Danielle Paige, Astrology Influencer
About This Project
Every year, this astrologer ran a live forecast workshop and handed out a PDF guide to her audience. She wanted an AI version that felt like her, not a chatbot bolted onto a PDF. As Conversational Designer and Systems Architect, I built and owned the whole thing end to end, a vibe coded product that made it into production.
The Challenge
The launch date was fixed to her workshop, so we didn't have time to build a full product. A GPT was the practical choice. First we checked the competition: 17 top-ranked astrology GPTs already live in ChatGPT's GPT Store. We ran them against real astrologers to see how accurate they were. None of them were. Wrong house systems, wrong regional and timezone handling, wrong assumptions in the underlying calculations. That was the actual problem to solve. Not the deadline. The math.
The Approach
I recruited a panel of vetted astrologers from her network, across different regions, and used them to build a golden dataset: charts with known-correct outputs to test against. Then I translated each astrological concept into the math behind it. Planetary positions came from Swiss Ephemeris data. Whole sign houses became 30° divisions of the ecliptic. The Ascendant and Midheaven became angle derivations using spherical geometry. Aspects became angular separation measured against orb tolerances. Birth location and time became geocoded coordinates with historical timezone and DST correction.
ChatGPT's context window couldn't process a full ephemeris in one turn, so I built middleware in Python and Flask, connected to the GPT through an OpenAPI schema. It ran every calculation outside the chat and returned finished results, so the model only talked about numbers that were already verified. I tested it with spot-check evals against the golden dataset, regression testing after every engine change, and accuracy benchmarking against the field.
The conversation layer drew on established therapeutic methods: Internal Family Systems, Somatic Experiencing, Polyvagal Theory, Narrative Therapy, and Jungian depth psychology. I mapped each one to a specific part of the experience and built a scaffolding protocol around them: one question at a time, reflection before response, space to answer. The hard rule was guidance over prediction. It reflects the chart back and lets the person decide for themselves.
The Outcome
The GPT launched with the workshop and reached over 3,000 users in its first two months. It became one of her highest-volume products, with its strongest sales on Black Friday, and sold through the end of March as the 2026 forecast.
Before launch, I ran a five-week beta from Black Friday to the workshop. Early feedback showed a 23% complaint rate, almost all of it tracing back to one issue: accuracy across geolocations and timezones, since countries handle time differently. I built an agent to monitor feedback across email, social, and her membership platform, triage issues, and propose fixes for my review. I tested with emulators and VPNs alongside the astrologer panel, rebuilt the middleware's timezone handling to adapt to each user's location, and shipped fixes through Claude Code and Replit. Complaints dropped to 6% by launch.
The forecast is now a product she can repeat every year. The calculation engine and psychological framework are becoming the foundation of her app, a componentized layer that supports the forecast and every other area her brand focuses on.
Danielle Paige, Astrology Influencer
About This Project
Every year, this astrologer ran a live forecast workshop and handed out a PDF guide to her audience. She wanted an AI version that felt like her, not a chatbot bolted onto a PDF. As Conversational Designer and Systems Architect, I built and owned the whole thing end to end, a vibe coded product that made it into production.
The Challenge
The launch date was fixed to her workshop, so we didn't have time to build a full product. A GPT was the practical choice. First we checked the competition: 17 top-ranked astrology GPTs already live in ChatGPT's GPT Store. We ran them against real astrologers to see how accurate they were. None of them were. Wrong house systems, wrong regional and timezone handling, wrong assumptions in the underlying calculations. That was the actual problem to solve. Not the deadline. The math.
The Approach
I recruited a panel of vetted astrologers from her network, across different regions, and used them to build a golden dataset: charts with known-correct outputs to test against. Then I translated each astrological concept into the math behind it. Planetary positions came from Swiss Ephemeris data. Whole sign houses became 30° divisions of the ecliptic. The Ascendant and Midheaven became angle derivations using spherical geometry. Aspects became angular separation measured against orb tolerances. Birth location and time became geocoded coordinates with historical timezone and DST correction.
ChatGPT's context window couldn't process a full ephemeris in one turn, so I built middleware in Python and Flask, connected to the GPT through an OpenAPI schema. It ran every calculation outside the chat and returned finished results, so the model only talked about numbers that were already verified. I tested it with spot-check evals against the golden dataset, regression testing after every engine change, and accuracy benchmarking against the field.
The conversation layer drew on established therapeutic methods: Internal Family Systems, Somatic Experiencing, Polyvagal Theory, Narrative Therapy, and Jungian depth psychology. I mapped each one to a specific part of the experience and built a scaffolding protocol around them: one question at a time, reflection before response, space to answer. The hard rule was guidance over prediction. It reflects the chart back and lets the person decide for themselves.
The Outcome
The GPT launched with the workshop and reached over 3,000 users in its first two months. It became one of her highest-volume products, with its strongest sales on Black Friday, and sold through the end of March as the 2026 forecast.
Before launch, I ran a five-week beta from Black Friday to the workshop. Early feedback showed a 23% complaint rate, almost all of it tracing back to one issue: accuracy across geolocations and timezones, since countries handle time differently. I built an agent to monitor feedback across email, social, and her membership platform, triage issues, and propose fixes for my review. I tested with emulators and VPNs alongside the astrologer panel, rebuilt the middleware's timezone handling to adapt to each user's location, and shipped fixes through Claude Code and Replit. Complaints dropped to 6% by launch.
The forecast is now a product she can repeat every year. The calculation engine and psychological framework are becoming the foundation of her app, a componentized layer that supports the forecast and every other area her brand focuses on.
