Domain AI agents that answer from your knowledge — not generic guesswork.
A LangGraph RAG multi-agent marketplace where HR, Finance, Support, and Operations each maintain their own knowledge base. Built with Qdrant, Google Gemini, NestJS microservices, and React — turning fragmented enterprise documents and emails into conversational intelligence.
Enterprise knowledge was abundant — and almost impossible to reach.
Critical information lived across documents, emails, and department silos. Traditional search lacked context, and a single generalised AI model could not meet the accuracy demands of multiple business domains.
Knowledge silos across departments
HR, Finance, Support, and Operations each held separate, inaccessible information stores.
Manual document search without context
Search was slow and returned results without meaningful relevance to the question.
Unstructured data left behind
Emails, scanned files, images, and PDFs sat outside any intelligent retrieval system.
No scalable multi-domain AI
Adding AI for a new department meant duplicating infrastructure and engineering effort.
Client
Enterprise knowledge organisation
Industry
AI & Intelligent Systems
Integrations
LangGraph · Qdrant · Gemini · Gmail
Engagement
Hours → seconds retrieval
What we learned before we designed anything
Findings from HR, Finance, Support, and Operations teams — the basis for every multi-agent, RAG, and unstructured-data decision that followed.
Generic AI lost trust immediately
A single model serving all departments produced answers too broad to act on — domain specificity drove adoption.
Most knowledge lived in unstructured formats
Scanned PDFs, email threads, and Word docs were inaccessible to standard search.
First-result accuracy decided adoption
Employees abandoned tools that required further manual filtering after the first hit.
Email analysis consumed Support and Operations time
Automated summarisation and action-item extraction offered immediate measurable return.
Departments needed to own their knowledge bases
Agent configuration had to be admin-accessible, not developer-dependent.
One search box over one store
Most enterprise knowledge tools lack per-domain agents with independent knowledge bases.
RAG built for one use case
No marketplace-style platform let organisations deploy and scale multiple domain agents from one system.
Email and OCR stay standalone
Competitors rarely fold email and document intelligence into the same conversational AI layer.
Rohit
HR Manager · Mid-Sized Enterprise
Goals
- • Give employees instant accurate policy answers
- • Reduce time spent searching internal documents
- • Get email summaries and action items without reading every thread
Pain Points
- • Policies spread across SharePoint, email, and manuals
- • Generic AI that ignores company-specific rules
- • No way to query scanned onboarding docs or email history
One marketplace, many agents, one RAG core.
LangGraph orchestration, Qdrant vector collections per domain, NestJS microservices, and Gemini generation — with OCR and email intelligence feeding every agent's knowledge base.
One connected marketplace, built agent by agent.
Every domain agent shares orchestration, retrieval, conversation, and security — only the knowledge base and configuration differ.
A single AI model produced generic, low-trust responses across all enterprise domains.
HR, Finance, Support, and Operations agents with their own knowledge
- ✓Independent Qdrant vector collections per domain eliminate cross-contamination
- ✓Global Intelligence Agent handles cross-domain fallback so no query goes unanswered
Enterprise knowledge was scattered across documents, scanned files, and emails — none of it queryable.
Upload, index, and make unstructured data conversational
- ✓PDF, DOCX, and TXT documents auto-indexed into the relevant agent's vector collection
- ✓OCR converts scanned documents and images into searchable, embeddable content
Traditional search returned results without contextual relevance or reasoning.
From query to grounded Gemini response
- ✓Query → embed → Qdrant semantic search → context retrieval → LangGraph → Gemini
- ✓Multi-step reasoning synthesises multiple retrieved documents in one coherent pipeline
Scaling AI to new departments required new infrastructure and engineering each time.
Admin-configurable agents plus Gmail intelligence
- ✓Admins create agents, upload knowledge, and map collections without developer involvement
- ✓Gmail integration summarises threads and extracts action items for Support and Operations
Built for grounded multi-agent enterprise scale.
LangGraph, Qdrant, Gemini, NestJS, React, Docker, and Kubernetes — chosen so every response stays fast, secure, and grounded in retrieved enterprise data.
RAG first, then marketplace agents on shared infrastructure.
Retrieval accuracy was treated as the prerequisite for every domain agent — hallucination was not acceptable in HR, Finance, or Support contexts.
Discovery
Validated that domain specificity — not a single generalised model — was the primary driver of user trust across departments.
RAG pipeline first
Proved query → embed → semantic search → context → Gemini end-to-end before any agent specialisation began.
Marketplace architecture
Built independent domain agents with per-agent Qdrant collections, plus LangGraph orchestration for multi-step reasoning.
Unstructured data coverage
Shipped OCR and Gmail email intelligence as first-class services so scanned docs and inboxes became queryable.
Secure cloud launch
Deployed NestJS microservices on Docker and Kubernetes with JWT, RBAC, CI/CD, and React real-time conversation UI.
What employees ask every morning
What changed after launch
Information retrieval from hours to seconds - Employees get accurate, domain-specific answers in a single conversational query.
Document and email analysis automated - Support and Operations freed from repetitive information extraction tasks.
Hallucination structurally minimised - RAG grounding ties every response to retrieved enterprise data, not model guesswork.
New AI domains without engineering effort - Admin-configurable agent creation scales across departments on shared infrastructure.
Scanned and unstructured data made conversational - OCR brings legacy and physical documents into the intelligence layer.
Connected multi-domain intelligence ecosystem - HR, Finance, Support, and Operations each have dedicated, trusted AI expertise on demand.
Challenges & Learnings
Vector collection management at scale
Independent Qdrant collections per agent needed careful indexing and query optimisation to stay fast.
RAG accuracy vs. recall trade-off
Embedding models and similarity thresholds required iteration to retrieve enough context without noise.
OCR quality variation
Robust pre-processing was essential for clean, embeddable text across varied scan quality and layouts.
Email intelligence scope
Meaningful insights and action items needed a dedicated extraction pipeline, not generic summarisation.
LangGraph multi-step latency
Pipeline design and response streaming kept multi-step reasoning responsive for users.
Where the marketplace goes from here
WhizCloud built something we had been unable to find off the shelf — a platform where each department gets its own intelligent assistant, grounded in our actual data, with responses we can trust.
“The HR agent alone has transformed how employees access policy information. The email intelligence capability was an unexpected highlight — our Support team's time spent on communication analysis dropped significantly within the first month. The WhizCloud team understood the architecture requirements deeply and delivered a platform that scales with us, not against us.”