标签：#ev

Hoi hoi! I'm @nyaomaru, a frontend engineer who recently watched a flock of geese march through...

Uploading a PDF and calling a translation model looks like a two-step feature. In production, it is a job pipeline with untrusted input, two different extraction paths, several expensive stages, and an output that can be fluent while still being wrong. That distinction matters for a small SaaS team. The translation request may come from support, sales, or an internal operations task. Nobody wants to operate a document platform, but the workflow still needs to answer basic questions: Was the upload actually a PDF? Does the file contain selectable text or scanned page images? Can a retry create a second charge or a conflicting result? What happens when page 37 fails after the first 36 pages succeed? How do we know the translated PDF is not blank or visually broken? When are the source and result deleted? The translation model is one component. Reliability comes from the system around it. Define the Job Contract First I would not let a file reach an extractor until the API has established a narrow contract. For example, a translation request might include: type TranslationStyle = " general " | " technical " | " academic " ; interface CreateTranslationJob { uploadId : string ; sourceLanguage : string | " auto " ; targetLanguage : string ; style : TranslationStyle ; idempotencyKey : string ; containsRestrictedData : boolean ; } The request should be rejected when the source and target languages are identical, the upload is missing, the target language is unsupported, or policy says the document cannot leave an approved environment. File validation should also be explicit. Do not trust the filename or browser-supplied MIME type. Check at least: the actual byte size; the file signature; whether the parser can open the document; whether the PDF is encrypted; the page count; whether the job fits the account or product limit. A 20 MB limit is simple to explain in a user interface, but size alone is not a good predictor of work. A compressed 200-page text PDF can be smaller th

Indoor maps are often treated as a rendering problem: take a floor plan, extrude a few walls, and display the result. That is useful for a viewer, but it breaks down when a team needs to edit a real space, place assets, or hand the result to another application. We are building KiMap around a different boundary: turn a floor plan into an editable indoor scene in the browser, then keep the resulting structure useful for an SDK consumer. Why a floor plan is not enough A production indoor workflow needs more than a textured image on a plane. At minimum, the editor has to preserve the relationships between walls, floors, rooms, openings, and the objects placed in the space. Those relationships determine whether the result can later support navigation, facility workflows, a digital twin, or a custom web experience. That is why the current KiMap workflow starts with structure. You can define the indoor geometry, inspect it in 2D and 3D, and keep editing instead of committing to a static export too early. The browser editor boundary The editor is built with React and Three.js. The goal is not to replace every DCC tool. It is to make the early spatial workflow accessible to teams that need to test an indoor experience before investing in a full custom pipeline. The parts we are concentrating on are: editable floor-plan structure and bounded spaces 2D and 3D scene inspection in the same workflow reusable 3D furniture and local asset handling saving an indoor project without dropping the referenced model data a path toward SDK-oriented rendering and integration The last point matters. A scene that looks correct in an editor is not automatically useful to an application. We want the data boundary to be explicit enough that an SDK consumer can load the geometry and assets without rebuilding the scene from scratch. What we are testing next KiMap is in free early access. The most useful feedback is not generic interest; it is a concrete blocker from someone building an indoor-nav

Week 3. "The deploy is done. Everything's green. Now what am I actually supposed to be looking at?" Previously Runner ↓ Cache ↓ Artifact ↓ Deployment Today ↓ Monitoring Junior Engineer: The canary rolled out fine yesterday. 100% traffic, all healthy. I closed my laptop. Was that wrong? Senior Engineer: Not wrong, exactly. But let me ask you something first. Your service is running on a server somewhere. Right now, this second — is it healthy? Junior Engineer: I mean... I assume so? Nobody's messaged me. Senior Engineer: "Nobody's messaged me" isn't an answer. It's the absence of one. That's the entire problem monitoring exists to solve. The Thing Nobody Says Out Loud Senior Engineer: Here's an uncomfortable fact about production systems: you cannot see them. Not directly. You're not standing next to the server, watching electricity move through it. Everything you know about whether it's healthy is a claim — something a piece of software told you, that you're choosing to trust. Junior Engineer: That sounds obvious when you say it, but I don't think I've ever actually thought about it that way. Senior Engineer: Most engineers don't, until the gap between "the system told me it's fine" and "the system is actually fine" bites them. Monitoring is the discipline of shrinking that gap — of making sure what you're told is close to what's actually true, and told to you fast enough to matter. 📒 Senior Engineer's Notebook You don't monitor a system because you don't trust it. You monitor it because you can't see it. Trust isn't the issue — visibility is. The Car Dashboard Analogy Junior Engineer: Can you make this concrete? Senior Engineer: Think about driving a car. You can't see the engine. You can't see the oil level, the coolant temperature, how much fuel is actually left in the tank, mid-drive. All of that is invisible to you, sealed inside metal, while you're doing 100 km/h. So the car gives you a dashboard. Speed, fuel, engine temperature, warning lights. You're not wat

Protocols let you define what a type can do without caring about what it actually is. Once...

