Mohammad Daoud Farooqi
Partner Solutions Architect, AI at MongoDB
AI agents in production are stitched together from four systems that were never meant to talk to each other: a vector database, a keyword engine, a graph store, and a system of record. My job is to collapse that into one. I help banks, insurers, and large data teams ship agents in the industries where “move fast and break things” was never an option.
Ai4 2026 · Venetian Ballroom H, Las Vegas
Fraud Detection Agents at Scale
Built on MongoDB + Mastra
Fraud detection is a hard place to put an agent. The evidence behind a decision is scattered, only some calls should reach a human, and every one of them has to survive an audit. Co-presented with Brandon Barros of Mastra, this session walked through a fraud investigation console that assembles its evidence server-side in MongoDB Atlas, running vector, full-text, and graph retrieval over a single collection. Mastra's durable workflows wrap that in a deterministic flow that suspends for an approver and resumes where it left off, which is what produces the audit trail, the PII boundary, and a path to self-improvement from production traces.
What the session covered
- A working reference architecture: a TypeScript agent built on Mastra, backed by a single MongoDB Atlas cluster for vectors, durable state, and long-term memory.
- Evidence assembled server-side: vector, full-text, and graph retrieval over one collection, with change streams keeping an open investigation current.
- The governance layer fraud work actually needs: a deterministic workflow that suspends for a human approver and resumes, an audit trail, and PII protection.
- The open-source patterns behind the demo: memory services, hybrid retrieval, and agentic quickstarts you can run today.
About
Agents That Reach Production
I'm a Partner Solutions Architect for AI at MongoDB and the author of the technical whitepaper on Constitutional AI ethical governance with MongoDB Atlas. My work is fixing a problem most teams know too well: AI agents in production are stitched together from four systems that were never meant to talk to each other. I help enterprises move agents into production in the industries where reliability, governance, and compliance are non-negotiable.
I spent years delivering for banks and insurers, so I worked inside data governance, security, and compliance constraints long before I had to architect around them. Today I author the reference architecture for MongoDB's internal Database-as-a-Service platform, and I build the open-source AI reference architectures that run on Atlas: long-term memory, hybrid vector and full-text search, and agentic quickstarts adopted across the partner ecosystem.
Beyond the Constitutional AI whitepaper, I wrote MongoDB's guide to building AI memory systems with AWS and Claude, and an official Anthropic cookbook, published on platform.claude.com and in anthropics/claude-cookbooks, showing MongoDB Atlas as the retrieval, graph, and system-of-record layer for Claude Managed Agents. At Ai4 2026 I took that further on stage with Mastra: a fraud investigation console where vector search, graph traversal, durable state, and human approval all sit on one operational data layer.
Focus Areas
Writing
Publications
Published technical writing on governing, remembering, and retrieving for production AI systems
Fraud Review Agent with MongoDB Atlas and Claude Managed Agents
Published by Anthropic on the Claude platform docs and in the anthropics/claude-cookbooks repository: MongoDB Atlas as the retrieval engine, graph store, and system of record for a human-in-the-loop fraud-review agent. Covers all four retrieval patterns ($vectorSearch, $search, $rankFusion, $graphLookup), a credential-safe host-side data path, and audit persistence.
Read →Constitutional AI: Ethical Governance with MongoDB Atlas
A complete guide to building self-governing AI systems on Atlas: RLAIF self-critique, hybrid search with two-stage retrieval and reranking, Change Streams audit trails, and Queryable Encryption. Written for regulated finance, healthcare, and legal.
Read →Constitutional AI: Ethical Governance with MongoDB Atlas
The original blog introducing Constitutional AI on Atlas: how Anthropic’s self-governance approach scales when constitutional principles live in the database, with Voyage AI embeddings for semantic rule matching and a reference architecture for responsible AI systems.
Read →Build AI Memory Systems with MongoDB Atlas, AWS, and Claude
Long-term memory for AI agents on Atlas, AWS Bedrock, and Claude: hierarchical memory structures, importance scoring, and semantic recall. Pairs with the open-source agent-memory service below.
Read →The Converged Datastore for Agentic AI
Why agent runtimes collapse onto one operational datastore instead of a sprawl of vector databases, caches, and relational stores. The thesis behind the Ai4 session.
Read →AI Fraud Detection with MongoDB Atlas and Temporal
Subsecond fraud detection without static rules: Atlas Vector Search over finance-tuned Voyage AI embeddings to find similar historical transactions, aggregation pipelines for money-laundering network analysis, and Temporal durable execution for crash-safe transaction orchestration. Pairs with the open-source mongodb-temporal-ai-agent-qs quickstart below.
