Zartis ads

Guiding enterprise & mid-market firms from AI ambition to production

📉 Reducing LLM API Costs by 30% LLM costs are climbing rapidly — particularly in RAG systems where input tokens dominate overall spend. In our latest technical analysis, we unpack how a straightforward preprocessing step can reduce API costs by around 30% while maintaining high model quality across real production scenarios. This research, led by our own Adrián Sánchez de la Sierra, draws on 280 controlled experiments comparing lemmatisation, stemming, and stopword removal. The findings are data-driven, counterintuitive, and highly relevant for any team scaling AI workloads. Key insights include: 🔹 Lemmatisation and stemming increase token counts due to BPE fragmentation 🔹 Stopword removal consistently achieves ~30% token reduction 🔹 RAG quality retention at 95.9% 🔹 Summarisation quality at 93.9% 🔹 Retrieval degradation of <1% — a result that challenges common industry assumptions 🔹 For teams processing 10M tokens per day, annual savings can reach $13.5k–$135k For organisations building RAG systems or operating at scale, this is one of those rare optimisations that is low-risk, quick to implement, and delivers immediate ROI. 📘 Read the full analysis: https://lnkd.in/dXDW_tvC If you’re experimenting with preprocessing in your own pipelines, we’d be keen to hear your experience.

💡 Is your organisation truly ready for AI? Our State of AI Adoption 2025 Survey reveals a challenging reality: whilst AI ambition is universal, data readiness is not. The breakdown: 🔷 29% have clean, AI-ready data 🔷 55% need "some preprocessing" (often hiding significant data debt) 🔷 14% face fundamental data quality challenges The gap between AI ambition and execution comes down to one thing: data quality. This article breaks down what each level of readiness really means—and the strategic steps needed to move forward. Read the full analysis: https://lnkd.in/d4ba_rtu

Tech debt isn't just an engineering problem anymore, it's a board-level risk. In 2025, legacy systems are no longer just "annoying" - they're actively blocking innovation, driving up cloud costs, creating security vulnerabilities, and reducing your company's valuation. The numbers tell the story: → 23–42% of dev time consumed by technical debt → 62% of developers cite it as their top frustration → Companies that modernise see up to 40% faster feature delivery But here's the shift: Leading organisations aren't treating modernisation as a "clean-up project." They're reframing it as a strategic growth enabler. The question isn't "How much tech debt do we have?" It's "How fast can we evolve?" Read the full article: https://lnkd.in/dZrXwVnV #TechnicalDebt #SoftwareModernisation #DevOps #TechStrategy

💡 What if we could measure #how AI thinks, not just whether it gets the right answer? Adrián Sánchez de la Sierra, Head of AI Consulting at Zartis, explores a fascinating discovery: when LLMs reason, they create geometric paths through high-dimensional concept space - and the shape of those paths reveals reasoning quality. The breakthrough? Curvature matters more than correctness. Sharp turns in reasoning trajectories signal logical inconsistency, whilst smooth paths indicate coherent thinking. This works across any LLM, measuring through text output alone. The implications are profound: → Compare models on reasoning quality, not just accuracy → Monitor AI systems for degradation before failures occur → Debug where logic breaks, not just where answers are wrong This isn't just theoretical elegance, it's a practical tool for building more reliable AI systems. Open source, universal, and grounded in empirical validation. For engineers and leaders navigating AI deployment, this offers a new lens: geometry as a window into machine reasoning. Read the full technical essay: https://lnkd.in/dNVSamwj #AI #MachineLearning #LLMs #ReasoningAI #TechLeadership

48% of companies we surveyed are experimenting with AI without a dedicated budget. That's not innovation. That's hope. Our latest research reveals a striking gap: nearly half of organisations are running AI pilots with no formal plan, no clear KPIs, and no leadership alignment. The result? Expensive experiments that quietly fade away. Here's the truth: AI only delivers when it's married to your business strategy, not treated as a standalone tech experiment. In this article, we explore: ✓ Why AI ambition doesn't equal AI results ✓ The Growth vs. Efficiency framework for prioritisation ✓ How to translate AI language into business outcomes executives actually care about Because in 2025, hope is not a strategy. Alignment is. Read more: https://lnkd.in/dwPBYaSA