<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0">
  <channel>
    <title>Pintoed AI — Field Notes</title>
    <link>https://pintoedai.com</link>
    <description>Working AI systems. Honest reviews. New posts monthly.</description>
    <language>en-us</language>
    <item>
      <title><![CDATA[The minimum viable eval: how we ship AI features without a regression suite from hell]]></title>
      <link>https://pintoedai.com/blog/minimum-viable-eval</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/minimum-viable-eval</guid>
      <pubDate>Mon, 27 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[You don't need LangSmith, you don't need a 400-prompt golden set, you don't need a quarterly eval summit. Twelve prompts, twelve good answers, and a threshold. The exact recipe we install on every engagement.]]></description>
    </item>
    <item>
      <title><![CDATA[What our clients actually pay for Claude per month: 10 anonymized stacks]]></title>
      <link>https://pintoedai.com/blog/real-claude-bills-10-stacks</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/real-claude-bills-10-stacks</guid>
      <pubDate>Sun, 26 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Ten real client Claude bills, anonymized but unredacted. Three profile shapes — Opus-heavy, Sonnet-default, Haiku-batch — what each costs at typical workload size, and the five leaks we keep finding when we audit.]]></description>
    </item>
    <item>
      <title><![CDATA[Sora vs Kling vs Runway: 50 client clips, one verdict]]></title>
      <link>https://pintoedai.com/blog/sora-kling-runway-50-clips</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/sora-kling-runway-50-clips</guid>
      <pubDate>Sat, 25 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[We shipped roughly 50 finished clips for paying clients across all three. Each won a different category. Here's the decision tree we now use, with the failure modes the marketing reels don't show.]]></description>
    </item>
    <item>
      <title><![CDATA[Cursor vs Claude Code: how we actually split work between them in 2026]]></title>
      <link>https://pintoedai.com/blog/cursor-vs-claude-code-2026</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/cursor-vs-claude-code-2026</guid>
      <pubDate>Fri, 24 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Two tools, both at the top of their class. Not competitors — a division of labour. Here's the line we draw between them, what went wrong every time we tried to use one for both jobs, and the hybrid workflow we settled on.]]></description>
    </item>
    <item>
      <title><![CDATA[The $0.20/hr GPU is lying to you: Vast vs Modal vs RunPod on a real fine-tune]]></title>
      <link>https://pintoedai.com/blog/gpu-vast-modal-runpod-fine-tune</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/gpu-vast-modal-runpod-fine-tune</guid>
      <pubDate>Thu, 23 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Same Llama-3.1-8B LoRA fine-tune, three serverless GPU providers. The cheapest hourly rate didn't win the job. Wall-clock numbers, the gotchas, and the one we keep using for production.]]></description>
    </item>
    <item>
      <title><![CDATA[Tool-use vs structured output: when each one is the right tool]]></title>
      <link>https://pintoedai.com/blog/tool-use-vs-structured-output</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/tool-use-vs-structured-output</guid>
      <pubDate>Tue, 21 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Two ways to make a model produce machine-readable output. They look like the same feature with different syntax — they're not. The decision matrix we use, with the failure modes for each.]]></description>
    </item>
    <item>
      <title><![CDATA[AI customer support: the 80/20 we keep recommending]]></title>
      <link>https://pintoedai.com/blog/ai-customer-support-80-20</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/ai-customer-support-80-20</guid>
      <pubDate>Mon, 20 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Triage classifier, drafting assistant, auto-deflection — in that order, and never the brief the buyer pitched. The architecture that flat-lines CSAT and cuts cost-per-ticket 35-60%.]]></description>
    </item>
    <item>
      <title><![CDATA[From POC to production: 7 things that always break the week after demo day]]></title>
      <link>https://pintoedai.com/blog/poc-to-production-7-things-break</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/poc-to-production-7-things-break</guid>
      <pubDate>Sun, 19 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Rate limits, retries, eval drift, cost spikes, prompt rot, tool-call timeouts, observability gaps. Each one with a fix, plus the pre-launch checklist we run on every system.]]></description>
    </item>
    <item>
      <title><![CDATA[Long-context patterns that replaced our vector DB (the actual diffs)]]></title>
      <link>https://pintoedai.com/blog/long-context-replaced-our-vector-db</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/long-context-replaced-our-vector-db</guid>
      <pubDate>Sat, 18 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Companion to RAG-is-mostly-dead. Four concrete patterns, what each one looks like as code, the decision tree we run, and the small vector store we still keep around for the cases where it earns its keep.]]></description>