Read →I also write about AI systems, MongoDB, and system design on my personal blog, threadwaiting.com →
Open Source
Agent Infrastructure You Can Run Today
Reference architectures and developer tooling for production AI agents, built at MongoDB and adopted across the partner ecosystem
agent-memory
Four kinds of agent memory in one Atlas cluster: short-term session state, importance-scored long-term semantic recall, episodic history, and a semantic cache. Served over MCP and REST. Supersedes ai-memory and memory-mcp, which are folded in.
mongodb-llama-stack
MongoDB integration for Llama Stack: vector, full-text, hybrid search, and graph-enhanced retrieval built on native Atlas features. Registered upstream as an external provider in OGX, formerly Meta’s Llama Stack.
mongodb-mastra-agentic-ai-qs
Fraud Investigation Console built on MongoDB and Mastra: a TypeScript agent with vectors, state, and memory in one Atlas cluster. The pattern behind the Ai4 session.
maap-anthropic-qs
End-to-end MAAP stack combining MongoDB Atlas with Anthropic’s Claude models for an agentic conversational interface. A production-shaped quickstart.
mongodb-gemini-extension
The official Gemini CLI extension for MongoDB, wrapping the mongodb-mcp-server npm package for one-command installation with automatic upstream updates.
More Reference Architectures
Platform Work at MongoDB
Internal platform engineering, not open source, but very much production.
MongoDB Database-as-a-Service Platform
The platform layer that provisions and runs MongoDB as a managed service, plus the control plane automating provisioning, lifecycle, and tenant orchestration for partners and OEMs.
Partner Pavilion
A full-stack production storefront and SSO admin dashboard showcasing partner AI, streaming, workflow-orchestration, and semantic-search integrations with Atlas.
Industrial Predictive Maintenance
Predictive maintenance system for industrial operations built on AWS, MongoDB Atlas, and Voyage AI embeddings.
Speaking & Judging
On Stage
Talks and judging on production agent infrastructure, AI governance, and the converged data layer
Ai4 2026 — Fraud Detection Agents at Scale, Built on MongoDB + Mastra
Co-presented with Brandon Barros of Mastra: a fraud investigation console that assembles evidence server-side in Atlas, pauses for a human approver, and leaves an audit trail. Session details
Give Your Agents a Memory
Sessions at the MongoDB booth on what agents remember and how: short-term state, long-term semantic memory, memory over MCP, and hybrid retrieval on Atlas.
Google Cloud Rapid Agent Hackathon
Judge, MongoDB Track
Judged Google’s global hackathon on building agents for real-world challenges: 14,000+ participants and $60,000 in prizes, with agents built on Gemini, Google Cloud Agent Builder, and partner MCP servers.
View on Devpost →Agentic Apps on Real-Time Data with MongoDB Atlas & Confluent
Public co-presented webinar on closing the loop between live streams and agents: Kafka and Flink for real-time anomaly detection, Atlas Vector Search for context, and Atlas Stream Processing feeding actions back into the system.
MongoDB Developer Day at a Global Investment Bank
Hands-on developer day with a runnable cookbook: Atlas AI workloads, LangChain and LangSmith agent patterns, AgentCore, Claude Code, and the MongoDB MCP server.
Build for the Agentic Era
Public joint webinar with LangChain on the future of LLM application development for a leading US grocery retailer: MongoDB for agentic apps, the agent development lifecycle, and a live MongoDB + LangSmith demo.
From Prompt to Production: Using MongoDB Agent Skills with Vibe Coding Tools
Public webinar on going from an idea in a prompt to a production-ready, governed AI workflow with MongoDB Agent Skills, MCP, and vibe coding tools, delivered to a 400+ audience.
View event →The Converged Datastore: Agentic for Insurance & Other Industries
Featured speaker in front of 500+ developers.
Booking talks on production agents, AI memory, and governance for regulated industries. Get in touch
Contact
Let's Talk Agents
Following up from Ai4? Exploring a partnership or a speaking slot? I'd love to hear from you.
farooqimdd@gmail.com
linkedin.com/in/farooqimdd
GitHub
github.com/mohammaddaoudfarooqi
Phone
+91-9599555734
Launch a chat →
Met me at Ai4 2026?
If we talked at the MongoDB booth, after the fraud-detection session, or over a drink at the Mastra happy hour, pick “Ai4 2026 follow-up” in the form and I'll pick the thread back up.