Building an AI That Tells Me When I’m the Problem

May 2025 | by Indu Arimilli

The Problem: I needed an AI for my own overthinking

Somewhere between finals week and accidentally ghosting three group chats, I realized something: I had no idea when I was the problem.

You know that moment when your group project goes sideways and you start spiraling — “Was I too controlling? Too passive? Too emoji-heavy?” Yeah, that.

I decided to apply product thinking to myself. My hypothesis? Maybe the data (aka my journaling, texts, and to-do logs) could tell me if I’m actually burnt out or just being dramatic.

So, like any normal person with access to AI tools and too much self-awareness, I decided to vibe code an emotional reflection AI, one that gently tells me when I’m spiraling vs. when I have a valid point.

Step 1: User Research (aka me, myself, and I)

In product management, it is called user empathy. In real life, it looked like me interviewing… myself.
I asked:

• “What does ‘being the problem’ actually look like for me?”

• “What signals show I’m stressed or projecting?”

• “What kind of feedback tone would I actually listen to — comforting or brutally honest?”

Then I took a mini usability test with two friends: I shared a few anonymized journal snippets and asked them to label them “reasonable,” “dramatic,” or “needs context.”
(Yes, they laughed way too hard at some of the dramatic ones.)

This became my training dataset.

Step 2: MVP Idea — “SelfScope”

Goal: Help me reflect on my emotional patterns through tone analysis and contextual feedback.

Features list:

1. Journal Input: Daily text or voice note entry.

2. Tone Detection: Uses sentiment + emotion classifiers to detect frustration, guilt, or calm.

3. Feedback Mode: Chooses tone (e.g. empathetic vs. direct) based on your day’s pattern.

4. Pattern Insights: Graphs that show emotional trends over time.

5. Reality Check” button: A one-tap summary: “You’re probably fine.” or “Hey, maybe apologize.”

User Stories:

• As a student juggling classes, I want to see emotional trends over time so I can notice burnout early.

• As someone prone to overanalyzing, I want soft but honest feedback so I can adjust my reactions calmly.

• As a PM-in-training, I want my reflection data to turn into actionable insights (without cringe).

Step 3: The Vibe Code Prompt

Here’s the prompt I gave ChatGPT (yes, I’m literally vibe coding my feelings):

Prompt:
“Build a lightweight Python script using sentiment analysis (HuggingFace or TextBlob) that takes text input, analyzes tone (positive/negative/neutral), and provides one-line feedback based on thresholds. Add personality to responses (supportive, direct, or sarcastic) depending on sentiment patterns over time.”

Step 4: Visuals & UI

Visual Concept: A clean, journal-style interface:

• Minimalist gradient background (calm blues + warm oranges).

• Central text box for journaling.

• Emoji-based tone meter (😇 → 😤).

• Feedback bubble: “You’re valid, but hydrate.”

• Insights chart: emotion trends across days.

(Think Notion meets Clippy, but with boundaries.)

Step 5: Lessons Learned

• Self-research is still user research. You can be both the product and the tester, just don’t skip data validation.

• Tone > accuracy. I didn’t need perfect emotion detection; I needed tone that felt human.

• PM thinking applies to everything. I scoped, prototyped, and iterated on my own self-awareness.

I didn’t just build a tool; I built an emotionally intelligent mirror.
And sometimes, it still tells me: “You’re fine, you just need sleep.”