Project
Splice AI
My Role
Co-founder
Category
Startup
Year
2024
Summary
Built a Data Analytics platform with Multi Agent System.
Overview
I co-founded Splice AI with Sahil Kumar, former Director of AI Products at Twilio. In our previous roles, (Sahil in product leadership and myself as Head of Design at ThousandEyes) we both depended heavily on data to make informed decisions. Yet time and again, we ran into the same frustrating pattern.
The Problem
- Low bandwidth for Data Analysts to handle ad-hoc requests.
Product and UX teams often needed point-in-time data to make key decisions. These could range from evaluating a feature launch to prioritizing design changes. These requests, while urgent for the requestor, often sat in the backlog. - Leadership requests were always high priority, forcing analysts to shift focus constantly and leaving even less time to work on strategic, long-term projects.
- For SMBs, the challenge was even more acute, many didn’t have dedicated analysts at all, meaning business-critical decisions were made without reliable data.
We hypothesized that if this problem was persistent and costly at large enterprises with established analytics teams, it was a significant blocker for SMBs, especially those operating multiple SaaS tools without an integrated way to analyze their data.
Our Approach
We set out to build Splice AI: a plug-and-play, AI-powered analytics platform that could take on the entire data analysis process end to end.
Splice AI High-Level Architecture
- Multi-agent system: Built on Autogen, with specialized agents acting as a Project Lead, Data Engineer, Data Cleaning Specialist, Data Analyst, Visualization Expert, and Industry Expert.
- End-to-end workflow: From data discovery → cleaning & preparation → analysis → visualization & reporting — all handled autonomously by multiple AI agents.
- Human-like output: Easy for business (non-technical) users to run analyses without relying on an expert data analyst.
Connecting data
User can connect various datasets together.
Analysis
Users can choose from preset analyses or ask a question.
Provide additional context.
Pull in the correct datasets.
Clean data and prepare it for analysis.
Analyze data and create a sharable report.
Dashboards
Users can choose a dashboard or prompt to create anew one.
Interact with the data or ask follow up questions
Request new visualizations and ways to look at the data.
Get instant updates to the dashbaord with language commands.
Problem Validation
Before we started building, we wanted to validate whether our own experiences reflected a wider industry reality. We spoke to ~30 data analysts and analytics leaders from companies like Cisco, Twilio, Tesla, PayJoy, Gojek, PG&E, and a range of mid- to small-sized businesses.
Our goal was to:
- Understand the day-to-day friction and pain points in delivering analytics.
- Identify which parts of the data workflow were most painful.
- Learn how constraints differed between enterprises and SMBs.
Key Insights from Interviews
1. Overwhelming Ad-hoc Demand
- Analysts wanted to cut down time spent answering ad-hoc questions, but leadership requests were always prioritized, leaving less bandwidth for strategic projects.
2. Data Cleaning & Discovery Are the Biggest Time Sinks
- Cleaning and validating data was described as “one of the most difficult and time-consuming parts” of the process.
- Multiple tables with no relation mapping and no data catalog in most cases made discovery inefficient.
3. Manual Data Wrangling is the Norm
- Many still relied on VLOOKUPs, filtering, and manual file validation across multiple Excel sheets.
- A lack of version control (most analysts don’t use Git) led to inconsistencies and duplicated effort.
4. Data Quality & Trust Issues
- Users were often suspicious about the accuracy and completeness of source data.
- Self-service tools existed, but analysts didn’t trust non-technical users to use them correctly.
Impact on Our Design Process
- Narrowed our scope to focus on three high-impact areas: data discovery, data cleaning & validation, and report generation.
- Ensured the system could show both the data and its source to build trust.
- Prioritized multi-agent specialization so different agents could own distinct roles (mirroring a human analytics team).
- Designed the UI to make conversational interface open but guided, reducing the risk of misuse.
v0: Conversational Interface
From the start, we knew speed of iteration would be critical. Instead of spending months building a fully functional prototype, we decided to design and test rapidly in Figma, validate assumptions, and refine based on direct feedback.
