project-sensei

time management assistant for overwhelmed college students

Comment

LIVE LINK: https://sensei-app.team-deetsuite.workers.dev NOTE: MUST BE AN INVITED TESTER TO DEMO - HIT US UP ON DISCORD!!!! @.skyrmion

Sensei: Hacklahoma 2026 Project Report

Our Story

This team is made up of four nontraditional students. Wellington is a PhD student in Cyber Studies at The University of Tulsa (TU), and originally from Zimbabwe. Mouzam is a Masters student in Business Analytics, also at TU, but originally from Pakistan. Dallas is another PhD student in Cyber Studies at TU, a Tulsa native, but an aspiring world traveler. Mahaelani is a first year graduate student in Computer Science at The University of Oklahoma, originally from Hawaii. Wellington and Dallas knew each other before Hacklahoma, but had never worked together before. Mouzam was new to the group, and almost didn't come because he thought he'd have nothing to offer. Mahaelani showed up late, and walked into the Team Building room, where the three of us from TU were scouting for a fourth member.

After just a few minutes of conversation, it was obvious that we were onto something. What started as a quick chat turned into instant alignment. We came from different backgrounds and different life paths, but we shared the same frustrations, the same curiosity, and the same drive. Within an hour, there was magic happening.

Inspiration

Students are overwhelmed and stressed out. Missed deadlines, time management challenges, constant context switching, planning systems that collapse after a few days. Universities tell students what classes to take, but offer little guidance on managing life. There's no real help balancing study, work, health, and family. No support for discovering paths when you don't even know where you're going. No personal coaching unless you're already doing well.

Meanwhile, executives have assistants and coaches. Students are expected to figure it out alone. Most students who struggle with productivity are dealing with cognitive overload. Planning your life shouldn't feel like another full-time job.

We built Sensei, an AI-powered student assistant that conducts a quick structured intake to understand how each user thinks, what drains their energy, what motivates them, challenges them, what's behind their avoidance. But people change, and Sensei changes with them. When the user is ready, they can re-intake to build a new profile that reflects their current reality. Sensei uses this information to modulate how it talks, what it suggests, and when and how it follows up. Sensei always asks you: What's the one thing that you want to do better in your life? Then, Sensei helps users reflect on why they haven't already done that one thing, and helps them carve out time to take a single step in that direction.

Given the 24-hour constraint, we focused on one thing done well: an AI time-management mentor that helps students plan, focus, and follow through. No wasted time on illusory productivity theater, just something a time-strapped student can sit down with and quickly feel helped by.

How We Built It

Stack

Layer Technology Rationale
Framework TanStack Start 1.159 + Vite 7.3 Full-stack, edge-native, type-safe server functions
UI React 19, Tailwind CSS v4, Framer Motion Industry standard, utility-first styling, smooth animations
Compute Cloudflare Workers Global edge deployment, serverless, generous free tier
Database Cloudflare D1 (SQLite) + Drizzle ORM Edge-native, zero warm-up, type-safe schema
AI Claude Sonnet 4.5 (Anthropic SDK) Strong instruction-following, native tool-use for scheduling
Auth Google OAuth 2.0 + HMAC-signed sessions Single flow grants identity + calendar access
Calendar Google Calendar API v3 Two-way: reads schedule to avoid conflicts, writes events for actions
Linting Biome 2.3.14 Single tool for linting + formatting, 10x faster than ESLint + Prettier
Deployment Wrangler CLI One command: npm run deploy

Development Methodology

The team followed a structured approach to AI-assisted development:

  1. Research: Understand the problem deeply. Read all team members' conversation outlines, define personas, identify edge cases.
  2. Plan: Write detailed specifications before coding, including file paths, function signatures, and conflict resolution logic.
  3. Implement: Hand the plan to Claude with full context; monitor output quality.
  4. Verify: Run Biome linting, TypeScript type checking, and manual testing after every feature.
  5. Commit and deploy: Ship incrementally; iterate on real feedback.

We kept Claude in the role of builder. Humans owned the vision; AI executed with precision.

Architecture

The app runs entirely on Cloudflare infrastructure. A custom server entry (app/server.ts) exports both a fetch handler for HTTP requests (TanStack Start SSR + server functions) and a scheduled handler for cron-triggered follow-ups every 15 minutes. The database is a 7-table SQLite schema (users, profiles, interactions, messages, scheduled actions, reflections, follow-up checks) managed through Drizzle ORM migrations.

