Posted by Sa-ryong Kang and Caren Chang, Developer Relations Engineers
Kakao Mobility is South Korea’s main mobility enterprise, providing a variety of transportation and supply companies, together with taxi-hailing, navigation, bike and scooter-sharing, parking, and parcel supply, by its Kakao T app. The crew at Kakao Mobility utilized Gemini Nano by way of ML Package’s GenAI Immediate API to supply parking help for its bike-sharing service and an improved deal with entry expertise for its navigation and supply companies.
The Kakao T app serves over 30 million complete customers, and its bike-sharing service is one in all its hottest companies. However sadly, many customers had been improperly parking the bikes or scooters when not in use. This conduct led to an inflow of parking violations and security issues, leading to public complaints, fines, and towing. These points started to negatively have an effect on public notion of each Kakao Mobility and its bike-sharing companies.
“By leveraging the ML Package’s GenAI Immediate API and Gemini Nano, we had been in a position to rapidly implement options that enhance social worth with out compromising consumer expertise. Kakao Mobility will proceed to actively undertake on-device AI to offer safer and extra handy mobility companies.” — Wisuk Ryu, Head of Consumer Improvement Div
To handle these issues, the crew initially designed a picture recognition mannequin to inform customers if their bike or scooter was parked accurately in response to native legal guidelines and security requirements. Operating this mannequin by the cloud would have incurred important server prices. As well as, the customers’ uploaded pictures contained details about their parking location, so the crew needed to keep away from any privateness or safety issues. The crew wanted to discover a extra dependable and cost-effective resolution.
The crew additionally needed to enhance the entity extraction expertise for the parcel supply service inside the Kakao T app. Beforehand, customers had been in a position to simply order parcel supply on a chat interface, however drivers wanted to enter the deal with into an order type manually to provoke the supply order—a course of which was cumbersome and liable to human error. The crew sought to streamline this course of, making order varieties quicker and fewer irritating for supply personnel.
Enhancing the consumer expertise with ML Package’s GenAI Immediate API
The crew examined and in contrast cloud-based Gemini fashions in opposition to Gemini Nano, accessed by way of ML Package’s GenAI Immediate API. “After reviewing privateness, value, accuracy, and response velocity, ML Package’s GenAI Immediate API was clearly the optimum selection,” stated Jinwoo Park, Android utility developer at Kakao Mobility.
To handle the difficulty of improperly parked bikes or scooters, the crew used Gemini Nano’s multimodal functionality by way of the ML Package GenAI API SDK to detect when a motorbike or scooter violates native rules by parking on yellow tactile paving. With a fastidiously crafted immediate, they had been in a position to consider greater than 200 labeled photos of parking pictures whereas frequently refining the inputs. This analysis, measured by well-known metrics like accuracy, precision, recall, and the F1 rating, ensured the characteristic met production-level high quality and reliability requirements.
Now customers can take a photograph of their parked bike or scooter, and the app will inform them whether it is parked correctly, or present steerage if it isn’t. Your complete course of occurs in seconds on the gadget, defending the consumer’s location and knowledge.
To create a streamlined entity extraction characteristic, the crew once more used ML Package’s GenAI Immediate API to course of customers’ supply orders written in pure language. If they’d employed conventional machine studying, it will have required a big studying dataset and particular experience in machine studying. As an alternative, they might merely begin with a immediate like, “Extract the recipient’s title, deal with, and telephone quantity from the message.” The crew ready round 200 high-quality analysis examples, and evaluated their immediate by many rounds of iteration to get one of the best end result. The simplest technique employed was a method known as few-shot prompting, and the outcomes had been fastidiously analyzed to make sure the output contained minimal hallucinations.
“ML Package’s Immediate API reduces developer overhead whereas providing robust safety and reliability on-device. It permits speedy prototyping, lowers infrastructure dependency, and incurs no further value. There is no such thing as a cause to not advocate it.” — Jinwoo Park, Android utility developer at Kakao Mobility
Delivering huge outcomes with ML Package’s GenAI Immediate API
In consequence, the entity extraction characteristic accurately identifies the mandatory particulars of every order, even when a number of names and addresses are entered. To maximise the characteristic’s attain and supply a strong fallback, the crew additionally applied a cloud-based path utilizing Gemini Flash.
Implementing ML Package’s GenAI Immediate API has yielded a big quantity of value financial savings for the Kakao Mobility crew by shifting to on-device AI. Whereas the bike parking evaluation characteristic has not but launched, the deal with entry enchancment has already delivered glorious outcomes:
-
Order completion time for supply orders has been lowered by 24%.
-
The conversion fee has elevated by 45% for brand spanking new customers and 6% for present customers.
-
Throughout peak seasons, AI-powered orders enhance by over 200%.
“Small enterprise house owners specifically have shared very constructive suggestions, saying the characteristic has made their work way more environment friendly and considerably lowered stress,” Wisuk added.
After the picture recognition characteristic for bike and scooter parking launches, the Kakao Mobility crew is keen to enhance it additional. City parking environments may be difficult, and the crew is exploring methods to filter out pointless areas from photos.
“ML Package’s GenAI Immediate API presents high-quality options with out further overhead,” stated Jinwoo. “This lowered developer effort, shortened total improvement time, and allowed us to concentrate on immediate tuning for higher-quality outcomes.”
Attempt ML Package’s GenAI Immediate API for your self