Project 01 — Making a live event easy to follow
Finals Central
Built for the person who doesn't follow basketball but still wants to know what happened — every game translated into plain, story-first language, with an AI narrator that reads the rundown aloud like a friend catching you up.
No experience required
Plain-language recaps
AI narration
Shipped on Vercel
🏆 2026 NBA CHAMPIONS 🏆
New York Knicks
NBA Champions · 2026 · Won in 5 games
0
NYK
–
0
SAS
Finals MVP · Jalen Brunson · 31.6 PPG · 5 games
The Rundown
Overview
Head to Head
Narrative
Players
No basketball knowledge required
Most "AI-powered" sports recaps just ask a model to write about a game and hope the tone lands. The design problem here was narrower and more deliberate: keep a strict source of truth — the real box score, real quotes, real sequence of events — and use AI for exactly one job, translating that into a story a non-fan can follow, without inventing anything along the way.
How the AI's role is scoped
Editor, not reporter — the model restructures real game data into "what happened, the hero, the moment that decided it." It organizes and explains facts that already happened; it never generates ones that didn't.
Progressive disclosure — The Rundown is the default, zero-context layer. Overview, Head to Head, Narrative, and Players exist for anyone who wants to go deeper, without forcing that depth on a first-time reader.
Multimodal by design — the same recap can be read or listened to. AI narration exists because "catching someone up" is naturally something a friend says out loud, not just something they type.
Judged on comprehension, not stat density — "94–90 Knicks win" becomes a story about a team down 16 that clawed back. Success is a non-fan understanding what happened, not how many numbers made it onto the screen.
The throughline with Ataraxia
Same underlying principle, different domain: AI's job is translation, not authority. Ataraxia never diagnoses, it explains a pattern in your own data. Finals Central never invents a storyline, it explains a game that already happened. Neither system gets to make things up — it only gets to make things clearer.
Shipped, not simulated
Built and deployed live on Vercel for the 2026 Finals — a real, running product, not a mockup.
Project 02 — A framework for interactable AI
Ataraxia
Every product with AI baked into the interface has to answer one question: how much authority does the AI get, relative to the person using it? Ataraxia is my worked answer — a set of interaction techniques, validated across 14 structured interviews, for building AI that stays useful without overstepping.
AI product design
Interaction framework
User research
Working prototype
0
wanted earlier pattern detection
0
already wear a fitness tracker
9:41
Good morning
Alex
Tuesday, Mar 14
Weekly summary
Steps down 25%, sleep stable — Tuesday was your toughest day
Low activity on Tue & Thu correlates with your shorter sleep nights. One short walk today could shift the pattern.
Learn more with AI →
1,512
Steps avg
↓ 25%
6.8h
Sleep avg
stable
4
Workouts
↑ 1
Steps
↓ 25%
1,512/day
Goal: 8k
Sleep
Stable
6.8hrs
Score: 84
Workouts
↑ 1
4this wk
On streak
Heart rate
↑ 4
72bpm
Resting
AI suggestions are advisory only — always yours to dismiss or question.
Why bad Tuesday?
Sleep + steps?
Improve week
9:41
Steps
Daily steps — tap a bar
Steps
7-day avg
1,512
Avg
↓ 25%
3,840
Best
Fri
8k
Goal
19%
Ataraxia AIadvisory only
Step count dropped 30% this week. Your two lowest days both followed short-sleep nights — there's a clear pattern worth addressing.
Helpful
Not relevant
Steps AI
Tap any bar to explore that day, or ask me anything about your steps this week.
AI insights are suggestions — you decide what to act on.
Why was Tuesday so low?
What's my step pattern?
How do I hit 8k?
Best vs worst day
9:41
Sleep
84
Sleep score
Nightly stages — tap a night
Deep
REM
Light
Avg duration
6.8h
+12m vs prior wk
Score
84
Optimal
Deep sleep
18%
↓ 5%
REM avg
24%
On target
Sleep AI
Tap any night to see details, or ask me about your sleep patterns.
AI insights are suggestions — you decide what to act on.
Why was Tuesday so bad?
