Context AI: From Information to Relationship Intelligence
Inspiration
Every year, millions of students move from structured education into an unstructured world of careers, relationships, and opportunity.
We are taught accounting, engineering, computer science, marketing, and economics. Yet many are never taught how to:
- Understand what they genuinely want from their careers
- Identify the right people to learn from
- Approach someone without sounding transactional
- Turn a first conversation into a meaningful relationship
- Understand another person without making careless assumptions
- Follow up naturally
- Maintain relationships across months or years
- Learn from each interaction
We are repeatedly told, “Networking is important.” But we are rarely taught what good networking actually is.
This matters because relationships do influence opportunity. A large randomized study involving more than 20 million LinkedIn users found that weak professional ties can causally increase job transmission, although the benefit eventually reaches diminishing returns. (Science)
Networking is therefore not merely a social activity. It is part of the infrastructure through which knowledge, mentorship, confidence, referrals, and opportunities move.
However, access to that infrastructure is unequal.
Some students inherit professional vocabulary, introductions, role models, and knowledge of unwritten career rules through their families and communities. Others – especially first-generation students, international students, newcomers, and people entering unfamiliar industries – must construct this understanding themselves.
We may have the same talent and work ethic, but not the same relational map.
The Personal Problem Behind Context AI
I encountered this problem through my own transition from university into professional environments at Microsoft and Restaurant Brands International.
I learned that career development rarely happens through job postings alone. It also happens through conversations:
- A mentor explaining how an industry actually works
- A manager helping someone recognize an overlooked strength
- An alumnus revealing the reality behind a job title
- A peer introducing an opportunity
- A thoughtful follow-up turning one meeting into a continuing relationship
But I also experienced how cognitively difficult these interactions can be.
Before an important conversation, information is usually fragmented across LinkedIn profiles, company pages, emails, calendar invitations, previous messages, notes, and personal memory. Even when the information is available, the harder questions remain:
- What does this information actually mean?
- Why am I speaking with this person?
- What am I genuinely trying to understand?
- What might matter to them?
- Where could our experiences meaningfully intersect?
- How do I remain authentic while still being intentional?
- What should happen after the conversation?
As a Rotman Commerce Career Peer, I supported more than 40 students with career development. Through those conversations, informal interviews, coffee chats, and discussions with classmates and early-career professionals, I repeatedly encountered similar difficulties.
Students often knew that they “should network,” but struggled with:
- Not knowing whom to approach
- Feeling intimidated by senior professionals
- Sending generic outreach messages
- Asking questions already answered online
- Over-preparing and making conversations feel like interviews
- Worrying that their motives would appear transactional
- Failing to connect their own experience to the other person
- Forgetting what was discussed afterward
- Losing momentum after a promising first interaction
- Repeating the same context to different AI tools
I initially saw this as a conversation-preparation problem. The deeper I explored it, the more I realized it was a relationship-intelligence problem.
The Core Insight
Most existing AI tools begin with the other person:
Who are they? Where do they work? What have they accomplished?
That information is useful, but incomplete.
A meaningful interaction depends on three forms of understanding:
$$
\text{Relational judgment}
\text{Know yourself} + \text{Know the other} + \text{Know the situation} $$
This became the foundation of Context AI.
The idea draws from Sun Tzu’s principle of knowing oneself and knowing others, but it does not treat relationships as battles. Context AI is not designed to help users manipulate people or manufacture intimacy.
It is designed to help people see more clearly:
- What they truly need
- What is known and unknown about the other person
- What the relationship can realistically support
- Where mutual value may exist
- What action would be appropriate at this moment
What Context AI Does
Context AI is a relationship-intelligence partner that converts fragmented personal and professional context into practical understanding.
1. Know Yourself
A user may ask:
“Give me five questions for my coffee chat.” However, that request may contain several deeper needs:
- Functional: Avoid awkward silence
- Emotional: Feel prepared and intellectually confident
- Relational: Build trust without appearing transactional
- Strategic: Understand a career path or preserve a future opportunity
- Identity: Be seen as thoughtful and capable
- Developmental: Become more comfortable navigating professional relationships
Instead of immediately generating generic questions, Context AI identifies the possible underlying need and asks the user to confirm it.