Our Approach
- Built the first set of high-fidelity mockups in Figma, simulating the complete Splice AI workflow from connecting a data source to receiving a visualized report.
- Used these mockups to walk through scenarios with the same group of ~30 analysts and leaders we had interviewed during the problem validation phase.
- Kept testing sessions short and focused, aiming to spot dealbreakers and must-have features before writing production code.
User can ask their questions in a chat interface
Project Manager Agent clarifies requirements and adds other agents as needed.
Data Collection Agent writes a script to retrive the relevant tables from the warehouse.
Data Cleaning Agent addresses null values and offers recommendations.
Data Analyst answers the question and Visualizer creates relevant visualizations to tell the story.
Key Insights from Testing
1. Visibility into data was essential
Analysts didn’t just want the final visualizations, they wanted to see the raw data being used. This transparency was necessary for them to trust the output and verify accuracy.
2. Access to the code was non-negotiable
Any SQL or Python code generated by the multi-agent system needed to be visible and editable. Analysts wanted to tweak queries, optimize joins, and validate logic without leaving the tool.
Impact on the Next Iteration
- Added a data preview panel alongside visualizations so users could inspect underlying datasets.
- Integrated a code editor (similar to Jupyter notebooks) where generated SQL or Python scripts could be reviewed and modified before execution.
- Created a canvas in the UI to allocate dedicated space for “Data” and “Code” tabs, ensuring transparency and control.
v1: Adding a Canvas
As we moved from inital mockups to a build-ready design, we stayed anchored to our primary user: the data analyst. Our goal was to give analysts a central platform where they could consolidate their data, scripts, and analyses, and then seamlessly share results with business stakeholders.
Evolving the Concept
- Based on feedback, we expanded the vision to create a workspace where analysis could evolve in real time as analysts interacted with the multi-agent system.
- The solution we landed on was a Jupyter/Colab-style notebook, embedded in a canvas area:
- As the analyst worked with Splice AI’s multi-agent system, each step, from data import to transformation to visualization, was logged in the notebook.
- Generated SQL or Python code appeared inline and could be edited directly by the analyst.
- This created an interactive, documented analysis trail.
User can ask their questions in a chat interface.
Project Manager Agent clarifies requirements for the analysis. Populates the notebook as per the discussion.
Project Manager Agent pulls in other Agents as required.
Data Collector Agent helps retrive the correct datasets.
Data Cleaning Agent recommends actions and cleans the data.
Data Analysis Agent performs the analysis and creates visualizations to answer the question.
Selected visualizations can added to a report to be shared with relevant stakeholders.
Key Features of the Canvas Notebook
- Data + Code + Output in One View
Analysts could see raw data, generated code, and resulting visualizations all in the same space. - Real-time Updates from Multi-Agent System
Agents handled different stages of analysis (data discovery, cleaning, visualization, etc.), posting their contributions directly into the notebook. - Seamless Sharing & Collaboration
Once complete, notebooks could be shared with PMs, sales leaders, or other stakeholders, who could interact with the results without editing core logic.
Technical Choices
- Frontend: Built with Next.js for performance and flexibility.
- Canvas Area: Designed to handle both structured outputs (charts, tables) and unstructured conversational logs from the agents.
- Backend Integration:
- The UI connected to a multi-agent Autogen backend.
- Each specialized agent (Project Lead, Data Engineer, Cleaning Specialist, Analyst, Visualization Expert, Industry Expert) posted results back to the UI through a real-time API.
- This allowed analysts to interact conversationally while maintaining full transparency into the process.
v2: Focus on SMBs
As we continued our research and conversations, a new opportunity emerged. While our initial focus was on empowering data analysts, we realized Splice AI could be equally, if not more, valuable for small to medium-sized businesses (SMBs) that didn’t have analysts at all.
The SMB Reality
SMBs still had a strong need to make data-driven decisions: understanding user behavior, tracking product adoption, analyzing financial performance. But these tasks were often handled by founders, sales leaders, or engineers - people already wearing multiple hats. Without the time or bandwidth to run full-blown analyses, decisions were often made with partial information or outdated reports.
