标签：#AI

If you have worked with APIs for a while, you have probably realized that designing an API is only...

My job-search tool had five functions whose only purpose was fixing JSON that Claude had just written. _clean_json_text , _fix_unescaped_newlines , _fix_single_quotes , _strip_markdown_wrapper , _extract_and_parse_json . There was also a sixth, _retry_json_fix , which took the broken JSON and sent it back to the model with a polite request to fix its own mess. I wrote every one of them, one bug at a time, over weeks. I was a little proud of them. That was the problem. How you end up with five parsers JobSearch is my personal tool, in production, single user: me. It ingests job offers from nine boards, and when I press "Analyze", Claude reads the offer against my CV and returns a structured verdict: score, recommendation, career track, the English level the ad actually requires. That verdict has to be JSON, because everything downstream is a database row, not prose. The first version did what every tutorial does. Ask the model for JSON in the prompt, take response.content[0].text , run json.loads on it. It worked in the demo and then production started teaching me things. The model wrapped the JSON in markdown fences, so I wrote a function to strip them. Sometimes it used single quotes, so I wrote a function to fix those. Then a description with a line break inside a string, so I wrote _fix_unescaped_newlines . Then a NaN where a number should be. Every fix was five lines, obviously correct, and came with its own tests. I still have the test names in the git history and they read like a confession: test_removes_trailing_commas , test_replaces_nan_with_null , test_replaces_infinity , test_unclosed_fence_still_strips_opening . By April the parsing layer was around 250 lines with seven strategies, chained, each catching what the previous one let through. The last resort was the AI self-repair call: if nothing parsed, send the broken output back and ask the model to repair it. A second API call, with real latency and real cost, to fix a formatting problem the first call

That 130-file PR that shipped KeyEcho 1.0 contained a decision I never wrote about: the desktop app moved from Tauri 1 + Vue to Tauri 2 + SolidJS. The same summer, upweb.dev moved from Nuxt 4 to SolidStart, and keyecho.app (SSR, five languages, Stripe checkout) was built on SolidStart from scratch. Three separate decisions, same answer every time. Scope first. This post is about defaults for my personal projects. Near the end I'll cover two places where I still use React, one of which runs very well. The criteria changed Most of the code in my repos is AI-generated now. My job has shifted from writing code to reviewing it. The old framework checklist put most of its weight on learning curve and developer experience. Both are worth zero now: the model writes five frameworks fluently and knows the rules of hooks better than most humans do. AI crushed the price of writing code. Two costs didn't move: the runtime bytes every user downloads, and the human hours spent reviewing. Pick your stack by the new prices. Runtime cost lives in the architecture. No prompt removes it, so it decides the framework. Review bandwidth is finite, and AI throughput will grind down any consistency that is maintained by verbal agreement, so it decides the styling layer. I have seen more than one React codebase a few years into its life: two animation libraries, two carousel components, three styling systems, each introduced for a perfectly good reason on some ticket, and the sum is a mess nobody can clean up. AI did not invent that drift. It made it an order of magnitude faster. React is no longer my default: size and performance Let me discard the weak arguments first. Dependency arrays, stale closures, re-render storms: I am not going to relitigate any of it. The model knows the rules of hooks cold, eslint-plugin-react-hooks catches most slips, and React Compiler now handles memoization. The ergonomics debate no longer produces a winner. What still produces a winner is the client. react-do

