BreakpointSnap explains how CLAD combines Lens Studio with the SPECS Interaction Kit and SPECS UI Kit to accelerate spatial application development. The post details closed-loop scene construction, debugging, automated interaction testing with LEAF, and performance optimization for wearable experiences.
Snap engineering blog
Snap describes Casper as a fleet of autonomous AI agents that can take engineering tasks from Slack, the web portal, or scheduled triggers and turn them into validated, reviewable pull requests. The post explains how Casper fits into Snap’s broader AI-assisted development workflow, emphasizing safe sandboxed execution, identity-based access control, Code Search context, and automatic review via CodePal. It also outlines Casper Agent Templates and the MCP Gateway as the mechanisms that let the system scale beyond coding into broader operational and analytical workflows.
This post explains how Snap built a GNSS-based tracking system for Spectacles to support geo-referenced 6DoF augmented reality. It describes how GNSS, VIO, magnetometer, and IMU data are fused to provide stable world-anchored tracking for outdoor AR use cases such as navigation, points of interest, and location-based experiences.
The post explains how Snap built an agent-first code search platform to help AI agents and engineers navigate thousands of repositories quickly and accurately. It details the decision to favor exact search over RAG, the Zoekt-based sharded architecture, and how the system integrates with MCP and CodePal to support cross-repo review, access control, and low-latency indexing.
The post introduces Spatial Benchmark, an open-source evaluation suite for measuring how well language models reason about 3D space. It reports results across multiple frontier and open models, showing strong performance on topological and constraint-based tasks, but persistent weaknesses in orientation-sensitive vector math, depth placement, and large-scene hierarchy handling.
Snap describes how it built CodePal, an internal AI code review assistant designed to handle the growing bottleneck created by AI-written code. The post explains its architecture, including symbolic context retrieval, a multi-stage review loop, and a verifier, and shows how these techniques helped Snap scale AI review to over 90% of pull requests while improving recall and reducing false positives.
Snap Cloud introduces an integrated backend for Spectacles developers, powered by Supabase and built directly into Lens Studio. The post explains how the platform provides PostgreSQL, storage, Edge Functions, realtime sync, and seamless OIDC-based authentication while reducing setup complexity and preserving security boundaries. It also covers the Lens Studio plugin, developer dashboard, local CLI workflows, and the alpha program for early access.
This post explains how Snapchat treats performance as a core product feature, with special attention to protecting the p90 tail latency of critical mobile journeys like open-to-camera. It details a custom production tracing system, including bounded buffers, retroactive spans, and smart sampling, and shows how that system helps diagnose issues such as disk contention, priority inversion, and language interoperability overhead.
Snap describes how its workflow orchestration platform, Flowrida, evolved from a single Kubernetes cluster into a tiered multi-cluster architecture to improve resilience, scalability, and recovery. The post details the constraints of Apache Airflow at Snap’s scale, the design of Force Switch and Continuous Switch mechanisms, and supporting components such as metadata synchronization, a unified Cloud Console UI, and cross-cluster task dependency handling.
Snap details how it migrated key AB experiment-processing pipelines from CPU-based Spark to GPU-accelerated execution using NVIDIA RAPIDS on Google Cloud. The post explains the performance, infrastructure, and scheduling challenges involved, including Kubernetes migration, storage changes, and fallback logic, and shows how these changes delivered stable operation with major runtime and cost improvements.
This post introduces Snap’s Spectacles Interaction Kit (SIK) and Spectacles UI Kit (UIKit), two open, TypeScript-based frameworks designed to make spatial interaction and UI development for Specs more intuitive and efficient. It explains the core design principles behind the kits and dives into technical challenges such as near-field gesture disambiguation, far-field targeting, wearable performance optimization, raycast accuracy, and reusable UI component architecture.
Snap introduces Agent Format, a declarative, vendor-neutral standard for defining AI agents independent of the runtime that executes them. The post explains how the schema separates agent identity, interface, tools, execution policy, and constraints to improve portability, shared tooling, and governance across frameworks such as LangChain and Google ADK. It also describes how adapters, runtime enforcement, and human-in-the-loop approval gates enable scalable, auditable production agent systems.
EyeConnect is a shared-AR alignment system for Spectacles that lets two or more users join a common virtual coordinate frame simply by looking at each other, exchanging only 3D head poses and 2D keypoint detections (no images) and deleting data after alignment to protect privacy. Technically, it relies on an on-device quantized low-parameter CNN to detect five Spectacles keypoints (3 ms runtime on the DSP), an Egomotion Alignment solver formulated as a Quadratic Eigenvalue Problem with gravity-constrained simplification, RANSAC-based association, and refined clock synchronization; in tests it produced a median initial error of ~15 cm and high-quality accuracy of ~2.2 cm within 5 m with a 90% time-to-first-fix of 2.6 s. The authors note limitations in crowded environments and scenarios with little user motion, where association ambiguity and short stereo baselines make refinement harder.
The post describes Snapchat's Universal User Modeling (UUM), a foundation model that produces long-term, cross-surface user embeddings by aggregating year-long behavioral sequences across Content, Ads, Lens, and other domains. It covers the scalable data pipeline (Spark + Iceberg), transformer-based sequence encoders with information-bottleneck tokens, and a multi-task next-k event training objective, and explains how UUM embeddings enhance real-time ranking and personalization while complying with Snap's privacy policy.
Prism is a Spark-powered platform at Snap that simplifies ML data processing and improves consistency across teams. It is designed to accelerate ML iteration by providing a unified, scalable data processing foundation for feature and model pipelines.
Snap receives a large volume of new ads daily and needs scalable systems to review creatives for policy compliance, safety, and relevance. This article discusses approaches used to analyze ad content at scale and ensure high-quality, relevant ad experiences for Snapchatters.
Bento is Snap’s ML platform designed to help build and deploy machine learning at scale, powering personalized experiences across the product. The platform focuses on MLOps capabilities to streamline model development, deployment, and operationalization for teams across Snap.
On Spotlight, Snap personalizes short-video recommendations by using large-scale embedding-based retrieval to select candidate stories in real time. This post covers the challenges and techniques applied—such as two-tower models and large-scale embedding infrastructure—to deliver relevant content based on user interests and watch history.
Privacy is central to Snap’s product philosophy, and Differential Privacy (DP) is one of the tools used to protect user data. This article summarizes privacy concerns and explains how Snap applies DP to features like friend recommendations and Maps to better protect Snapchatters' data.
This post explains how supplemental feeds help advertisers bring AR experiences into Snap ads to reach Snapchatters at scale. It describes the systems and platform features that enable advertisers to integrate AR into their creative workflows and deliver immersive experiences.