engblogs

A concise, unit-by-unit guide to 8th grade math topics—lines and slopes, functions, radicals and scientific notation, exponent rules, angles and triangles, transformations, and real-world data—with a focus on practical Duolingo Math practice.

A four-layer production stack that makes LLM evaluation fast and deterministic, enabling same-day iteration with per-sample caching, micro-LoRA patches, and end-to-end validation to surface and fix seam-level issues.

Four-layer, end-to-end LLM evaluation pipeline that enables same-day iteration through deterministic benchmarking, rapid micro-adapters, and seam-aware validation.

Blueberry is a Slack-native, durable reasoning harness that acts as a force multiplier for on-call engineers by auto-triaging alerts, surfacing grounded explanations, and coordinating parallel investigations to reach first insight and tested theories quickly.

A deep-dive into Netflix's real-time service topology system that uses streaming-first pipelines, backpressure, and a three-stage architecture to resolve network intermediaries, enrich graph data, and enable time-travel topology queries at scale.

A technical walkthrough of Netflix's real-time service topology at scale, detailing a streaming-first, three-stage pipeline that resolves intermediaries, uses backpressure, and supports time-travel topology queries with sub-second latency.

Netflix-scale, streaming-first architecture delivers real-time service topology at scale via a three-stage aggregation pipeline with backpressure, multi-layer graph storage, and time-travel topology reconstruction.

A scalable auditing method uses Gaussian probing of LoRA adaptors to detect whether AI models have been specialized to generate illegal content like CSAM, enabling platforms to flag or remove unsafe open-source models without prompting the models or generating outputs.

SceneSmith leverages a trio of vision-language-model agents (designer, critic, orchestrator) powered by GPT-5.2 to generate diverse, physics-aware indoor environments for robot training, enabling scalable simulation-first data collection and policy evaluation.

MIT's Cybersecurity Clinic mobilizes students from multiple disciplines to defend critical urban infrastructure against ransomware and other cyberattacks through hands-on, pro-bono risk assessments and the concept of defensive social engineering.

A scalable auditing approach uses Gaussian probing of LoRA adaptors to detect whether an open-source AI model has been specialized to generate CSAM, enabling platforms to flag or remove unsafe variants without prompting outputs.

SceneSmith uses a tri-agent vision-language model system to generate diverse, physics-aware indoor 3D scenes that accelerate robot training data generation and policy evaluation in simulation.