Why functional test-passing is an incomplete metric for coding agents, and what the SWE-CC benchmark reveals about repository governance, workflow compliance, and real-world software engineering.
Why standard AI world models dream impossible physics, and how I built PI-WorldModel: an open-source latent architecture that obeys basic motion laws and runs on microcontrollers in 2.94 microseconds.
Why standard code LLMs struggle with program logic, and how I built JanusIR-500M-Research: a 545.7M parameter hybrid model combining causal language decoders with relational graph attention over LLVM IR.
A deep dive into Google's TimesFM-3 - exploring native multivariate forecasting, dynamic past-future covariates, dual-stage attention, and empirical validation on synthetic retail datasets.
What the August 17, 2026 outage reveals about our developer toolchain: why distributed Git did not stop teams from freezing, the gap between runtime and delivery availability, and practical fallbacks for both individual devs and production teams.
Why error messages deserve the same design rigor as APIs and UIs - a look at diagnostics as a first-class engineering discipline, from Rust to Clang to your own CLI tools.
A head-to-head research benchmark comparing Retrieval-Augmented Generation (RAG) and Knowledge-Augmented Generation (KAG) on multi-hop reasoning queries, with an LLM-as-Judge scoring system and an interactive web dashboard.
Symtrace replaces line-based git diff with deterministic, AST-driven structural analysis — detecting MOVE, RENAME, and MODIFY at the syntax node level using tree-sitter and BLAKE3 hashing.
A systems-level semantic search engine implemented entirely in C, demonstrating vector embeddings, cosine similarity, and offline information retrieval, compiled to WebAssembly for portable execution.
A conceptual exploration of backend engineering fundamentals - state, flow, failure, and scale - grounded in real-world analogies and modern system behavior.
A privacy-first mobile note-taking application implementing complete Retrieval-Augmented Generation pipeline locally on Android devices using React Native and ONNX Runtime.
Embedding Python-trained ML models directly into legacy Java applications with sub-millisecond inference latency and zero network overhead.
A systems-level exploration of long-term memory, contextual retrieval, and cognitive-inspired architectures for persistent AI behavior.