engblogs

RAP (Random Access Parquet) introduces an external index for data lakes that maps keys directly to Parquet file locations and row numbers, enabling true online point queries with single or a few precise ranged reads and dramatically reduced latency for interactive AI agents and services.

Why durable, stateful workflows and boundary-driven retries are essential for production AI, demonstrating how persistent execution, layered orchestration, and meaningful units of work enable reliable results across long-running, multi-step AI pipelines.

RAP enables fast point-lookups on Parquet files in the data lake by using an external index that maps keys directly to file and row locations, collapsing multi-step I/O into a single or few range reads for interactive online queries and AI context retrieval.

Think beyond single-model prompts: build production AI with durable workflows that treat execution as persistent state, using bounded retries and Temporal-based orchestration to survive failures at scale.

Convergence Is Not Enough: a critical look at Automerge-based merging that shows convergence can violate program invariants when merging the heap, and proposes merge-aware datatypes that merge high-level intents rather than raw writes.

ABBEL introduces a belief-state framework that replaces long interaction histories with learnable, autoencoder-inspired belief summaries overseen by a belief-grading objective to enable memory-efficient, long-horizon LLM interaction.

ABBEL introduces belief-state memory to replace full interaction history with graded, autoencoded beliefs, enabling memory-efficient, long-horizon LLM interaction for complex tasks like collaborative coding.

Explores shifting nuclear plant operations toward autonomous, transparent supervisory control that blends human oversight with finite-state automation to enable safer, scalable, and cost-effective operation.

Explores autonomous supervisory control for nuclear plants, highlighting human-machine collaboration, finite state automata-based automation, and the path toward scalable microreactor deployment.

Reframing high-frequency trading data preprocessing as a Bayesian, representation-learning problem that adaptively captures market structure from historical limit order book data to preprocess across regimes and reduce compute.

A Bayesian, regime-aware representation-learning approach to adaptive preprocessing of limit order book data that learns market structure to cope with unseen conditions, reframing preprocessing as a learning problem rather than fixed normalization.

Fly.io pivots to Sprites—AI-powered, semi-disposable computers for agents that reimagine cloud infrastructure and redefine the company's future as a Platform as a Service provider.