    </item>
    <item>
      <title><![CDATA[AI SDR economics: when it pays back in 3 weeks vs when you should kill it]]></title>
      <link>https://pintoedai.com/blog/ai-sdr-economics-payback</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/ai-sdr-economics-payback</guid>
      <pubDate>Fri, 17 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Three real engagements, three different payback curves. One paid back in 3 weeks. One in 9 months. One we killed at week 6. The four variables that move the math, plus the kill criteria we now set on day one.]]></description>
    </item>
    <item>
      <title><![CDATA[Stop building chatbots. Build outcomes.]]></title>
      <link>https://pintoedai.com/blog/stop-building-chatbots</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/stop-building-chatbots</guid>
      <pubDate>Tue, 14 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[The chat box is a UI choice, not a product. Most of the time it is the wrong one. Build the outcome the user wanted; skip the conversational middleman.]]></description>
    </item>
    <item>
      <title><![CDATA[The buzzword graveyard: terms we won't use in client meetings in 2026]]></title>
      <link>https://pintoedai.com/blog/buzzword-graveyard-2026</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/buzzword-graveyard-2026</guid>
      <pubDate>Mon, 13 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Twelve words and phrases we've officially retired from client conversations — AI-first, agentic, copilot, hallucinate, knowledge base, hyper-personalised, and the rest. Why each one died, and what we say instead.]]></description>
    </item>
    <item>
      <title><![CDATA[What "AI strategy" means when you're under $10M ARR]]></title>
      <link>https://pintoedai.com/blog/ai-strategy-under-10m-arr</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/ai-strategy-under-10m-arr</guid>
      <pubDate>Sun, 12 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Most AI strategy advice is written for Fortune 500. At your scale you don't need a council, a transformation, or an offsite. Three concrete questions, no deck, and the things you explicitly don't need.]]></description>
    </item>
    <item>
      <title><![CDATA[The five AI demos that always look great and ship terribly]]></title>
      <link>https://pintoedai.com/blog/five-demos-ship-terribly</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/five-demos-ship-terribly</guid>
      <pubDate>Sat, 11 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Chat-with-your-PDF, AI lead scoring, AI dashboard insights, AI voice on cold leads, AI SEO content. Five categories that wow on stage and crater in production. Why each one fails and what to build instead.]]></description>
    </item>
    <item>
      <title><![CDATA[Opus 4.7 1M context: what it actually unlocks vs the marketing]]></title>
      <link>https://pintoedai.com/blog/opus-4-7-1m-context-real</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/opus-4-7-1m-context-real</guid>
      <pubDate>Thu, 16 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[A million tokens of context is the headline. The reality is more interesting and more bounded. Four workloads where the bigger window changed our architecture, three where it didn't, and the cost discipline that keeps the bill from running away.]]></description>
    </item>
    <item>
      <title><![CDATA[Sonnet 4.6 vs Haiku 4.5: the routing decisions we make every week]]></title>
      <link>https://pintoedai.com/blog/sonnet-haiku-routing-weekly</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/sonnet-haiku-routing-weekly</guid>
      <pubDate>Fri, 10 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Two models, sat side by side, are the workhorses of our 2026 stack. Here's the honest cheat sheet for which workload pins to which, the breakeven points, and the Haiku-first / Sonnet-on-fail pattern we route most traffic through.]]></description>
    </item>
    <item>
      <title><![CDATA[We killed our newsletter SaaS. Here's why.]]></title>
      <link>https://pintoedai.com/blog/killed-our-newsletter-saas</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/killed-our-newsletter-saas</guid>
      <pubDate>Tue, 28 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[No Buttondown. No Substack. No Mailchimp. The subscribe buttons on this site go to RSS and to a mailto link. We ripped the SaaS out two weeks ago and we're not going back. The math, the brand-voice argument, and what we'd do if we did need a list.]]></description>
    </item>
    <item>
      <title><![CDATA[Prompt caching deep-dive: the 4 patterns that compound savings]]></title>
      <link>https://pintoedai.com/blog/prompt-caching-deep-dive</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/prompt-caching-deep-dive</guid>
      <pubDate>Thu, 09 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Companion to the 71% bill-cut case. What to cache, where to put breakpoints, the four prompt shapes that produce the largest sustained savings, and the two anti-patterns we keep finding in audits.]]></description>
    </item>
    <item>
      <title><![CDATA[ElevenLabs vs Suno: shipping audio for clients in 2026]]></title>
