How Koo thinks about his tech stack

Launched in 2020, local microblogging platform Koo has grown at a surprisingly fast pace, given that its competitor is the ubiquitous Twitter. However, for the technical team behind the application, “house” does not indicate archaic in any way. In a statement released to PTI, co-founder and CEO Aprameya Radhakrishna said the app has seen its user base grow 10-fold over the past year and expects that figure to surpass 100 million by next year.

Analytics India Magazine caught up with the app’s chief technology officer, Phaneesh Gururaj. Gururaj joined Koo just over a year ago after a seven-year stint as a data science product manager at redBus. Here are the excerpts:

Phaneesh Gururaj, CTO at Koo

Tech stack

Under Gururaj, platform technology evolved rapidly. “Koo is built on a robust technology stack that uses Kotlin for our BE systems and our Android. For our iOS, we use Swift,” he said. The website is built on node.js while its systems ML use Python. Video engineering is essential for Koo. According to Gururaj, in a single day, more than 20TB of video is streamed from the app. “The CDN plays a very important role and we are leveraging the cloud here. We also use transcoders to convert videos to other formats.Our video engineering is above Elixir.

Gururaj says the cloud has helped the platform scale quickly in terms of user growth and functionality. “We are hosted on AWS and also use GCP for some use cases. Our infrastructure runs on Kubernetes and consists of Datastores – PostGreSQL, Open Search, AEROSPIKE, ArangoDB, GPU ML systems and data pipelines. Data Lake is the where we hydrate all our data points and signals and from there we extract intelligence and run other analytical pipelines.This is built on top of open source Apache frameworks like Kafka, Hudi, Spark, NiFi and Flink. use different types of databases ranging from OLTP, OLAP, Graph and NoSQL – Aerospike, Arango, Elastic Search, PostgreSql which power our functionality,” he added.

Estimated monthly active users of Koo in February and April 2021, source: Statista

Localized Features and Benefits

Acknowledging the natural comparison with its rivals, Gururaj is confident that the platform’s localized features are unlike any other. Earlier this week, the app launched a new in-app feature called “Topics” available in 10 languages ​​including Hindi, Bengali, Marathi, Gujarati, Kannada, Tamil, Telugu, Assamese, Punjabi and English. The “Topics” feature will allow users to select their own areas of interest in the language they are most comfortable with. “As 90% of India speaks a native language, we have attracted creators from across ‘Bharat’, including millions of new users who were previously hesitant to be on English-speaking platforms. We believe Koo is well on its way to creating a digital experience that is a category unto itself,” he said.

For Gururaj, the fact that the app allows users to express themselves in their native language is its biggest bonus. “Multi-Lingual Kooing (MLK) is a pioneering feature, which enables real-time translation of a message into multiple languages, which improves reach and bridges the language gap between speakers of different languages,” he explains. he.

Koo has further empowered users by allowing them to self-verify their profiles and be recognized as real voices. “We are the first social media platform to introduce voluntary self-verification for all users, and we are also among the first to publish how our algorithms work – a move that reiterates our commitment to transparency in the platform and a user-centric approach.

Koo’s interface, Source: Bootcamp.uxdesign.cc

Moderation Policy

Asked about the contentious issue of moderation that has plagued social media platforms like Facebook and Twitter, Gururaj explains the need for a balance between regulation and freedom. “We are a social media intermediary in the truest sense of the word, and we do not editorialize, label or characterize content (unless required by law). Our nuanced understanding of native languages, adherence to the law of the land, and strict community guidelines form the foundation of our content moderation practice. We are currently building dictionaries of words and phrases in all languages ​​in partnership with the Central Institute of Indian Languages ​​(CIIL) to promote fair use of language and responsible behavior online,” Gururaj said.

Hiring

Gururaj’s experience building Koo’s AI/ML team from scratch involved dealing with the challenges of hiring in the Indian subcontinent. “Complex projects, like those executed at Koo, require a mix of professionals from different specializations. Even with the density of talent in India, we have not yet been able to meet the demand for specialized data skills,” he said. Koo has still managed to overcome the steep learning curves by hiring early career professionals and providing them with strong mentorship support from senior technical managers.

Future vision for the AI/ML team

Gururaj notes that while the team is currently working on its recommendation engines, content classifiers and content moderation systems, Koo will continue to invest in R&D for its key areas such as native language technologies. “We want to develop large language models and train personalized neural networks for native language comprehension and translations. These investments are necessary to ensure that we respond to the growing complexity of content created on our platform by millions of first-time creators using native languages,” said Gururaj.

Recommendation system

How does Koo’s recommendation system hold up against TikTok’s ultra-accurate recommendation engines? “For our recommendations, we rely on the Koo social graph. Graph theory states that adding a node to a graph increases the value of the entire graph, not linearly, but exponentially. This exponential increase in value requires effort to ensure that we continue to extract value for our users.

The AI/ML team has invested in the development of graph algorithms like large-scale graph embeddings, which is an area of ​​active research for the application. Core teams are also hard at work building our own graph technologies that will allow them to train state-of-the-art graph algorithms for link prediction and small community detection. These apps will directly translate to providing superior recommendations on Koo.

cyber security

Gururaj explains how constant improvements are made to improve Koo’s cybersecurity. “We have a bug bounty program and work with known ethical hackers around the world to constantly improve Koo’s security posture. We have also invested in and implemented several security tools to prevent DDOS and ransomware attacks.

Additionally, we have implemented an internal security team and built automation to continuously monitor anomalous traffic,” he said.

Gururaj’s firm belief in Koo’s mass appeal makes him a phenomenon unlike any of his counterparts. He sees Koo as a standout innovator in English-first social media landscape building technology that can be consumed by the world – technology that speaks a native language.

“We enable users to cross language barriers when creating or consuming content. We bring eminent personalities from all walks of life to Koo to speak and be heard. A large number of government and social organizations use Koo to stay connected and inform the masses about various initiatives. We are hyperlocal and global at the same time,” he signed.

Lance B. Holton