AI/ML Platform team at Roblox | Building SOTA LLM Safety Guardrails (RobloxGuard) | Reinforcement Learning (FAI-RL)

1. Developed RobloxGuard 1.0 Model — Roblox’s Open-Source, State-of-the-Art LLM Safety Guardrails

• Achieved SOTA performance, outperforming leading models such as Llama Guard, ShieldGemma, NVIDIA NeMo Guardrails, and even GPT-4o on key benchmarks

• Launched a Text Generation API powered by RobloxGuard 1.0

🔗 GitHub: https://github.com/Roblox/RobloxGuard-1.0

🔗 Paper: https://www.arxiv.org/abs/2512.05339

🔗 Text Generation API Announcement: https://devforum.roblox.com/t/beta-introducing-text-generation-api/3556520

2. Implemented & Open-Sourced FAI-RL — A Production-Ready Reinforcement Learning Framework for LLM Fine-Tuning

• Engineered a unified framework to support RL algorithms including DPO, PPO, GRPO, GSPO, and Supervised Fine-Tuning (SFT).

• Designed a highly extensible system with YAML-based configs and support for custom reward functions and dataset templates.

🔗 GitHub: https://github.com/Roblox/FAI-RL

3. Implemented a LLM Labeling Platform (Built from Scratch, Integrated with Label Studio)

• An automated data pipeline for LLM labeling tasks.

• Built a test case framework to streamline prompt engineering.

• Implemented a daily search quality evaluation.

• Functionality to sample and evaluate the quality of labeled data through human evaluators and LLMs.

• A fine-tuned CLIP model for image labeling, effectively addressing a cold start issue.

• Ability to retrieve internal datasets for use in a RAG system.

Impact:

• Created over 30 datasets with the help of 300 human evaluators.

• Trained more than 15 models across 5 different teams.

• Enhanced search quality by 2.2% through a fine-tuned CLIP model.

Built and scaled a real-time, multi-country A/B testing platform handling 200K QPS, implementing monitoring, circuit breakers

1. Implemented an in-house A/B test platform based on Java/Scala/Spring/Spark and Kafka/Hive/AWS for more than 3 years.

• Implemented Circuit Breaker to detect problematic A/B tests within a few minutes from scratch based on Scala/Spark Streaming/Kafka/S3/Hive/Oozie/Yarn/Sqoop.

• Implemented Exploration mode to reduce result update wait times from up to 24 hours to around 10 minutes for a full result set update from scratch based on Clickhouse/ZooKeeper.

• Implemented query-based monitoring system and rule-based message generator from scratch based on Prometheus/Kotlin/MySQL.

• Created a new data pipeline for exposure details widget to detect major exposure logging issue as early as possible from scratch.

• Set up infra/batch/deployment/monitoring to expand the A/B test platform from one country to multiple countries.

Carousell is a simple and easy way to buy and sell with anyone.

1. Implemented Search Service from monolithic towards microservices based on Go/Grpc/Protobuf/Envoy and Docker/Kubernetes/Elasticsearch.

• Implemented real-time item quality score, seller score for search ranking from scratch using Apache Beam/Redis/Kafka/Elasticsearch.

• Ingested user impressions into Elasticsearch and implemented simple random buckets to make A/B testing.

• Implemented boosting new seller’s items on the home page.

• Implemented a cache layer to reduce the number of Elasticsearch access for low latency and cost reduction.

Built high-performance messaging and data streaming platforms, managing massive user messages and company-wide Kafka/RabbitMQ clusters.

1. Implemented Messaging Platform to handle massive messages for users based on Java/Vertx/Spring and Redis/Cassandra.

• Implemented user service to collect user activity logs from connected sessions using Vertx/WebSocket/Redis.

• Implemented inbox service to send and receive massive messages using Spring/Redis/Cassandra.

2. Provided company-wide Kafka cluster to aggregate data from different teams based on Kafka/RabbitMQ/Spark and Mesos/Marathon from scratch.

• Set up and maintained Kafka, RabbitMQ cluster for the entire company on AWS.

• Implemented cluster migration/data transmission between topics based on Spark Streaming/Mesos/Marathon.

Built LINE Push Platform handling 100M+ concurrent users and 5B daily messages, designing high-throughput distributed messaging and Redis-based services.

1. Implemented LINE Push Platform for over 100M concurrent users and 5B messages per day from scratch based on Java/Netty/Spring and Redis/ZooKeeper/MySQL for 4 years.

• Implemented Session-Service to handle over 1M concurrent users per server using Netty.

• Implemented Message-Service to distribute messages for 5B messages per day using Spring/Redis.

• Implemented a high throughput, distributed Message Queue platform for asynchronous message -processing to support over 1M messages per second at peak time in front of API Gateway using Luxun.

• Implemented Service discovery to lookup Servers, support multiple regions, and failover for replacing load balancers.

• Implemented Redis cluster manager for high scalability and availability using Redis/ZooKeeper.