Healify

I've been in this work from three directions: as the person who built the tool, as the clinician who used tools like it with real patients, and as the patient who was tracked by therapists doing the same thing. That triple vantage point is what the 2025 rebuild is built on. This is what I learned.

Healify v1 was a real company. We had co-founders, a Slicing Pie equity arrangement, nearly two years of active development, real therapist relationships, and a platform that delivered over a million mood assessments. We applied for an SBIR grant — the review board gave us a green light on the short form. We applied to YC twice. The first time they told us we were in the top 10% and to apply again. We did.

None of it turned the corner. We pivoted too many times — therapist-side, individual-side, B2B, B2C. The SBIR application stalled when the co-founder running the clinical side got overwhelmed by his own growing practice and couldn't complete the 30-page form. The team drifted away. I was left with a buggy AWS codebase I hadn't built and couldn't maintain. The remaining engineers couldn't publish new versions. The core engineer had left.

I stepped away from Healify to work directly with patients. Not as a research method — as an attempt to understand what I'd been building a tool for. I trained in hypnotherapy under a well-known local teacher, developed a hybrid modality I called Hypnosomatics, and spent the following year treating chronic pain, anxiety, and stress through body-based nervous system work.

In late 2024 I came back. A former user reached out — missed the tool, offered to pay to have it rebuilt. And I'd started to see that building with AI was a real amplifier, not a productivity toy. Healify was the right vehicle: a problem I knew, a codebase I could reason about, and a product I still believed in.

A former associate of my father — David Burns, MD, who founded CBT and whose work Healify had always been built on — organized a 12-week book club cohort around his latest book. Sixty participants: general public and clinicians. A structured program, real check-ins, a twelve-week window. Small enough to instrument carefully. Real enough to actually test.

The hypothesis: does a mood measurement app work as B2C — individual-first, no therapist relationship required? And: what happens when AI generates a personalized summary of your mood patterns over time?

Sixty people checking in daily or weekly. Some had built routines around it. Real dependency, real stakes — these were participants in a therapeutic reading program, not beta testers who could be bounced to a maintenance page.

I discovered that AI had silently rewritten the auth layer during a refactor. All user data was still in the database. None of it was personally identifiable anymore. Sixty people, week four of twelve, and their mood histories were orphaned in a production system I'd built with an AI coding assistant that had quietly changed the architecture underneath me.

I did not stop the cohort.

I spent the next several days tracing how the data had been structured before the rewrite, mapping what was recoverable, and rebuilding the identity layer without disrupting live sessions. Some participants were checking in during the same window I was doing surgery on the database schema underneath them. Data intact; users unaware; cohort continuous.

The recovery taught me something concrete: when you build with AI, the AI doesn't know what it doesn't know about your production context. It can refactor auth correctly in isolation and break it completely in the system. That's not a bug in the AI — it's a constraint you have to architect around. Scope AI changes tightly. Review them at the boundary. Don't let a refactor touch the persistence layer without a before/after audit. I know this now because I had to learn it live, with users in the system.

By week twelve: approximately 500 check-ins across 60 participants. Consistent, genuine usage — not engagement-hacked. People tracking something they cared about, on a cadence they chose.

I built a feature late in the cohort: an AI-generated summary of each person's mood patterns. A simple offer in the app, no heavy promotion. The system used my father's structured mood items — ~25 scored dimensions — as input. Structured data, not chat transcripts. The output was a personalized analysis of patterns across all their check-ins.

More than half of participants requested one.

The reports were five pages. People discovered aspects of their internal life they hadn't articulated before — patterns across weeks they couldn't have seen from inside a single session. The feedback was consistently strong. One participant had 40+ check-ins; the analysis of their data was, by any measure, profound.

Four people converted to paid. Out of sixty initial participants, roughly thirty remained active by week twelve. Four converted. The conversion rate is too small to generalize from and the cohort too narrow to call conclusive.

But all four had received the AI-generated summary.

The conversion correlation pointed at something real: AI can do something therapeutically valuable when it has structured data to work from. Not AI chat. Not generic wellness responses. Pattern recognition across longitudinal structured data, returned as personalized natural language — that's a different thing. That's what people paid for.

That proof of concept is the seed of HabitualOS. If structured data plus AI analysis equals personalized insight that people genuinely value — what does that look like across all daily contexts, not just mood? Not measurement as a product. Measurement as the input to something larger.

Healify taught me where the leverage is. HabitualOS is what happens when you build the full system.