ResearchFlow AI
Inspiration
Research is rarely done in one place.
While working on a research project, we constantly move between tools like Kaggle, Google Colab, GitHub, Google Scholar, arXiv, Gemini, ChatGPT, Google Docs, Google Sheets, and Google Slides. The actual research may take hours, but the work around the research takes even more effort.
After a focused session, we still have to remember what we worked on, update our project tracker, write progress notes, prepare presentation updates, and figure out what to do next.
At the same time, one notification or a quick visit to social media can turn a focused research session into a distracted one.
We asked:
What if an AI agent could take care of the workflow around research, while the researcher focuses on the research itself?
That idea became ResearchFlow AI.
What ResearchFlow AI Does
ResearchFlow AI is an autonomous research workflow agent designed around the Taskmaster concept.
A researcher starts a focused session by specifying their project, objective, and duration.
For example:
Project: WiFi CSI Fall Detection
Objective: Develop a real-time fall detection model using WiFi CSI data.
During the session, ResearchFlow operates in the background and observes high-level browser activity such as visits to Kaggle, Colab, GitHub, Scholar, arXiv, and AI research tools.
At the same time, its Focus Mode can block configured distraction websites such as Instagram, Reddit, X/Twitter, TikTok, and Netflix.
When the session ends, the interesting part begins.
ResearchFlow follows an autonomous loop:
Observe → Understand → Decide → Act → Remember
The goal is not simply to give the researcher another AI chatbot.
The goal is to build an agent that takes action across the research workflow.
How It Works
- Focused Research Session
The user starts a session and defines:
Research project
Research objective
Session duration
Focus Mode
ResearchFlow then creates a session and begins collecting authorized activity metadata.
- Privacy-Preserving Activity Tracking
ResearchFlow intentionally works at the domain metadata level.
It records information such as:
Domain visited
Timestamp
Duration
Activity category
Session ID
Whether the visit was blocked
It does not capture:
Keystrokes
Passwords
Private messages
Form contents
Clipboard contents
Screenshots
Private page contents
Tracking only occurs during an explicitly started research session.
This gives the agent enough context to understand a research workflow without turning it into surveillance software.
- Focus Shield
During an active research session, configured distraction domains can be blocked by the Chrome Extension.
If the user tries to open a distracting website, ResearchFlow can:
Block the navigation.
Display a Focus Mode message.
Record the blocked attempt.
Continue the research session.
- Gemini-Powered Understanding
When the session ends, the collected activity and project context are passed to Gemini.
The AI produces structured analysis covering:
Session summary
Activity breakdown
Accomplishments
Distractions
Inferred progress
Challenges
Next steps
Recommended actions
ResearchFlow also explicitly separates:
OBSERVED — what the system actually knows.
INFERRED — what the AI reasonably concludes from the available evidence.
UNKNOWN — information the system cannot determine.
For example:
Observed: The researcher spent 30 minutes on Kaggle.
Inferred: This is consistent with coding or data-science activity.
Unknown: The exact code written during those 30 minutes.
This prevents the AI from pretending to know things that the system never observed.
The Taskmaster Agent
The core of ResearchFlow is the transition from AI analysis to autonomous action.
Instead of stopping at:
"Here is what you did today."
ResearchFlow asks:
"Based on what happened, what should happen next?"
For example:
SESSION COMPLETE ↓ GEMINI ANALYSIS ↓ AGENT DECISION ↓ "Research progress should be recorded." ↓ Update project tracker ↓ Update research notes ↓ Update project memory ↓ Action complete
The agent can route actions to different tools depending on what happened during the session.
This event-driven workflow is what makes ResearchFlow a Taskmaster, rather than simply a productivity tracker.
Project Memory
Research projects are continuous, so an agent that forgets everything after every session is not particularly useful.
ResearchFlow maintains persistent project context such as:
Project objective
Current research phase
Completed work
Current work
Known challenges
Previous session summaries
Next steps
This allows a future session to begin with context from previous work.
For example:
SESSION 1 ↓ Model development ↓ Progress recorded ↓ Project memory updated
SESSION 2 ↓ Retrieve previous context ↓ Continue from current research state ↓ Analyze new work ↓ Update project again ↓ Remember
ResearchFlow therefore becomes a continuous research companion, rather than a one-time session analyzer.
How We Built It
ResearchFlow combines several components:
React for the web application and command-center interface
Chrome Manifest V3 for browser-level activity monitoring and Focus Mode
Node.js / Express for the backend session and event APIs
Gemini for structured research-session analysis and agent reasoning
Google Workspace APIs for workflow automation across Sheets, Docs, and Slides
Persistent storage for session history and project memory
The application is designed around a modular agent architecture so that browser events, AI reasoning, external tools, and persistent memory can work together.
The Chrome Extension and web application communicate through a backend API, while sensitive credentials and API keys remain server-side.
What We Learned
Building ResearchFlow taught us that creating an agent is very different from creating a chatbot.
A chatbot primarily responds to a user.
An agent needs to:
Observe an event.
Understand the context.
Decide what should happen.
Select the appropriate tool.
Execute the action.
Record the result.
Remember the outcome for future decisions.
We also learned that good agent design requires clear boundaries.
Giving an AI access to everything on a user's computer would make the system powerful, but it would also create serious privacy concerns.
Instead, we designed ResearchFlow around the principle:
Give the agent only the minimum information it needs to perform its task.
That led us to domain-level telemetry, explicit session activation, and clear separation between observed facts and AI inference.
Challenges
Building a Real Agent Instead of a Simulation
One of our biggest challenges was distinguishing between an impressive-looking prototype and an actually functional agent.
It is easy to simulate:
User worked on Kaggle ↓ AI says "Good progress!"
It is much harder to create the complete chain:
Browser Event ↓ Backend ↓ Gemini ↓ Agent Decision ↓ Tool Selection ↓ External Application ↓ Verified Result
We therefore built the system incrementally, validating each layer before moving to the next.
Chrome Extension Constraints
Chrome extensions have their own security and permission model, especially with Manifest V3.
We had to carefully design how the extension could:
Detect active domains
Track activity only during active sessions
Block configured distractions
Communicate with the backend
Persist session state
while maintaining the privacy boundaries of the product.
Keeping AI Grounded
Another challenge was preventing Gemini from hallucinating research progress.
A visit to Kaggle does not prove that a model was successfully trained.
A visit to Google Scholar does not prove that a paper was read.
Therefore, we designed the AI output around:
Observed → Inferred → Unknown
This makes the agent more trustworthy and transparent.
Connecting Multiple Tools
The real value of ResearchFlow comes from its ability to move information between different applications.
Connecting a browser extension, AI model, backend, persistent memory, and Google Workspace introduces authentication, permissions, API, and state-management challenges.
This is also what makes the Taskmaster architecture valuable: the agent is not just generating information; it is intended to route information and actions across the user's workflow.
What's Next
Our vision for ResearchFlow goes beyond a single research session.
We want it to become a complete autonomous research operations layer that can:
Understand long-term research projects
Maintain project memory
Automatically update project trackers
Maintain research notes
Keep presentations synchronized with progress
Recommend the next research task
Coordinate multiple research tools
Learn the structure of different research workflows
Ultimately, we want researchers to spend less time maintaining the machinery around their work and more time doing the work itself.
ResearchFlow AI
You do the research. ResearchFlow handles the workflow.
Built With
- ai-agents
- automation
- chrome
- express.js
- gemini
- generative-ai
- google-ai-studio
- google-workspace
- javascript
- manifest-v3
- node.js
- productivity
- react
- research
Log in or sign up for Devpost to join the conversation.