Our Pivot for Business Users
We decided to create a version of Splice AI tailored to SMBs, with a different balance of complexity and usability:
- Conversational-first interface: Users could simply ask questions in plain language and receive visualized insights without touching raw data.
- Optional Transparency for Power Users: Business leaders who had the expertise could still open the hood to see or modify the data tables and scripts.
- Guided Insights: The system proactively suggested relevant metrics or follow-up questions, reducing the need for manual exploration.
Connecting data
User can connect various datasets together.
Analysis
Users can choose from preset analyses or ask a question.
Provide additional context.
Pull in the correct datasets.
Clean data and prepare it for analysis.
Analyze data and create a sharable report.
Dashboards
Users can choose a dashboard or prompt to create anew one.
Interact with the data or ask follow up questions
Request new visualizations and ways to look at the data.
Get instant updates to the dashbaord with language commands.
This evolution allowed Splice AI to serve two distinct user personas:
- Analysts: who needed full visibility, control, and an integrated workspace.
- Business Users: who needed speed, clarity, and actionable insights without deep technical involvement.
Demo
Before building the full UI, we created a terminal-based prototype to validate the multi-agent backend.
In this version, users could upload a dataset, ask a question, and work with the agents (Project Lead, Data Engineer, Cleaning Specialist, Analyst, Visualization Expert) to produce an answer.
The terminal demo gave us a clear, transparent view of how the agents collaborated step-by-step, and it became a powerful way to show early adopters and potential partners how Splice AI could turn a question into an actionable, visualized insight in minutes.
Go to Market
Our go-to-market strategy focused on small and medium-sized businesses that were already using multiple SaaS tools but lacked dedicated analytics resources.
Positioning
We positioned Splice AI as the “in-house analytics team you don’t have to hire”, a way to turn messy, siloed data into clear, actionable insights without technical overhead.
Competitive Analysis
While there were established players in the BI and analytics space, most were too complex, too expensive, or too resource-heavy for SMBs. There were many pieces of teh Modern Data Stack that needed to be stitched together, maintained and the cost of each one added up.
Our Differentiation
- End-to-end automation: Multi-agent system covering discovery, cleaning, analysis, and visualization.
- Dual user modes: Analyst view for full control, business-user view for simplicity.
- Plug-and-play setup: No need for extensive onboarding, infrastructure, or maintenance.
Pilot Program
To validate Splice AI in real-world conditions, we launched a private pilot program targeting a handful of SMBs that matched our ICP. We created a simple landing page with a signup form and reached out through our network, inviting companies to share a real business question they needed answered.
How It Worked
- Participants connected their core data sources (CRM, finance, product usage logs).
- They submitted 2–3 burning business questions via Slack.
- Our multi-agent system ran the end-to-end workflow and delivered results in PDF format, along with visualizations and supporting datasets.
Example Pilot Requests & Outputs
Sports Analytics Startup
- Question: “Which customer segments have the highest likelihood of churning this quarter?”
- Output: Churn risk model + table of high-risk accounts + recommended retention actions.
Packaging Supplier
- Question: “What products have the highest repeat purchase rate over the last 6 months?”
- Output: Ranked product list with repeat purchase percentage, plus a cohort analysis visualization.
Professional Services Firm
- Question: “Which marketing channels bring in the most profitable clients?”
- Output: ROI comparison chart across channels + SQL queries for transparency.
This pilot helped us understand the real world challenges of working with a diverse set of companies with varying data needs.
Conclusion
While pilot users loved the speed and quality of the insights, we faced a recurring challenge:
Every company’s data was different, stored in various formats, spread across multiple tools, and structured in unique ways.
This meant that, despite the automation in our multi-agent system, significant manual setup was still required for each new customer.
Unifying these disparate data sources into a consistent, analysis-ready format proved technically complex.
These challenges highlighted the need for more robust data integration and normalization capabilities before Splice AI could scale seamlessly across customers.
