FoundryAzure OpenAIRAGMulti-Agent Orchestration
Data
Microsoft FabricLakehouseMedallion ArchitectureDelta LakeDataflowsPipelines
Integration
Azure Logic AppsFunction AppsREST APIs
Retrieval & Security
Azure AI SearchVector SearchMicrosoft Entra IDManaged Identity
video demo

Supabase-hosted walkthrough

Startup Application Review demo

see it run

In-browser demo

A scripted walkthrough of the demo flow, traced step by step.

application_review_agent · run #APP-20581
press play to watch the trace
next up
Startup Advisor & Co-sell Intelligence Platformopen →