      <link>https://pintoedai.com/blog/elevenlabs-vs-suno-2026</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/elevenlabs-vs-suno-2026</guid>
      <pubDate>Tue, 07 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Same category in our reviews index, two completely different jobs. ElevenLabs is the voice studio. Suno is the music generator. Mostly we use both — for entirely different reasons. Decision tree, failure modes, real cost shape.]]></description>
    </item>
    <item>
      <title><![CDATA[Migrating from OpenAI to Claude: the 5 gotchas we hit every time]]></title>
      <link>https://pintoedai.com/blog/openai-to-claude-migration</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/openai-to-claude-migration</guid>
      <pubDate>Mon, 06 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Function calling vs tool-use, prompt caching as not-a-drop-in, output format guarantees, refusal behaviour, and the eval rebuild that catches the rest. The week-by-week sequence and what cost actually does after the swap.]]></description>
    </item>
    <item>
      <title><![CDATA[Pricing AI work: why we charge fixed-fee, not T&M]]></title>
      <link>https://pintoedai.com/blog/pricing-ai-work-fixed-fee</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/pricing-ai-work-fixed-fee</guid>
      <pubDate>Sun, 05 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[The first call with a new client almost always includes "what's this going to cost?" Why fixed-fee, what it costs us when we get it wrong, the engagement shapes we offer, and the two narrow cases where T&M is actually right.]]></description>
    </item>
    <item>
      <title><![CDATA[Apollo vs Lemlist vs Smartlead: which outbound stack we actually deploy]]></title>
      <link>https://pintoedai.com/blog/apollo-lemlist-smartlead-outbound-stack</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/apollo-lemlist-smartlead-outbound-stack</guid>
      <pubDate>Sat, 04 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Three platforms in our reviews index, three different jobs. We pick between them most weeks for client engagements. Here's how — with the deliverability gotcha that flips the answer faster than buyers expect.]]></description>
    </item>
    <item>
      <title><![CDATA[Writing tools the model actually uses]]></title>
      <link>https://pintoedai.com/blog/writing-tools-the-model-uses</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/writing-tools-the-model-uses</guid>
      <pubDate>Fri, 03 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Tool definitions are prompts. Most teams write them like API docs and the model uses them poorly. Six rules — name like a verb, describe the trigger, mark required params with formats, write errors as prompts, keep the surface small.]]></description>
    </item>
    <item>
      <title><![CDATA[Stop measuring deflection. Measure resolution.]]></title>
      <link>https://pintoedai.com/blog/stop-measuring-deflection</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/stop-measuring-deflection</guid>
      <pubDate>Thu, 02 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Deflection-rate is the metric the AI customer-support industry sells on. It also misleads more buyers than any other metric in board decks. What it actually counts, why it flatters bad systems, and what to track instead.]]></description>
    </item>
    <item>
      <title><![CDATA["AI agents" mostly means "scripts with model calls" and that's fine]]></title>
      <link>https://pintoedai.com/blog/ai-agents-are-scripts</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/ai-agents-are-scripts</guid>
      <pubDate>Tue, 31 Mar 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Most production AI agents we ship are 3-5 well-shaped function calls in a loop. The framing matters because it produces better systems than "autonomous agentic AI" does, even when the implementation is identical.]]></description>
    </item>
    <item>
      <title><![CDATA[The shape of a useful LLM logging pipeline]]></title>
      <link>https://pintoedai.com/blog/llm-logging-pipeline</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/llm-logging-pipeline</guid>
      <pubDate>Mon, 30 Mar 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Companion to the production-readiness checklist. The exact log schema, retention policy, sampling UI, and budget we install on every engagement — typically under $50/mo at small scale, scaling cleanly.]]></description>
    </item>
    <item>
      <title><![CDATA[Klaviyo vs Mailchimp vs Kit: the AI-era email decision]]></title>
      <link>https://pintoedai.com/blog/klaviyo-mailchimp-kit-ai-email</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/klaviyo-mailchimp-kit-ai-email</guid>
      <pubDate>Sun, 29 Mar 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Three platforms, three completely different audiences. Klaviyo for e-commerce, Kit for creators, Mailchimp for everybody-else SMB. Where each one earns its keep, the AI flows that pay back, and the migration paths.]]></description>
    </item>
    <item>
      <title><![CDATA[Synthesia vs HeyGen for L&D video at enterprise scale]]></title>
      <link>https://pintoedai.com/blog/synthesia-vs-heygen-l-and-d</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/synthesia-vs-heygen-l-and-d</guid>