The goal is not to psychoanalyse the user. It is to develop and test a more useful understanding of what they are trying to accomplish.
2. Know the Other Person
Context AI separates information into:
- Verified facts
- Expressed priorities
- Observable behaviour
- Professional incentives and constraints
- Working hypotheses
- Important unknowns
This distinction is essential.
For example, mentoring students is evidence that someone has spent time mentoring. It does not automatically prove that the person is generous, approachable, or motivated by altruism.
Context AI is instructed to distinguish facts from interpretation, communicate uncertainty, identify outdated information, and avoid constructing confident psychological profiles from limited public data.
3. Know the Situation
The same question can be appropriate in one relationship and awkward in another.
Context AI therefore considers:
- How the people know each other
- Whether this is a first interaction or an established relationship
- Previous conversations
- Current trust
- Power and status differences
- Shared interests
- Commitments and unfinished follow-ups
- Timing and setting
- What each person can genuinely offer
- What has changed since the previous interaction
This creates a model not only of two people, but of the relationship between them.
4. Form an Interaction Thesis
Rather than defaulting to a long profile or ten suggested questions, Context AI produces an interaction thesis:
- What the user appears to need beneath the request
- What may matter to the other person
- What is verified versus inferred
- Where their interests align
- Where tension or uncertainty exists
- What posture the user should adopt
- What the user should listen for
- What the most natural next step may be
Questions and message drafts come after this understanding, not before it.
5. Learn Through Debriefing
The long-term vision is a closed learning loop:
$$ \text{Understand} \rightarrow \text{Prepare} \rightarrow \text{Interact} \rightarrow \text{Debrief} \rightarrow \text{Remember} \rightarrow \text{Improve} $$
After an interaction, Context AI can help the user reflect on:
- What surprised them
- What energized or disengaged the other person
- What changed their original interpretation
- What they learned about themselves
- What was promised
- What should be remembered
- What they would do differently
- What next step would feel natural
This transforms networking from a series of disconnected transactions into a process of relational learning.
Who It Is For
Context AI begins with students and early-career professionals, but the underlying problem is broader.
Potential users include:
- University and college students
- First-generation students
- International students and newcomers
- Recent graduates
- Career switchers
- Professionals entering unfamiliar industries
- Founders speaking with customers, investors, and partners
- Sales professionals preparing for relationship-based conversations
- Mentors and career advisors supporting multiple people
Each group faces the same underlying challenge: information about people is abundant, but meaningful relational context is fragmented.
How We Built It
The current prototype was created using Google Cloud Agent Platform’s Agent Studio with:
- Gemini Flash
- Google Search
- URL Context
- A structured system-instruction layer
- Agent Studio’s visual flow and preview environment
The system was designed around six modes:
- Self-discovery
- Person understanding
- Interaction preparation
- Post-interaction debrief
- Follow-up
- Relationship reflection
The initial instruction set defined:
- How to identify deeper user needs
- How to distinguish facts from hypotheses
- How to assess relationship stage
- How to conduct responsible professional research
- How to produce natural rather than performative conversation guidance
- How to challenge unsupported user assumptions
- How to prioritize mutual value and user development
Testing the Prototype
Trial 1: Career Development and Networking
I asked Context AI to understand my background, help clarify my career direction, identify the types of people I should connect with, and explain how those relationships could support my development.
This tested whether the agent could move from user context to personalized career and networking strategy.
Trial 2: Course Research and Professor Outreach
I provided a course name and professor information without supplying the course URL. I asked the agent to:
- Research and explain the course
- Understand the professor’s background
- Discover public discussion about the course
- Identify meaningful conversation topics
- Draft an application-related outreach email
This tested independent research, social listening, evidence synthesis, and conversion of research into practical action.
Challenges We Faced
Interface and Capability Mismatch
When asked whether images could be uploaded, the agent suggested using an attachment icon or drag-and-drop feature. Neither was available in the interface.
This demonstrated that an agent must understand its actual environment rather than describe capabilities that may exist in other AI products.
Restricted Access to Professional Platforms
The agent could not reliably access LinkedIn profiles, possibly because of login requirements, platform restrictions, privacy protections, or limitations in URL Context.
It performed better when information was supplied directly, but this reduced independent research capability.
