Mobile · Consumer
A dating app engineered for real conversations
Swipe-based matching, realtime chat and trust & safety — built to turn matches into conversations rather than dead ends.
- React Native
- Node.js
- PostgreSQL
Mobile · AI
An AI astrology app that explains the chart, not just the horoscope
A deterministic ephemeris engine computes the chart; the language model only narrates what was actually computed — personal readings that stay internally consistent.
- Client
- Consumer wellness startup
- Industry
- Consumer wellness
- Timeline
- 4 months to launch
- Platforms
- iOS · Android
Overview
Most astrology apps generate text that sounds personal but is not — the same paragraph for everyone born in a month. We split the problem: astronomy is computed exactly, and the model is allowed to explain those computed placements and nothing else.
The challenge
Handing the whole job to a language model produces readings that are fluent, generic and contradict themselves between sessions. The astronomy has to be exact and the narration has to stay anchored to it.
- Generic sun-sign readings gave users no reason to return
- A model left to itself invented placements and contradicted yesterday's reading
- Chart maths must be precise for birth time, timezone and historical daylight saving
- A long generated reading per user per day is expensive at consumer scale
Our approach
How we built it
-
01
Compute the chart exactly
A Swiss Ephemeris service resolves birth time, place, timezone and historical DST into precise planetary positions, houses and aspects. This is arithmetic, and it is never delegated to a model.
-
02
Narrate only what was computed
The model receives the computed placements as structured facts and explains them in plain language. It cannot introduce a placement that is not in the chart.
-
03
Stay consistent over time
Each user's chart interpretation is generated once and cached, so daily readings build on a stable foundation instead of re-inventing the personality every morning.
-
04
Control the unit economics
Chart interpretations are cached per user and transit narratives per placement bucket, cutting per-reading model cost by roughly an order of magnitude versus naive generation.
-
05
Earn the notification
Daily prompts reference the actual transit driving them, so the push has a reason to exist beyond re-engagement.
What we shipped
In the delivered product
Precise natal chart from birth date, time and place
Plain-language explanations of placements, houses and aspects
Daily and weekly transit readings personal to the chart
Compatibility (synastry) between two charts
Conversational follow-up questions about your own chart
Subscription tier with deep-dive reports
The outcome
Results
~10x
Lower model cost per reading via caching
3.1x
Day-30 retention vs the generic-text baseline
<2s
Chart computation and first reading
0
Placements invented outside the computed chart
More work
View all projects
Other projects
AI · Public sector
Natural-language reporting across a government data warehouse
Analysts ask a question in plain language; the system writes verified, read-only SQL against a governed warehouse and returns the answer with its workings.
- Python
- FastAPI
- Self-hosted LLM
AI · Customer support
A website assistant that answers only from your own content
A retrieval-grounded support assistant that cites its sources, knows when to escalate, and never invents an answer.
- Node.js
- LLM API
- pgvector
Ready to build something people trust?
Book a free consultation. We'll map the highest-impact opportunity for your business — no obligation.