Back in 2023 when OpenAI first shipped tool use, I wanted an excuse to build something with it. I landed on a day planner. I’d dump my thoughts, plans and goals into GPT-3.5, it would read my Google Calendar and schedule my day around them. Fun little project. If I mentioned I’d been tired the day before it would always give me nap time after lunch, very cute.
I got almost nothing out of it and dropped it after a week or two.
But something else happened. To feed the thing I had to write about my day, every day. And at some point I realized I was journaling for the first time in my life. I’d always wanted to. It never stuck, because it never gave me anything back.
That was the whole insight, and everything since has been chasing it: journaling needs a value add.
What I actually wanted
Sending an entry to 3.5 and getting JSON back was fun, but it was so bad. I had way too many data types, my instructions were too complex, and most importantly I was telling OpenAI my daily schedule and habits for free.
Still, the shape was there. If I could reliably turn “slept badly, skipped breakfast, two coffees, ran 5k, knee hurt again” into structured data, I’d have something worth writing into every day. Not a diary. A queryable record of everything I do, eat and consume, built out of normal sentences.
That’s tinh. Say what happened, it files it.
The local model years
I kept journaling and tweaking the 3.5 setup until Llama 2 came out. Like everyone else I was thrilled to have a model running locally, so I immediately wired it up and rewrote all my prompts.
It was incredibly slow. I’d let it run overnight on my M1 MacBook Air, then in the morning I’d manually correct the JSON before committing it to my database. Terrible user experience. But I saw the light immediately.
I rode the early local LLMs up to around Gemma 3. By then I knew there was a product in here, and a product meant it had to run on a phone, and that meant the smallest models available, 4b parameters and under. Which meant token efficiency and model accuracy were everything.
Six months of not using an LLM
Around summer 2025 I started building pre-filters. I learned spaCy for named entity extraction and some tricks with dual threshold embeddings. I spent about six months on traditional NLP for one reason: cut the number of tokens reaching the model, and cut the complexity of the prompt.
Then Anthropic doubled usage limits over the holidays around the Opus 4.5 release, and I thought, why not see how far I can push this.
Turns out I could push it all the way.
December 2025: about 8 minutes to process a full journal entry end to end, still needing a manual correction at the end, needing around 4GB of RAM.
January 2026: 180ms per sentence, under 200MB of RAM.
Absolutely insane speedup. I literally couldn’t sleep for a few days after getting it wired up.
The part where it clicked
That speed changed what the product was. It wasn’t going to be a batch of text at the end of the day anymore. It was going to be short voice notes, throughout the day, on a phone.
Basically the holy grail of habit tracking.
And the stack could take it; Any data type, updates from user corrections without retraining anything. Fast enough that before I finished saying a sentence it was already categorized and waiting for me to confirm it.
For years I’d been writing everything out, waiting overnight, then checking in the morning to see what stupid mistakes the model had made. Now I had a colorful UI reacting to my voice, and I could confirm data was correct on the spot. It was actually a good tech demo.
So I scrapped the old project, moved to Swift, leaned on Apple’s ML stack and went to town. Last six months have been learning Swift and porting the whole thing to iPhone.
What people are finding in their own data
The private beta has been running for over a year, and the last cohort finally stuck around for continual use, which is the number I’ve been waiting on this entire time.
What I didn’t expect was the messages. People telling me about correlations the app surfaced that they’d never have found by hand. Triggers for chronic illnesses. Patterns across months of small stuff nobody would ever sit down and tabulate manually.
“tinh helped me figure out that barometric pressure has a huge impact on my joint pain.”
That’s the thing about logging everything. Individually none of it matters. In aggregate, over 600+ days, it’s the only honest record of my lif I’mm ever going to have.
tinh v1, September 2026
Coming to the App Store this September.
Still fully private, everything on device. One time fee for lifetime use, plus a year of updates.
I use it all day every day and it feels like a thing. A real thing.
It feels alive.