Weak Independent Social Listening
The agent could analyze supplied sources, but struggled to independently discover diverse conversations across forums, news, Reddit, and professional communities.
This revealed an important difference between analyzing a URL and conducting open-ended research.
Lost Conversation History
Refreshing the interface caused previous context to disappear, with no clear recovery mechanism.
For most chatbots, this is inconvenient. For Context AI, it undermines the central value proposition because relationship intelligence depends on continuity.
Prompt Complexity
The first system instruction was approximately 2,500 words. Although it told the model to avoid excessive headings and checklists, its own highly structured format encouraged the model to produce long, heavily segmented answers.
The prompt also asked Gemini Flash to perform several difficult tasks simultaneously:
- Infer deeper needs
- Evaluate evidence
- Communicate uncertainty
- Understand relational dynamics
- Select an interaction mode
- Produce concise guidance
This taught me that more instructions do not necessarily create better intelligence.
What We Learned
A Prompt Is Not a Product
The first version of Context AI was still close to a specialized chatbot. It could produce polished answers, but it did not yet possess durable memory, session recovery, or an ongoing model of the user.
A genuine Collaborative Partner must become more useful through interaction.
Memory Must Be Functional
Telling an agent to remember something does not create memory. Persistent personalization requires technical infrastructure for storing, retrieving, correcting, and deleting user-approved information.
Research Requires Boundaries
A relationship-intelligence product needs particularly careful safeguards.
Context AI should support legitimate professional, educational, and consensual contexts. It should not:
- Track or monitor individuals
- Research people who have withdrawn contact
- Infer protected or highly sensitive characteristics
- Collect irrelevant personal information
- Present speculation as psychological fact
- Optimize for manipulation
Relationship Intelligence Should Reduce Inequality, Not Intensify It
The deeper opportunity is not to help already-powerful users extract more from their networks.
It is to make some of the unwritten knowledge of relationship-building more accessible to people who did not inherit it.
Context AI cannot give every student the same network. It can help more students understand how relationships form, approach others with greater confidence, recognize mutual value, and learn from their experiences.
What’s Next
1. A Leaner and Stronger Agent
The system instruction will be shortened by approximately 40%, with:
- Clearer priorities
- Explicit search triggers
- Stronger research boundaries
- Fewer conflicting directions
- Several carefully designed examples
- More concise output defaults
2. Persistent Sessions and Memory
The next technical milestone is integrating Agent Platform Sessions and Memory Bank.
Context AI will maintain four user-controlled forms of context:
- User model: goals, preferences, strengths, constraints, and development areas
- Person model: verified facts, expressed priorities, observations, and hypotheses
- Relationship model: interactions, shared themes, commitments, and next steps
- Learning model: user corrections, successful advice, rejected interpretations, and changes over time
The critical test is whether a debrief from one conversation produces visibly better advice in a completely new session.
3. A Chrome Side-Panel Extension
Context currently lives where the user is already working: professional profiles, email, course pages, company websites, meeting notes, and documents.
The next interface will therefore be a Chrome side-panel extension.
Users will be able to:
- Deliberately select relevant webpage content
- Preview what will be shared
- Explain their goal and relationship context
- Choose an interaction mode
- Receive an interaction thesis
- Debrief afterward
- Approve what should be remembered
The extension will use explicit user selection rather than automatic background scraping.
4. Opt-In Productivity Integrations
Future integrations may include:
- Gmail for previous correspondence and follow-up
- Google Calendar for upcoming-interaction preparation
- Google Drive and Docs for meeting notes
- Conversation role-play for interviews and networking
- Career-service tools for advisor-supported student development
Vision
Context AI began as a tool for preparing better coffee-chat questions.
It evolved into a broader question:
Can AI help people develop the self-understanding, social context, and relational judgment that are often learned only through privilege, mentorship, and years of trial and error?
The goal is not to automate human relationships.
It is to help users participate in them more thoughtfully.
The strongest version of Context AI should not make people permanently dependent on generated scripts. It should help them become better at understanding themselves, perceiving others carefully, navigating uncertainty, and building authentic relationships on their own.
Built With
- ai
- cloud
- gemini
- research
Log in or sign up for Devpost to join the conversation.