Two specialized AI agents power the conversational experience:

  • Sensei Agent: Conducts the main coaching conversation. Has one tool, schedule_action, which creates a Google Calendar event, a reflection reminder, and a follow-up check. Receives rich context including the user profile, recent interactions, upcoming calendar (7-day window), and missed reflections.
  • Reflection Agent: Runs post-action conversations. Has one tool, complete_reflection, which records whether the user completed the action, their emotional state, barriers faced, and whether they want to repeat it.

Both agents use a tool-use loop capped at 3 iterations to prevent runaway behavior.

Timeline

Time Milestone
Feb 7, 12:24 PM Project initialized from template
Feb 7, 9:47 PM First React app draft (Wellington's initial UI)
Feb 8, 5:04 AM Phase 1: Full stack migration (TanStack Start, D1, OAuth, 7-table schema)
Feb 8, 5:24 AM Phase 2: Agent system + Google Calendar integration
Feb 8, 5:51 AM Phase 3: End-to-end wiring (auth guard, intake flow, protected routes)
Feb 8, 6:32 AM Phases 4-6: Biome linting, cron follow-ups, error boundaries
Feb 8, 6:50 AM Phase 7: UI polish (stat cards, skeletons, markdown, typing indicator)
Feb 8, 8:17 AM Deployment complete + first post-deploy fixes
Feb 8, 10:13 AM Six-feature release (dark mode, time awareness, editable actions, calendar context, bold opener, calendar cleanup)

From first framework commit to final feature: 5 hours and 9 minutes of implementation. From project creation to deployed product: ~22 hours.

Features

1. Personalized Intake Flow (10 Steps)

A structured onboarding questionnaire captures the user's goals, energy patterns, avoidance triggers, structure preferences, values, and accessibility needs. Answers feed into a resolution engine (shared/resolve.ts) that computes a persona mode (Coach, Ignition, Pacer, Stabilizer, or Adaptive) and detects contradictions. For example, if someone claims long focus sessions but also reports low energy, the system flags the conflict and adjusts.

This profile directly shapes how Sensei interacts with the user. The resolved intake data is injected into Sensei's context before every conversation, influencing its tone, the kinds of questions it asks, how much structure it offers, and how it frames suggestions. A user in Ignition mode gets quick, energizing prompts. A user in Pacer mode gets steadier, lower-pressure guidance. The intake acts as Sensei's understanding of who you are and how you work best.

Because people change, the intake can be retaken at any time from the profile page. If a user's energy levels shift, their challenges evolve, or they want a different interaction style, they can redo the questionnaire and Sensei's personality updates to reflect the new answers. This keeps the app useful over time rather than locking users into a static profile from day one.

Rationale: Generic advice fails. Understanding how someone thinks before coaching them makes every interaction more relevant, and letting users update that understanding keeps it accurate.

2. Conversational AI Agent (Sensei)

A 249-line system prompt defines Sensei as a curious, nonjudgmental mentor. It learns about the user first, then naturally guides toward scheduling one concrete action. The agent receives the user's full profile, recent history, and current calendar as context, but never explicitly references the profile data. Sensei uses the intake to inform its behavior quietly, so the conversation feels natural rather than formulaic.

Rationale: The invisible goal is a calendar event. If the conversation feels natural and the user wants to schedule something, the app has succeeded.

3. Google Calendar Integration

Sensei creates calendar events for scheduled actions and reflection reminders. The agent reads the user's upcoming 7-day calendar to avoid conflicts. Users can reschedule or delete actions from the dashboard, and changes propagate to Google Calendar.

Rationale: A commitment that doesn't exist on your calendar doesn't feel real. Integration makes Sensei's suggestions tangible.

4. Reflection System

12 hours after a scheduled action, the user receives a reflection prompt. A dedicated Reflection Agent asks how it went, captures completion status, emotional tone, barriers, and whether the user wants to repeat. This data feeds back into future Sensei conversations.

Rationale: Doing the thing is half the value. Reflecting on it honestly, without judgment, is the other half.

5. Automated Follow-Up (Cron)

A Cloudflare cron job runs every 15 minutes. If a user hasn't reflected after the scheduled window, the system uses Claude to compose a context-aware follow-up message tailored to the user's notification tolerance and persona mode.

Rationale: Users forget. A gentle, personalized nudge is better than silence.

6. Dashboard with Filtering

Stat cards (Total, Pending, Completed, Missed) are clickable filters. Action cards show title, description, scheduled time, goal area, and status badges. Pending actions have reschedule (pencil) and delete (trash) buttons. Completed actions link to their reflection.

