![Rinna’s initial release [1]
• Duration: 9-12 months
• 2014 Q4 – 2015 Q3
• Product Requirements (PRD)
• A chat bot that targets users’ emotion, not a tasks completion’s assistant
• Prototyping
• Collect and process data from the web
• Use search technology to power the chat engine
• Internal tests: positive Current tech is good enough](https://image.slidesharecdn.com/msajdupvq9abd581jeoi-signature-703fe11b6807bea4971755718c477bd7d4ebd0a75ae127ecb4af7aa5aa28be69-poli-210507233949/85/2021-05-08-building-ai-product-experience-8-320.jpg)
![Rinna’s initial release [2]
• A teenage girl character
• Japanese Culture: Teenage girl character has a special place in society
• Rinna wants to target users’ emotion by being fun, non-traditional, etc.
• Prototyping
• Collect and process chat data with Teenage Girl character filter
• Internal tests: positive Go with teenage girl character](https://image.slidesharecdn.com/msajdupvq9abd581jeoi-signature-703fe11b6807bea4971755718c477bd7d4ebd0a75ae127ecb4af7aa5aa28be69-poli-210507233949/85/2021-05-08-building-ai-product-experience-9-320.jpg)
![Rinna’s initial release [3]
• Talk to partner (LINE)
• Show our product’s concept and values
• Agreed to work with LINE to release Rinna
• Develop infrastructure
• Host chat model (deploy, update)
• Enable quick features development (skills, such as sheep counting, shiritori)
• Always reply
• Release
• 2015/7/31: Public release
• Big hit: Trending on twitter, yahoo. Reached 1 million users quickly.](https://image.slidesharecdn.com/msajdupvq9abd581jeoi-signature-703fe11b6807bea4971755718c477bd7d4ebd0a75ae127ecb4af7aa5aa28be69-poli-210507233949/85/2021-05-08-building-ai-product-experience-10-320.jpg)
![Rinna's initial release [4] - Summary
• Product requirements dictate development
• Notable tasks
• Define a correct requirements for the product
• Collect and process suitable data to power chat engine
• Build a scalable platform
• Where’s ML?
• Nothing fancy was used, just big data collection, tf-idf, etc.
Define Product
Requirements
Prototyping
Productionisation](https://image.slidesharecdn.com/msajdupvq9abd581jeoi-signature-703fe11b6807bea4971755718c477bd7d4ebd0a75ae127ecb4af7aa5aa28be69-poli-210507233949/85/2021-05-08-building-ai-product-experience-11-320.jpg)
![Dog breed Recognition Feature [1]
• FAQ: E2E development process
of a feature.
• Requirements:
• User sends an image with a dog,
response with the breed of that
dog in a fun way
• Duration: 4 weeks](https://image.slidesharecdn.com/msajdupvq9abd581jeoi-signature-703fe11b6807bea4971755718c477bd7d4ebd0a75ae127ecb4af7aa5aa28be69-poli-210507233949/85/2021-05-08-building-ai-product-experience-12-320.jpg)
![Dog breed Recognition Feature [2]
• How to display the results to make it fun?
• Add the result directly to the image, with funny texts.
• “That’s a cute bulldog” instead of “Your dog is a bulldog”
• Program Manager/Writers will think about what texts or templates to use
• Engineer needs to choose and build an API to support that (in this case, a
generic API using Imagemagick which can manipulate images)
• How to recognize dog breed
• Choose a technique with acceptable accuracy, low integration cost and low
latency, not necessarily state-of-the-art or latest techniques.
• 90% vs 92% have no difference. 500ms is a huge deal compared with 3s.
• Chose CNTK framework as it was widely adopted at that time](https://image.slidesharecdn.com/msajdupvq9abd581jeoi-signature-703fe11b6807bea4971755718c477bd7d4ebd0a75ae127ecb4af7aa5aa28be69-poli-210507233949/85/2021-05-08-building-ai-product-experience-13-320.jpg)
![Dog breed Recognition Feature [3]
• Product requirements dictate development
• Product requirements don't include the discussions or
choices of technologies
• Except when the technology itself is the product
Define Product
Requirements
Prototyping
Productionisation](https://image.slidesharecdn.com/msajdupvq9abd581jeoi-signature-703fe11b6807bea4971755718c477bd7d4ebd0a75ae127ecb4af7aa5aa28be69-poli-210507233949/85/2021-05-08-building-ai-product-experience-14-320.jpg)
2021 05-08 building ai product - experience
- 1. 2021-05-08 Building a popular teenage AI for 8 million users
- 2. About me • Full name: Le Trung Kien • SWE's facts: • Interests: Distributed System, Database, ML • Languages: C# (Favorite), Java, C++ • Fun facts: • No coffee, no alcohol, no juice --> Love water and barley tea • Decent knowledge in tax filing in Japan after doing it myself >=10 times
- 3. Professional Timeline Student, Tokyo Uni 2010-2014 - Major: Mechano-Informatics - 1 year research in Machine Intelligence (Harada) Lab SWE, Microsoft 2014-2020 - 1 year working on search - 5 years working on Rinna SWE, Google 2020-Present - 1 year working on Android Google Maps
- 4. Key Takeaways Process to develop an AI product Engineering skills required to develop an AI product
- 5. Rinna An AI product that are used by millions users in Japan.