Building Production-Grade LLM Evaluation Pipelines: From Vibes to Metrics How we replaced "looks good to me" with automated evaluation catching 92% of hallucinations before deployment The Problem: Why "Vibe Checks" Fail in Production Three months ago, our team shipped a RAG-based customer support assistant. It worked great in testing — we'd ask it questions, read the answers, and say "yeah, that looks right." Then it hit production. A customer asked about their billing cycle. The assistant confidently cited a policy that didn't exist. Another asked about API rate limits and got numbers from a competitor's documentation. By the time we caught it, 500+ users had seen hallucinated responses. The post-mortem was brutal: we had zero automated evaluation . Our test process was literally "ask 5 questions, read answers, thumbs up." What Production Evaluation Actually Needs Academic benchmarks (MMLU, HellaSwag) don't tell you if your system works for your use case. Production evaluation needs: Domain-specific judges — Your criteria, not generic "helpfulness" Speed — Evaluation must run in CI/CD, not overnight Regression detection — Know immediately when a prompt change breaks things CI/CD integration — Block merges that degrade quality Golden dataset management — Versioned, stratified, growing test cases Architecture: The Evaluation Pipeline ┌─────────────┐ ┌──────────────┐ ┌────────────────────┐ ┌──────────────┐ │ Test Cases │────▶│ LLM Under │────▶│ Judge Ensemble │────▶│ Metrics & │ │ (Golden Set)│ │ Test │ │ - Faithfulness │ │ Regression │ └─────────────┘ └──────────────┘ │ - Instruction F. │ │ Detection │ │ - JSON Schema │ └──────┬───────┘ │ - Custom LLM │ ▼ └────────────────────┘ ┌──────────────┐ │ Dashboard/ │ │ PR Comments │ └──────────────┘ Core Abstractions # eval/base.py @dataclass ( frozen = True ) class TestCase : id : str input : dict [ str , Any ] expected : dict [ str , Any ] | None = None tags : list [ str ] = field ( default_factory = list ) # ["edge-case",

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

Building Production-Grade LLM Evaluation Pipelines: From Vibes to Metrics How we replaced "looks good to me" with automated evaluation catching 92% of hallucinations before deployment The Problem: Why "Vibe Checks" Fail in Production Three months ago, our team shipped a RAG-based customer support assistant. It worked great in testing — we'd ask it questions, read the answers, and say "yeah, that looks right." Then it hit production. A customer asked about their billing cycle. The assistant confidently cited a policy that didn't exist. Another asked about API rate limits and got numbers from a competitor's documentation. By the time we caught it, 500+ users had seen hallucinated responses. The post-mortem was brutal: we had zero automated evaluation . Our test process was literally "ask 5 questions, read answers, thumbs up." What Production Evaluation Actually Needs Academic benchmarks (MMLU, HellaSwag) don't tell you if your system works for your use case. Production evaluation needs: Domain-specific judges — Your criteria, not generic "helpfulness" Speed — Evaluation must run in CI/CD, not overnight Regression detection — Know immediately when a prompt change breaks things CI/CD integration — Block merges that degrade quality Golden dataset management — Versioned, stratified, growing test cases Architecture: The Evaluation Pipeline ┌─────────────┐ ┌──────────────┐ ┌────────────────────┐ ┌──────────────┐ │ Test Cases │────▶│ LLM Under │────▶│ Judge Ensemble │────▶│ Metrics & │ │ (Golden Set)│ │ Test │ │ - Faithfulness │ │ Regression │ └─────────────┘ └──────────────┘ │ - Instruction F. │ │ Detection │ │ - JSON Schema │ └──────┬───────┘ │ - Custom LLM │ ▼ └────────────────────┘ ┌──────────────┐ │ Dashboard/ │ │ PR Comments │ └──────────────┘ Core Abstractions # eval/base.py @dataclass ( frozen = True ) class TestCase : id : str input : dict [ str , Any ] expected : dict [ str , Any ] | None = None tags : list [ str ] = field ( default_factory = list ) # ["edge-case",

The final approval settles one case, but it doesn't resolve the broader issue of using copyrighted works to train AI models.

Amazon's Adaptive Display accessibility feature for Fire TVs is officially rolling out today

The director role for the Center for AI Standards and Innovation (CAISI) has become a revolving door since David Sacks left his position as czar.

Sony Music Entertainment has filed another lawsuit against Udio, accusing the AI music generator of infringing the copyright of more than 30,000 of its songs, ranging from Elvis Presley's Hound Dog to Beyoncé's Say My Name, and Harry Styles' As It Was. The lawsuit, filed in a New York court on Monday, claims that this […]

California and XPRIZE competition tests whether drones can stop wildfires early.

