Home
Progress Counter
Organ Health
Craving Logs
Quit Plan
AI chat bot
Alternate solutions
Book Consultation
Resources

Inspiration

Quitting smoking is rarely a lack of information, it’s a lack of support in the hardest moments. Cravings hit late at night, motivation fades after a few days, and “just one more” feels reasonable when nobody is there to push back.

We built BreathFresh because we wanted a companion that turns quitting into something visible and winnable: track how long you’ve been smoke-free, see your body recovering, log cravings, and talk to an AI coach that never encourages a slip. The goal wasn’t another generic wellness checklist, it was a product that feels present when willpower is lowest.

What it does

BreathFresh is a progressive web app for smoking cessation. After onboarding with smoking history, users get:

Live progress — smoke-free time, cigarettes avoided, money saved, and life regained
Organ health visualization — recovery views for lungs, heart, oxygen, and circulation
Craving log & quit plan — track triggers and generate a personalized 7-day reduction plan
BreathFresh AI — a AI-powered coach with FAQ-grounded answers for cravings, relapse recovery, and motivation
Resources & consultations — curated content, alternatives, and specialist booking
Admin tools — manage users, content, and specialists

How we built it

Frontend

We built the app with React, TypeScript, Vite, Tailwind/shadcn UI, React Query, and Supabase Auth. It ships as a PWA, so it can install on mobile with a native feel and support push notifications.

Backend and AI

A FastAPI service exposes chat and quit-plan endpoints. The chat pipeline works in three steps:

Lightweight FAQ keyword retrieval for grounded context
LLM completion with a strict stop-smoking system prompt
Fallback responses when the model is unavailable

Personalized quit plans are returned as structured JSON: 7-day targets, craving windows, and interventions.

Health recovery math

User profiles and progress live in Supabase. Health recovery is computed from smoking intensity and days since quit.

Pack-years:

$$ \text{PackYears} = \left(\frac{\text{cigarettes/day}}{20}\right) \times \text{years smoked} $$

Smoking intensity (capped at 1):

$$ \text{Intensity} = \min\left(\frac{\text{PackYears}}{40},\, 1\right) $$

Estimated full-recovery horizon in days:

$$ \text{FullRecoveryDays} = 180 + \left(\frac{\min(\text{PackYears},\,40)}{40}\right) \times 1280 $$

Each health metric at day $ t $ recovers from a damaged baseline $ B $ toward 100:

$$ \text{Metric}(t) = B + (100 - B)\cdot\min\left(\frac{t}{\text{FullRecoveryDays}},\, 1\right) $$

Money saved is estimated from daily cigarette use and pack cost over smoke-free time.

Stack

Layer	Tech
App	React, Vite, TypeScript, PWA
Auth / DB	Supabase
AI API	FastAPI + Groq
Visualization	Three.js organ viewers, Recharts
Deploy	Railway (API) + frontend hosting

Challenges we faced

Keeping the AI firmly anti-smoking — LLMs can sound overly “balanced.” We hardened the system prompt so questions like “should I try one?” always get a clear refusal plus coping steps, never permission.
Heavy ML vs. deployability — We explored local RAG with embeddings/FAISS. For reliability and cost, we moved to Groq + keyword FAQ retrieval so the coach stays fast and hostable.
Personalization without false precision — Recovery visuals and pack-year math need to feel meaningful without claiming clinical diagnosis. We framed metrics as motivational progress models, not medical advice.
Full-product surface area — Auth, onboarding, chat history, plans, admin, and PWA install all had to work together. The hard part was polish and consistency, not a single feature.
Windows ML tooling — Early Torch/transformers experiments were painful to set up; simplifying the inference path removed a major blocker.

What we learned

Quitting support is a product design problem as much as an AI problem: timing, tone, and refusal matter.
Grounding the coach with FAQ context improves usefulness without a full heavyweight RAG stack.
Structured JSON quit plans are more actionable than free-form advice.
Making recovery visible (money, organs, milestones) reinforces behavior better than text alone.
Shipping a PWA + cloud API is a practical way to put health tools on phones quickly.

What’s next

Richer craving analytics and smarter notification timing
Stronger retrieval (hybrid search) while keeping latency low
Deeper specialist workflows and localization
Optional habit integrations for real-world trigger detection

BreathFresh is our bet that quitting gets easier when progress is measurable, recovery is visible, and support is available the moment a craving starts.