How do I get more deep sleep?
Best night this week
Caffeine + sleep link?
9:41
Ataraxia AI
Online · advisory only
Hi Alex. Steps and energy dipped this week. What would you like to explore?
Break it down
Show numbers
Ataraxia AI · advisory only, not a diagnosis
Why did my energy drop Tuesday?
Tuesday your sleep was 5.2h — nearly 2h below average. Short sleep is the strongest predictor of low energy in your data. Elevated caffeine that evening likely delayed sleep onset.
View sleep data
What should I do?
Ataraxia AI · advisory only
Suggestions are advisory only — always yours to question or dismiss.
9:41
MetricsMar 10–16
Steps
↓ 25%
1,512/day avg
Below 8k goal · tap to explore
Sleep
Score 84
6.8havg
Stable · tap to explore
Correlations this week
Steps → energy
+0.72
Sleep → recovery
+0.80
Caffeine → sleep
−0.61
HR → sleep quality
−0.55
Metrics AI
Ask about any metric, correlation, or how this week compares to your history.
AI insights are suggestions — always yours to question or dismiss.
What's my biggest pattern?
Compare to last week
Why is HR elevated?
What should I focus on?
9:41
Workouts4 this week
Morning Run
Moderate
154 bpm
380 kcal
5.2k
Dist
32m
Time
5:10
Pace
138ft
Elev
Basketball Practice
High
168 bpm
642 kcal
74m
Time
178
Max HR
88
Recovery
High
Intensity
Yoga Flow
Recovery
98 kcal
45m
Time
88
Avg HR
Low
Intensity
94
Recovery
9:41
Log
Optional — all inputs improve AI context
Lifestyle
Caffeine
Not logged
+ Add intake
Nutrition
Not logged
+ Log meals
Screen time
Not logged
+ Add hours
Energy level
Not logged
+ Log feeling
How are you feeling?
Low
Okay
Good
Great
Amazing
Open reflection
Save & get AI insight →
All inputs are optional. You're always in control of what you share.
AI Intelligence
"Resting HR up +4 bpm this week — likely linked to reduced sleep quality. Want to explore?"
Take a look →
Not now
A central challenge in designing AI-powered products is deciding how much authority the system gets, relative to the person using it. Excessive automation erodes agency and breeds overreliance; too little, and the AI never delivers real value. Research on algorithm aversion backs this up: people lose confidence in AI fast after one visible mistake, but people who trust it too completely outsource judgment they shouldn't. Ataraxia's interaction model is built to sit between those two failure modes.
0
Structured interviews conducted
0
Most common confidence score, out of 5
0
Wanted earlier pattern detection
Five techniques for building interactable AI
Progressive disclosure — home screen shows 3 metrics and a simple summary; deeper AI observations, day-level detail, and full chat only surface when the person asks for them.
Human-in-the-loop feedback — every AI suggestion carries "Helpful" / "Not relevant" buttons, so the system's authority is never assumed, only earned.
Mixed interaction modes — casual users get a passive weekly summary; engaged users get hourly breakdowns and free-form chat. Same product, two depths, neither shallow nor overwhelming.
Transparent explainability — the AI never says "your energy was low." It says "your steps dropped 25% on the same days your sleep was shortest" — a claim you can check against your own data.
Direct manipulation over gatekeeping — every chart stays tappable and readable on its own. The AI is an optional layer on top of the data, never the only way to read it.
Grounded in the interviews
Finding: data is visible, meaning isn't → Decision: explain why, not just what
Finding: "somewhat personalized" was the ceiling → Decision: compare to self, never to population
Finding: burnout recognized only in hindsight → Decision: flag shifts the moment they're meaningful
The resulting trust architecture
Observational language — "may help," never "you should"
Persistent disclaimers — advisory only, always dismissable
Data provenance — every insight cites the user's own history
Safety guardrails — distress language routes to real resources
User agency — all logging optional, nothing forced
Calibrated confidence — correlations shown as numbers, hedged
"A doctor or a real human feels more trustworthy than a machine."— Participant A, on why AI insights need visible certainty, not conclusions