After seeing Christopher Nolan's The Odyssey for the first time early last week, I came away impressed, but somewhat conflicted about two of the film's more fantastical set pieces. While those scenes were beautifully crafted and tremendously acted, there was something a little bit off about the way they hit my eye. I spent days […]

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

Over the past few years, AI has fundamentally changed how software gets built. Teams can go from an idea to a working application in a fraction of the time it used to take, and founders can create products with resources that would have been unimaginable just a few years ago. That's an incredible shift, and I think it's one of the most exciting changes our industry has seen. What hasn't changed, though, is the question that comes after the application is built: Is it actually ready? Throughout my career, I've been involved in delivering enterprise software across many different industries and organizations. One thing I've learned is that there isn't a single definition of what makes an application "ready." If you ask six different stakeholders whether a system is ready, you'll probably get six different answers—and they're all likely to be valid. That's because each person is looking at the software through the lens of the outcome they're responsible for: A founder may be wondering whether the application can handle the growth they're hoping for over the next year. A CTO is often focused on where the biggest technical risks are and what should be improved first. An engineering leader is thinking about production readiness, security, reliability, and operational support. An agency inheriting a client application wants to understand what they're taking ownership of before making commitments. An acquirer is trying to estimate the cost of technical debt An Investor wants confidence that the technology is creating long-term value rather than future expense. Those perspectives are different because the questions they're trying to answer are different. The challenge is that we often evaluate all software the same way. Traditional assessments tend to focus on the health of the codebase. They look at architecture, security, maintainability, testing, complexity, and technical debt. Those are all important, and they should absolutely be part of any technical review. But they'r

Vi um post legal esses dias sobre rodar duas contas do Claude Code na mesma máquina compartilhando tudo entre elas: mesmas skills, mesmos servidores MCP, mesmos hooks. O artigo era: "Um cérebro, duas carteiras" Eu rodo, exatamente o oposto, e acho que pra muita gente o oposto é a escolha certa. Minhas duas contas não são uma pessoal e uma reserva pra quando os créditos acabam. Uma é pessoal, outra é de trabalho. A última coisa que eu quero é meus MCP servers de trabalho, meus hooks de trabalho e meu histórico de projeto de trabalho vazando pras sessões pessoais. Então em vez de ligar as duas, eu mantenho elas separadas de propósito . O setup inteiro O Claude Code guarda o estado num diretório de config ( ~/.claude por padrão) e lê a variável CLAUDE_CONFIG_DIR pra apontar pra outro lugar. Esse é o único mecanismo que você precisa. Sem shim, sem symlink, sem jq . bash # ~/.zshrc # Claude Code: contas isoladas (pessoal vs trabalho) # Cada uma usa um CLAUDE_CONFIG_DIR proprio -> credenciais/sessao separadas. # ~/.claude = pessoal (default) # ~/.claude-work = trabalho claude-work () { CLAUDE_CONFIG_DIR = " $HOME /.claude-work" claude " $@ " ; } claude-personal () { CLAUDE_CONFIG_DIR = " $HOME /.claude" claude " $@ " ; } # 'claude' sozinho continua sendo a conta pessoal. source ~/.zshrc claude-work # pede login OAuth da conta de trabalho, uma vez só Usei funções de shell em vez de alias por um motivo: a função repassa "$@" limpo, então claude-work --resume abc e claude-work chat funcionam sem a variável de ambiente vazar pra nada mais no shell. Um alias faria quase o mesmo aqui, mas a função deixa a passagem de argumentos explícita. É isso. claude puro é pessoal. claude-work é trabalho. Nada é compartilhado, e é aí que mora a graça. Por que eu não compartilho o cérebro A versão de cérebro compartilhado faz symlink de skills , plugins , settings.json e dá merge no bloco mcpServers entre as duas contas pra elas se comportarem igual. Se as suas duas contas são de fato a mesm

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

Building Production-Grade LLM Evaluation Pipelines: From Vibes to Metrics How we replaced "looks good to me" with automated evaluation catching 92% of hallucinations before deployment The Problem: Why "Vibe Checks" Fail in Production Three months ago, our team shipped a RAG-based customer support assistant. It worked great in testing — we'd ask it questions, read the answers, and say "yeah, that looks right." Then it hit production. A customer asked about their billing cycle. The assistant confidently cited a policy that didn't exist. Another asked about API rate limits and got numbers from a competitor's documentation. By the time we caught it, 500+ users had seen hallucinated responses. The post-mortem was brutal: we had zero automated evaluation . Our test process was literally "ask 5 questions, read answers, thumbs up." What Production Evaluation Actually Needs Academic benchmarks (MMLU, HellaSwag) don't tell you if your system works for your use case. Production evaluation needs: Domain-specific judges — Your criteria, not generic "helpfulness" Speed — Evaluation must run in CI/CD, not overnight Regression detection — Know immediately when a prompt change breaks things CI/CD integration — Block merges that degrade quality Golden dataset management — Versioned, stratified, growing test cases Architecture: The Evaluation Pipeline ┌─────────────┐ ┌──────────────┐ ┌────────────────────┐ ┌──────────────┐ │ Test Cases │────▶│ LLM Under │────▶│ Judge Ensemble │────▶│ Metrics & │ │ (Golden Set)│ │ Test │ │ - Faithfulness │ │ Regression │ └─────────────┘ └──────────────┘ │ - Instruction F. │ │ Detection │ │ - JSON Schema │ └──────┬───────┘ │ - Custom LLM │ ▼ └────────────────────┘ ┌──────────────┐ │ Dashboard/ │ │ PR Comments │ └──────────────┘ Core Abstractions # eval/base.py @dataclass ( frozen = True ) class TestCase : id : str input : dict [ str , Any ] expected : dict [ str , Any ] | None = None tags : list [ str ] = field ( default_factory = list ) # ["edge-case",