      <pubDate>Sat, 28 Mar 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Two avatar-video platforms, similar feature surfaces, very different procurement profiles. Synthesia ships in regulated enterprise. HeyGen ships in agencies and creator-adjacent businesses. Decision tree, cost shape, what neither one is doing.]]></description>
    </item>
    <item>
      <title><![CDATA[Citations in production: the hard requirement most teams skip]]></title>
      <link>https://pintoedai.com/blog/citations-in-production</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/citations-in-production</guid>
      <pubDate>Fri, 27 Mar 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Citations are the difference between an AI feature you can put in front of a regulated buyer and one you can't. Four patterns that work — quote-then-attribute, anchored offsets, tool-call provenance, confidence-gated — with the eval bar for each.]]></description>
    </item>
    <item>
      <title><![CDATA[Fine-tuning vs prompting in 2026: when each one wins]]></title>
      <link>https://pintoedai.com/blog/fine-tuning-vs-prompting-2026</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/fine-tuning-vs-prompting-2026</guid>
      <pubDate>Thu, 26 Mar 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Fine-tuning is back, partially, mostly on smaller models. The four cases where it beats a well-prompted Haiku, the breakeven volume math, and the 70% of fine-tune projects we get asked to scope that should have been better prompts.]]></description>
    </item>
    <item>
      <title><![CDATA[Gemini vs Claude long-context: the parts the benchmarks miss]]></title>
      <link>https://pintoedai.com/blog/gemini-vs-claude-long-context</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/gemini-vs-claude-long-context</guid>
      <pubDate>Tue, 24 Mar 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Both providers ship 1M-token context. The benchmark numbers are close. The production behaviour is not. Where each one wins on real workloads, the routing decisions we make weekly, and the gotcha that flips the answer based on what's in your prompt.]]></description>
    </item>
    <item>
      <title><![CDATA[The AI build checklist: 23 questions before you write a line of code]]></title>
      <link>https://pintoedai.com/blog/ai-build-checklist</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/ai-build-checklist</guid>
      <pubDate>Wed, 22 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Most failed AI projects fail in scoping, not in shipping. Here are the 23 questions we answer before any keyboard touches a repo — covering data, model choice, eval, fallback, cost ceiling, and kill criteria.]]></description>
    </item>
    <item>
      <title><![CDATA[How we cut a client's LLM bill 71% with prompt caching (real numbers)]]></title>
      <link>https://pintoedai.com/blog/cut-llm-bill-71-percent</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/cut-llm-bill-71-percent</guid>
      <pubDate>Wed, 15 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[A real client engagement: $40K/month down to $11.5K/month, no quality drop. The receipts, the gotchas, and the three things we'd do differently.]]></description>
    </item>
    <item>
      <title><![CDATA[Claude vs ChatGPT for production agents in 2026 — what actually breaks]]></title>
      <link>https://pintoedai.com/blog/claude-vs-chatgpt-production-agents</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/claude-vs-chatgpt-production-agents</guid>
      <pubDate>Wed, 08 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[We've shipped agents on both for paying clients across the last 18 months. Here's where each one breaks first under real load — and why we keep coming back to Claude for tool-heavy workloads.]]></description>
    </item>
    <item>
      <title><![CDATA[When NOT to build with AI: three engagements we turned down]]></title>
      <link>https://pintoedai.com/blog/when-not-to-build-with-ai</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/when-not-to-build-with-ai</guid>
      <pubDate>Wed, 01 Apr 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Three real client conversations where we said "don't build this with AI" — and one where we said "build it without AI first." The pattern that connects them.]]></description>
    </item>
    <item>
      <title><![CDATA[RAG is mostly dead. Here's what we use instead in 2026.]]></title>
      <link>https://pintoedai.com/blog/rag-is-dead-long-context</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/rag-is-dead-long-context</guid>
      <pubDate>Wed, 25 Mar 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Two years ago every project started with a vector database. Today most of ours don't. The combination of long context, prompt caching, and tool-use has eaten 70% of what RAG was for.]]></description>
    </item>
    <item>
      <title><![CDATA[Anatomy of a working AI SDR (and why most fail)]]></title>
      <link>https://pintoedai.com/blog/ai-sdr-anatomy</link>
      <guid isPermaLink="true">https://pintoedai.com/blog/ai-sdr-anatomy</guid>
      <pubDate>Wed, 18 Mar 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Most AI SDRs we've seen in the wild fail in the same three ways. Here's the architecture we use that doesn't — and the budget you actually need to ship one.]]></description>
    </item>
  </channel>
</rss>