Danielle Paige, Astrology Influencer
About This Project
Every year, this astrologer ran a live forecast workshop and handed out a PDF guide to her audience. She wanted an AI version that felt like her, not a chatbot bolted onto a PDF. As Conversational Designer and Systems Architect, I built and owned the whole thing end to end, a vibe coded product that made it into production.
The Challenge
The launch date was fixed to her workshop, so we didn't have time to build a full product. A GPT was the practical choice. First we checked the competition: 17 top-ranked astrology GPTs already live in ChatGPT's GPT Store. We ran them against real astrologers to see how accurate they were. None of them were. Wrong house systems, wrong regional and timezone handling, wrong assumptions in the underlying calculations. That was the actual problem to solve. Not the deadline. The math.
The Approach
I recruited a panel of vetted astrologers from her network, across different regions, and used them to build a golden dataset: charts with known-correct outputs to test against. Then I translated each astrological concept into the math behind it. Planetary positions came from Swiss Ephemeris data. Whole sign houses became 30° divisions of the ecliptic. The Ascendant and Midheaven became angle derivations using spherical geometry. Aspects became angular separation measured against orb tolerances. Birth location and time became geocoded coordinates with historical timezone and DST correction.
ChatGPT's context window couldn't process a full ephemeris in one turn, so I built middleware in Python and Flask, connected to the GPT through an OpenAPI schema. It ran every calculation outside the chat and returned finished results, so the model only talked about numbers that were already verified. I tested it with spot-check evals against the golden dataset, regression testing after every engine change, and accuracy benchmarking against the field.
The conversation layer drew on established therapeutic methods: Internal Family Systems, Somatic Experiencing, Polyvagal Theory, Narrative Therapy, and Jungian depth psychology. I mapped each one to a specific part of the experience and built a scaffolding protocol around them: one question at a time, reflection before response, space to answer. The hard rule was guidance over prediction. It reflects the chart back and lets the person decide for themselves.
The Outcome
The GPT launched with the workshop and reached over 3,000 users in its first two months. It became one of her highest-volume products, with its strongest sales on Black Friday, and sold through the end of March as the 2026 forecast.
Before launch, I ran a five-week beta from Black Friday to the workshop. Early feedback showed a 23% complaint rate, almost all of it tracing back to one issue: accuracy across geolocations and timezones, since countries handle time differently. I built an agent to monitor feedback across email, social, and her membership platform, triage issues, and propose fixes for my review. I tested with emulators and VPNs alongside the astrologer panel, rebuilt the middleware's timezone handling to adapt to each user's location, and shipped fixes through Claude Code and Replit. Complaints dropped to 6% by launch.
The forecast is now a product she can repeat every year. The calculation engine and psychological framework are becoming the foundation of her app, a componentized layer that supports the forecast and every other area her brand focuses on.
Danielle Paige, Astrology Influencer
About This Project
Every year, this astrologer ran a live forecast workshop and handed out a PDF guide to her audience. She wanted an AI version that felt like her, not a chatbot bolted onto a PDF. As Conversational Designer and Systems Architect, I built and owned the whole thing end to end, a vibe coded product that made it into production.
The Challenge
The launch date was fixed to her workshop, so we didn't have time to build a full product. A GPT was the practical choice. First we checked the competition: 17 top-ranked astrology GPTs already live in ChatGPT's GPT Store. We ran them against real astrologers to see how accurate they were. None of them were. Wrong house systems, wrong regional and timezone handling, wrong assumptions in the underlying calculations. That was the actual problem to solve. Not the deadline. The math.
The Approach
I recruited a panel of vetted astrologers from her network, across different regions, and used them to build a golden dataset: charts with known-correct outputs to test against. Then I translated each astrological concept into the math behind it. Planetary positions came from Swiss Ephemeris data. Whole sign houses became 30° divisions of the ecliptic. The Ascendant and Midheaven became angle derivations using spherical geometry. Aspects became angular separation measured against orb tolerances. Birth location and time became geocoded coordinates with historical timezone and DST correction.
ChatGPT's context window couldn't process a full ephemeris in one turn, so I built middleware in Python and Flask, connected to the GPT through an OpenAPI schema. It ran every calculation outside the chat and returned finished results, so the model only talked about numbers that were already verified. I tested it with spot-check evals against the golden dataset, regression testing after every engine change, and accuracy benchmarking against the field.
The conversation layer drew on established therapeutic methods: Internal Family Systems, Somatic Experiencing, Polyvagal Theory, Narrative Therapy, and Jungian depth psychology. I mapped each one to a specific part of the experience and built a scaffolding protocol around them: one question at a time, reflection before response, space to answer. The hard rule was guidance over prediction. It reflects the chart back and lets the person decide for themselves.
The Outcome
The GPT launched with the workshop and reached over 3,000 users in its first two months. It became one of her highest-volume products, with its strongest sales on Black Friday, and sold through the end of March as the 2026 forecast.
Before launch, I ran a five-week beta from Black Friday to the workshop. Early feedback showed a 23% complaint rate, almost all of it tracing back to one issue: accuracy across geolocations and timezones, since countries handle time differently. I built an agent to monitor feedback across email, social, and her membership platform, triage issues, and propose fixes for my review. I tested with emulators and VPNs alongside the astrologer panel, rebuilt the middleware's timezone handling to adapt to each user's location, and shipped fixes through Claude Code and Replit. Complaints dropped to 6% by launch.
The forecast is now a product she can repeat every year. The calculation engine and psychological framework are becoming the foundation of her app, a componentized layer that supports the forecast and every other area her brand focuses on.