Vendo AI

• 

  Dashboard - see the current state of your analytics at a glance

• 

  Annotations - bring context for what happpened in the business

• 

  Data Dictionary - see what data you have access to.

• 

  Agent Library - fast-start with premade agents (coming soon)

Inspiration

• We are a data and analytics consultancy that helps organisations get the most out of their data.
• We’ve observed that while businesses invest heavily in data infrastructure, only a handful of people use these tools, leaving most data underutilised.
• For organisations without dedicated analysts, this challenge is even bigger. For those with analysts, capacity remains limited.
• Many clients pay us thousands per month to get access to additional analyst capability.
• We believe Google’s new Agent Development Kit enables us to quickly build an AI solution that automates this service at scale.

What It Does

• Vendo AI builds on our core product, Vendo Data, to give growth teams a self-serve AI assistant for exploring, understanding, and acting on their data.
  • Vendo Data is a no-code data pipeline that connects all key data sources and automatically prepares the data for analysis in BigQuery.
  • Supports: Mixpanel, Google Analytics, Shopify, Google Ads, Meta Ads, Stripe, PayPal, HubSpot, Salesforce, and more.
  • Includes a diagnostic module that scans all events, properties, schemas, and anomalies, creating a Data Dictionary that the AI uses to understand what’s available.
• Vendo AI adds a powerful multi-agent system on top:
  • Users can ask business questions in plain English.
  • The agents plan how to answer, locate the correct data, run the analysis, and present the results as text, charts, or reports.
  • Agents handle context, cross-check with the Data Dictionary, and escalate complex requests to human analysts if needed.
  • This means faster insights, reduced analyst workload, and better marketing decisions — without needing to write SQL or know the schema details.

How We Built It

• We designed a modular multi-agent architecture using Google Agent Development Kit (ADK) with a root agent coordinating specialised sub-agents:
  • Data Planner: suggests which events and properties to track.
  • Data Retrieval Agent: runs SQL/API queries across BigQuery, Mixpanel, etc.
• Data Integrations: Connectors for Google Ads, Meta Ads, TikTok Ads, Shopify, Google Analytics, Microsoft Ads, Mixpanel, Stripe, PayPal. Runs via Cloud Scheduler, Tasks, Cloud Run to BigQuery.
• Data Quality & Context: Automated checks, Data Dictionary in Firestore, session context stored with chat history and annotations.
• Frontend & Backend: Backend on Cloud Run; frontend on Next.js + Vercel.
• Agent Tools:
  • schema_tools based on user request, examines the schema to identify the correct event, event property, and user properties.
  • run_bigquery to convert natural language into SQL
  • build_chart to create charts based on the data.
  • notify_vendo escalates unresolved tasks to human analysts. Sends a Slack message.
  • google_search fetches external context when needed.

Challenges We Ran Into

• Agent Orchestration: Routing to sub-agents and routing back to the root agent when the sub-agents are stuck, and escalating to human analysts.
• Schema Complexity: Adding all the schemas to the context blew up the context usage.
• Slack Integration: Handling threads, permissions, logging in Firestore.
• Inconsistent Agent Behaviour: The agents refused to comply with user requests!
• Error Handling: Fallbacks and escalation to human analysts.

Accomplishments That We're Proud Of

• We build the tooling to answer complex questions that require advanced SQL knowledge.
• Built the current prototype in about three weeks.

What We Learned

• Data quality is everything.
• Efficient context management reduces token cost drastically.
• Few-shot examples are critical.
• Better models simplify prompts.
• Monitoring and EVALs are vital for continuous improvement.

What's Next for Vendo AI

• Test with real customers: Conduct an alpha release with a select group of customers.
• Merge into main Vendo App: The agent will be integrated into the main Vendo App.
• Better EVALs: Creating more granular EVALs for each step to measure and improve.
  
• Deeper Slack integration: Slash commands, real-time alerts.
• Agent Library: Creating the agent library and new skills for things such as budget reallocation, anomaly detection, and automatic reporting.
• Build Dashboard: Starting point to see business overview

Built With

• bigquery
• cloud-run
• fastapi
• firebase
• gemini-api
• google-adk
• next.js
• python
• vercel
• vertex