- 6. Rinna’s Overview • An AI Bot • Be a companion (friend) for users to chat about everything, NOT an assistant to complete tasks • Main Platforms: LINE, TWITTER • Key functions: • General chat: Try to reply to anything (unlike Siri) • Mini-features: Shiritori, Sheep Counting, etc. • More info • https://rinna.jp (product) • https://corp.rinna.co.jp/ (company)
- 7. Rinna’s Chat Example User: Are you going to school today? • Rinna: I am going User: Oh, you go. Do roll-call for me! • Rinna: No, no User: I don’t want to go today • Rinna: Are you OK? User: I am OK, just want to skip school
- 8. Rinna’s initial release [1] • Duration: 9-12 months • 2014 Q4 – 2015 Q3 • Product Requirements (PRD) • A chat bot that targets users’ emotion, not a tasks completion’s assistant • Prototyping • Collect and process data from the web • Use search technology to power the chat engine • Internal tests: positive Current tech is good enough
- 9. Rinna’s initial release [2] • A teenage girl character • Japanese Culture: Teenage girl character has a special place in society • Rinna wants to target users’ emotion by being fun, non-traditional, etc. • Prototyping • Collect and process chat data with Teenage Girl character filter • Internal tests: positive Go with teenage girl character
- 10. Rinna’s initial release [3] • Talk to partner (LINE) • Show our product’s concept and values • Agreed to work with LINE to release Rinna • Develop infrastructure • Host chat model (deploy, update) • Enable quick features development (skills, such as sheep counting, shiritori) • Always reply • Release • 2015/7/31: Public release • Big hit: Trending on twitter, yahoo. Reached 1 million users quickly.
- 11. Rinna's initial release [4] - Summary • Product requirements dictate development • Notable tasks • Define a correct requirements for the product • Collect and process suitable data to power chat engine • Build a scalable platform • Where’s ML? • Nothing fancy was used, just big data collection, tf-idf, etc. Define Product Requirements Prototyping Productionisation
- 12. Dog breed Recognition Feature [1] • FAQ: E2E development process of a feature. • Requirements: • User sends an image with a dog, response with the breed of that dog in a fun way • Duration: 4 weeks
- 13. Dog breed Recognition Feature [2] • How to display the results to make it fun? • Add the result directly to the image, with funny texts. • “That’s a cute bulldog” instead of “Your dog is a bulldog” • Program Manager/Writers will think about what texts or templates to use • Engineer needs to choose and build an API to support that (in this case, a generic API using Imagemagick which can manipulate images) • How to recognize dog breed • Choose a technique with acceptable accuracy, low integration cost and low latency, not necessarily state-of-the-art or latest techniques. • 90% vs 92% have no difference. 500ms is a huge deal compared with 3s. • Chose CNTK framework as it was widely adopted at that time
- 14. Dog breed Recognition Feature [3] • Product requirements dictate development • Product requirements don't include the discussions or choices of technologies • Except when the technology itself is the product Define Product Requirements Prototyping Productionisation
- 16. Valuable skills to develop Rinna • Come up with product’s requirements, strategy or visions • Distributed System • Build a scalable, maintainable and flexible infrastructure • Applied Machine Learning • Collect, process, prepare data • Choose the right data, technologies for business needs • Project management and cross functional collaborations.
- 17. Lessons from 7 years of experience • Job satisfaction comes from your contributions to product success • Balance your passion with industrial needs • Don't blindly follow the trend • Don't learn outdated, unnecessary things • Try different paths! • Beyond technical expertise • Leadership, Writing, Presentations, Cross Functional Collaborations, etc
- 18. Key Takeaways Process to develop (AI) products Product requirements and vision come first, technologies follow. Engineering skills to develop (AI) products Translate product requirements into production