The front-running tax that bled crypto for a decade needs only observable intent and a party that controls order. Agent marketplaces are rebuilding both. In September 2020, a security researcher who goes by samczsun found about $12 million of someone else's cryptocurrency sitting in a vulnerable contract, exposed, and realized he had a few minutes to rescue it before someone less friendly noticed. He wrote the rescue transaction. Then he stopped, because he understood the problem with sending it. The moment his transaction hit Ethereum's public waiting area, the mempool, every bot watching that space would see a profitable move spelled out in plain code, copy it, pay a higher fee to jump ahead of him, and take the $12 million themselves. His rescue would become their heist, and he would have personally handed them the map. He wrote about this later in an essay called "Escaping the Dark Forest," borrowing a metaphor from Dan Robinson and Georgios Konstantopoulos at Paradigm, who had borrowed it from Liu Cixin's science fiction: an environment where any signal of your presence gets you killed, so the only survivors are the ones who stay silent and shoot first. The mempool is a dark forest. Broadcasting a valuable intention into it is detection, and detection is death. Samczsun survived only by refusing to play the open game. He submitted his rescue privately, straight to a miner, bypassing the public mempool entirely, so the predators never saw it coming. That story is usually told as a piece of crypto lore. I want to tell it as something else, because the thing that killed transactions in the dark forest was never really about blockchains. It was about a shape, and that shape is quietly being rebuilt inside the AI agent marketplaces that a lot of people are racing to launch right now. When it finishes, the same predators will be back, and this time the prey will be your agents. The three conditions, and why blockchain was just the extreme case The phenomenon samczsun

Alphabet, Google's parent company, is reportedly working on a new chip designed to make its Gemini models run much more efficiently.

Introducing vernLLM: A Resilience Layer for LLM Applications Building production-ready LLM applications is not just about sending prompts and receiving responses. Real-world AI systems need to handle timeouts, provider failures, rate limits, inconsistent outputs, and reliability issues. That is where vernLLM comes in. vernLLM is a lightweight resilience layer for OpenAI-compatible chat completion APIs , providing a single interface with built-in retries, timeouts, circuit breaking, caching, structured output, and usage tracking. Instead of rebuilding the same reliability features for every LLM project, vernLLM gives you the tools needed to make your AI integrations more robust from the start. Features Automatic retries with backoff Transient failures happen. vernLLM automatically retries recoverable errors while failing fast on validation errors and non-retryable responses. Timeouts & cancellation Prevent hanging requests with configurable timeouts and cancellation support. Circuit breaker protection Automatically stop sending requests to failing providers and recover when the service becomes healthy again. Structured output with type safety Pass a Zod schema and receive validated, typed results back. const result = await llm . call ({ systemPrompt : ' Return JSON: { "skills": string[] } ' , userContent : ' Extract skills from: ... ' , schema : SkillsSchema // zod schema }); Provider-native JSON Schema support Constrain model generation itself instead of only validating responses afterward. Built-in caching support Cache LLM responses using your own cache adapter with cachedCall and cachedLLMCall . One interface across providers Use the same API across multiple providers: OpenAI Groq Mistral DeepSeek Cerebras Together AI Fireworks AI Ollama Anthropic Gemini AWS Bedrock Any HTTP-compatible provider through fromFetch Why vernLLM? Many LLM applications end up creating their own wrappers around provider SDKs to handle: retry logic API failures provider switching respons