Rationale: This is mission control. See everything at a glance, act on anything with one click.

7. Profile Transparency

A dedicated profile page shows users exactly how their intake answers were interpreted: persona mode, friction strategy, confidence loop, avoidance reframe, scaffolding approach, and value alignment. Conflict overrides are called out explicitly.

Rationale: Users should see how they're being understood. Transparency builds trust.

8. Dark Mode

Class-based dark mode with localStorage persistence and system preference fallback. Toggle in the header; an inline script prevents flash of wrong theme.

Rationale: Accessibility and user preference. Many students work late.

9. Editable Pending Actions

Inline datetime editing on dashboard cards. Save triggers updateScheduledAction, which updates both the database and Google Calendar events.

Rationale: Life changes. Rescheduling should be easier than deleting and recreating.

10. Time-Aware Scheduling

The agent receives the current timestamp in its context, enabling relative scheduling ("in 2 hours," "tomorrow morning") rather than requiring absolute times.

Rationale: Natural conversation requires temporal awareness.

11. Schedule-Aware Context

The agent receives the user's next 7 days of calendar events, preventing it from suggesting times that conflict with existing commitments.

Rationale: Scheduling over an existing class or meeting would immediately break trust.

Challenges

Scope Management

The initial vision included 5+ agents, gamification, career pathfinding, skill trees, LMS integrations, leaderboards, and Solana NFTs. The team was overwhelmed. The breakthrough came when Dallas asked for help narrowing down, and the decision crystallized: one pillar (time management), one agent (Sensei), one demo-worthy flow. That single scoping decision enabled a polished MVP instead of a fragmented prototype.

Technology Stack Pivot

The team started with Firebase + Vite + React, then realized they needed tighter backend integration. Pivoting to TanStack Start + Cloudflare Workers + D1 + Drizzle mid-hackathon (thanks to Fazil's advice) required learning four new technologies simultaneously. The payoff was a type-safe, edge-deployed, serverless architecture, but the learning curve was steep.

Google Calendar API Subtleties

Calendar integration proved deceptively tricky:

  • Initial implementation used Base64-encoded event URLs that broke; switching to Google's native htmlLink fixed it
  • Multi-account users couldn't access events because token handling needed per-account awareness
  • The calendar scope scared test users during consent; narrowing to calendar.events solved it
  • Deleting actions from the dashboard left orphaned calendar events until cleanup logic was added

Agent Behavior Tuning

Getting the AI agent to behave naturally required iterative prompt engineering:

  • Without time awareness, the agent couldn't schedule "in 2 hours"
  • Without calendar context, it double-booked users
  • Without deduplication logic, follow-up messages in the same conversation created duplicate scheduled actions
  • The agent had to be coached to ask before scheduling, rather than just scheduling immediately

Authentication Edge Cases

The OAuth callback initially used TanStack's navigate(), which didn't trigger a full page reload. The session cookie was set, but the AuthProvider didn't pick it up until the page was manually refreshed. Switching to window.location.href forced a full reload and fixed the issue.

What We Learned

Simplicity wins at hackathons. The final MVP has one agent, one intake flow, one calendar integration, and a clean UI. Every deferred feature would have diluted the demo without adding proportional value.

AI pair programming requires active oversight. The team specified exact file paths, function signatures, and edge cases in plans. We read every line of generated code before committing. When something broke, we debugged together with Claude, understanding root causes rather than patching symptoms.

Transcript-driven development creates accountability. Saving conversation transcripts meant the team could trace every decision, reproduce every bug, and avoid re-solving the same problems. Future developers can understand why the code looks the way it does.

Type safety pays off under pressure. TypeScript + Drizzle ORM caught schema mismatches at build time. Shared types between client and server prevented divergence. When moving fast, the compiler is your best reviewer.

Edge deployment simplifies operations. Cloudflare Workers + D1 meant no database management, no server provisioning, no cold start anxiety. One command (npm run deploy) put the app live. The team spent zero time on infrastructure and all their time on product.

Production

URL: https://sensei-app.team-deetsuite.workers.dev

Stats: 30+ commits, 7 development phases, 8 post-deploy fix sessions, 100+ files, ~2,400 lines of core application code.

Built at Hacklahoma 2026 by the Dream Team: Dallas, Wellington, Mouzam, and Mahaelani.

Built With

Share this project:

Updates