Senkovi is Josh Angel's technical portfolio and research notes: AI models, agents, software and DevOps infrastructure, developed through experiments and lessons from running systems.https://senkovi.comenhttps://senkovi.com/ledgerhttps://senkovi.com/ledgerA developer-preview leadership workspace: profile asks, sourced event history, owner-approved goals and scheduled agent workflows, with macOS and web clients.Mon, 07 Sep 2026 20:21:55 GMThttps://senkovi.com/when-a-frontier-lab-ships-a-model-in-your-domain-benchmark-ithttps://senkovi.com/when-a-frontier-lab-ships-a-model-in-your-domain-benchmark-itNVIDIA shipped a strong 4B translation model. We benchmarked it against our customer-tuned models on three client domains. The result says more about metrics than about models.Tue, 01 Sep 2026 21:00:00 GMThttps://senkovi.com/a-dawn-of-new-metrics-to-define-model-successhttps://senkovi.com/a-dawn-of-new-metrics-to-define-model-successTraditional metrics are not keeping up with modern model progression. How do we tackle this, and how is arbitr pushing bringing generation evaluation into the future.Tue, 18 Aug 2026 21:00:00 GMThttps://senkovi.com/rivet-1b-training-a-small-model-from-scratchhttps://senkovi.com/rivet-1b-training-a-small-model-from-scratchA backdated note on rivet-1b, a 1.09B parameter model trained locally from tokenizer to post-training, with paper and weights published.Mon, 03 Aug 2026 21:00:00 GMThttps://senkovi.com/ibm-think-2026-bostonhttps://senkovi.com/ibm-think-2026-bostonThe conversation at Think, across sessions, hallway conversations, and the people building on the platform, reinforced something we've been believing for a while: the era of "one large model to do everything" is (or arguably, already has) giving way to something more nuanced.Tue, 23 Jun 2026 21:00:00 GMThttps://senkovi.com/the-foundation-layerhttps://senkovi.com/the-foundation-layerEarlier this year we stopped building one base model per language and trained a single one across ten. We call it tahi, "one" in Māori, because it's the first of the family and the thing every model after it stands on.Sun, 21 Jun 2026 21:00:00 GMThttps://senkovi.com/closing-the-style-gap-dpo-and-beyondhttps://senkovi.com/closing-the-style-gap-dpo-and-beyondIt isn't always about accuracy for a translation - it's about domain-specific "vibe."Mon, 18 May 2026 21:00:00 GMThttps://senkovi.com/why-we-re-building-a-fast-path-system-for-live-interpretationhttps://senkovi.com/why-we-re-building-a-fast-path-system-for-live-interpretationReal-time interpretation is one of those problems that sounds solved until you actually try to build it for production.Wed, 25 Feb 2026 20:00:00 GMThttps://senkovi.com/a-foundation-of-excellence-why-training-data-quality-determines-model-successhttps://senkovi.com/a-foundation-of-excellence-why-training-data-quality-determines-model-success"Nothing could fix the building's flawed foundation." – Enter ShikaiFri, 29 Aug 2025 21:00:00 GMThttps://senkovi.com/zero-strategies-for-scaled-model-training-a-quick-lookhttps://senkovi.com/zero-strategies-for-scaled-model-training-a-quick-lookA Technical Deep-Dive into DeepSpeed ZeRO Performance Characteristics for training AI SystemsMon, 25 Aug 2025 21:00:00 GMThttps://senkovi.com/ai-hardware-is-a-tale-of-two-toolshttps://senkovi.com/ai-hardware-is-a-tale-of-two-toolsApple or Linux for AI development? An age old question, or a simple answer?Mon, 21 Jul 2025 21:00:00 GMThttps://senkovi.com/when-your-ai-model-works-better-than-you-thinkhttps://senkovi.com/when-your-ai-model-works-better-than-you-thinkA Translation Model Discovery StoryWed, 16 Jul 2025 21:00:00 GMThttps://senkovi.com/beam-search-and-translationhttps://senkovi.com/beam-search-and-translationA Study of Quality and Resource Tradeoffs in Straker Tiri ModelsSun, 06 Jul 2025 21:00:00 GMT