Our point of view on shipping AI.
Practical writing on enterprise AI — strategy, evaluation, and the last mile into production. Plus field notes on security, commerce, and building teams.
Why Enterprise AI Doesn’t Ship
By Petru Popa ·Most enterprise AI dies between the demo and production. The blocker is rarely the model — it’s integration, data, and evaluation. Here’s where it breaks and how to close the gap.
- Dev Tools6 min
LiteLLM Supply Chain Attack: The Hardened Pipeline Published the Malware
By Petru Popa ·litellm 1.107.0 shipped a credential stealer through LiteLLM's own release workflow, pointed at a lookalike repository. PyPI's integrity API shows the project stopped publishing attestations in April, and even an attestation would have named the same workflow the attacker ran. The one record that told the truth was the Docker build provenance, which nobody checks.
Read - Partner Spotlight7 min
Stereo 70 Exports From CadSense: Give the DXF EPSG:3844, Not the Code Retired in 2008
By Petru Popa ·CadSense hands a surveyor a Romanian property's geometry as a Stereo 70 DXF. What the drawing does not carry is its coordinate system code, and the one Terra Apps' own product summary prints, EPSG:31700, was deprecated by the EPSG registry in 2008.
Read - Dev Tools6 min
Codex CLI Starts a Self-Updating Daemon by Default. The config.toml Switch Is Not in the Docs.
By Petru Popa ·Codex CLI 0.157.0 turned on automatic background-server startup, and the source at 0.162.1 still marks daemon_auto_start as stable and enabled by default. The daemon outlives the terminal, keeps the environment of the session that started it, and runs an updater that checks hourly and stops only when you tell it the right way. OpenAI's config reference does not list the key. Here is what the source says, and the two files that turn it off.
Read - AI Engineering6 min
Decision Model d1-3B Leads Its Size Class on Public Tests. On the Held-Out Split It Ranks 53rd.
By Petru Popa ·Liquid AI released d1-3B on 7 October as the best decision model under 10B parameters, citing a self-scored 48.57 on Decision Index 0.2.1. The public leaderboard had moved to edition 0.3 the day before, and that edition scores mostly on private tests. On it, d1-3B places 53rd of 117, inside the tie band with Decider 4B, and 76th on tasks from new domains. The public lead is real. It does not travel.
Read - Dev Tools6 min
CodeQL Swift Scanning Has No Supported Setup on Xcode 27. Some Runs Still Pass.
By Petru Popa ·GitHub's CodeQL 2.27.2 notes say traced builds stop working on macOS 27 and with Xcode 27, because Apple no longer ships the Intel slices the tracer needs. The workaround offered is build mode none, which Swift does not have. On the xcode-27 runner, one public Swift scan dies on a Bad CPU type error and another finishes green with no warning at all. The guard that would flag it is still an open pull request.
Read - Dev Tools6 min
GitHub Secret Scanning Got a New AI Model. Its Documented Blind Spots Predate It.
By Petru Popa ·On October 7 GitHub moved every repository with AI-detected password alerts onto a purpose-built secret detection model, and priced push-time checks in AI Credits that are spent whether or not a push is blocked. The only published account of what the detector skips was last edited two weeks before the swap, and a June docs migration had already cut it from five exclusions to one, losing the word that made that one readable.
Read - Dev Tools6 min
GitHub Copilot With Ollama: Going Offline Switches Off Content Exclusion
By Petru Popa ·Copilot CLI now discovers models from a running Ollama instance. GitHub's own docs show that the setting which keeps prompts off GitHub's servers also cuts the CLI off from the content exclusion rules it fetches there, and the CLI changelog says unreachable rules mean files stay readable.
Read - Dev Tools6 min
AI Agent Security After Copilot's Sandbox Went GA: The Default Policy Still Lets Agent Commands Read Your Registry Tokens
By Petru Popa ·GitHub made local sandboxing for Copilot generally available on 7 October. The docs behind the announcement show it ships off, that its default policy grants read access to package-registry tokens while leaving outbound traffic open, and that the CLI's own file tools are checked in software rather than by the operating system.
Read - AI Engineering6 min
Multimodal RAG on EmbeddingGemma 2: The 6x Storage Cut Is a Text-Only Number
By Petru Popa ·EmbeddingGemma 2 shipped on 6 October under Apache 2.0, with Matryoshka truncation pitched as up to a 6x storage saving. The model card's own truncation table shows the 128-dimension setting costs 13.4 points on the multimodal benchmark and gives back most of the code-retrieval gain. For a multimodal index the safe floor is 256 dimensions, and the model only loads on a transformers release that came out the same afternoon.
Read - AI Engineering6 min
AI Agent Evaluation by Repetition: ThinkingBox's 20-of-20 Column Is Not What Its Own Tool Prints
By Petru Popa ·Microsoft's ThinkingBox ranks agents by how many of 507 business workflows they pass on all 20 attempts, and counts that literally. The aggregator it ships reports a different number that credits near misses, a pass usually means ten ticket fields matched exactly, and an open issue leaves custom agents able to reach the golden state.
Read - AI Engineering6 min
A Local LLM for Coding on a 12 GB Card: Strata Is MIT, the Weights Are Not, and It Serves One Request at a Time
By Petru Popa ·Strata runs Qwen3.8-Flash-Next, a 125B mixture-of-experts model, on a gaming PC with a 12 GB card, and serves it as an OpenAI- and Anthropic-compatible endpoint for coding agents. The repository is eleven days old, has shipped 40 releases, and supports only the latest. Its issue tracker, its batching docs and the model's license say who should run it and who should not.
Read - AI News5 min
To Run Kolibri Locally, the License Is Apache 2.0 and the Runtime Is a Plugin Pinned Below vLLM 0.30
By Petru Popa ·Aleph Alpha released Kolibri, a 78B-total, 3.46B-active English-German MoE, under Apache 2.0 on 3 October. The model card answers what the announcement left open: the only supported runtime is Aleph Alpha's own vLLM plugin, which refuses vLLM 0.30; the 1M context is extrapolated past a 262,144-token config; and the chat template reasons at high effort unless told otherwise.
Read