I built a free tool that checks a supplier before you pay them. The part that took most of the work, and taught me most, was reading ten national company registers instead of relying on the EU's own VIES service. This is what I found out, mostly so the next person doesn't have to. The problem with "the VAT number is valid" VIES — the European Commission's VAT Information Exchange System — answers one question: is this VAT number currently registered. That sounds like the question you want answered. It isn't. A company that has gone into liquidation keeps a cleanly resolving VAT number in VIES. So does one that has been struck off the register. Deregistration and insolvency are run by different authorities on different timetables, and the gap between "this company has stopped being a going concern" and "the VAT number stops validating" can be months. So you can check a supplier, get a green tick, and be looking at an insolvency estate. The national registers know. VIES doesn't ask them. Ten registers, and what each actually gives you I found free, public, machine-readable-enough sources for ten countries: Bulgaria, Czechia, Estonia, Finland, France, Greece, Latvia, Poland, Romania and Slovenia. They are not equivalent, and this is the thing I'd have liked written down somewhere before I started: Six of them report company *state * — inactive, in liquidation, bankrupt, insolvent, terminated, ceased, struck off: Romania, Estonia, France, Greece, Bulgaria, Latvia. This is the valuable one. Three report whether the company is actually VAT-active — Poland, Romania, Slovenia. That matters more than it sounds, because VIES does not distinguish "this is a real company that isn't VAT-registered" from "this number belongs to nobody". The rest give you a name and not much more. Czechia, for instance, is in the ten but in neither of the other two groups. It confirms a name. That's it. Worth knowing before you build a feature around it. Poland is the interesting one Poland is the

Telnyx ships over 700 Ultra voices across 36 languages with sub-100ms time-to-first-byte. The voices are not the problem. Hearing them is. The docs list three. The Voices API returns 4,000+ across every provider. Voice pickers play a fixed sample sentence per voice. None of that tells you how a voice handles emotion, pacing, or character inside a real scene. So I built a small app that lets you do exactly that. You write a short scene with a few characters, assign each character a different Telnyx Ultra voice and an SSML emotion, and render the whole thing into one MP3. Every voice speaks in character, in context, in one continuous audio file. The Telnyx code example is: https://github.com/team-telnyx/telnyx-code-examples/tree/main/multi-character-narrator-python The Use Case Voice pickers exist. They play a fixed sample sentence per voice. What they do not do is let you hear a voice inside a real scene — a tense argument, a calm narrator, a panicked character, a reassuring guide — because a single sample sentence does not tell you how a voice handles emotion, pacing, or character. This example solves that. You write a short scene with a few characters. Each character gets a different Telnyx Ultra voice. Each character gets an SSML emotion. You hit render. The app fans out parallel TTS calls, stitches the per-line audio in script order, and plays you one continuous MP3 with every voice speaking in character. The default scene is the Ides of March from Julius Caesar. Five characters, ten lines, five distinct voices, five different emotions: Cassius — determined, plotting the assassination Caesar — surprised, realizing the betrayal Brutus — apologetic, justifying the act Mark Antony — angry, mourning the fallen leader Narrator — calm, setting the scene One render, one MP3, every voice in context. That is the demo. The Eight Curated Ultra Voices The app ships with eight pre-built Telnyx Ultra voices curated for the most common use cases. Each one is a real Telnyx voice

The Silent Disruptor The silence in the room was absolute, broken only by the rhythmic scraping of pens on paper during a high-stakes meeting. Then, it happened. My pocket erupted into a frantic, brassy ringtone that seemed to last an eternity before I could fumble to silence it. My face turned crimson as the room’s focus shifted from the presentation to my vibrating trouser pocket. I had remembered to check my calendar, but I had completely forgotten to toggle my phone to silent mode. That moment of pure, concentrated embarrassment was the catalyst for me building Muffle. The Friction of Manual Control We live in an age of automation, yet our phones—the very devices meant to assist us—remain stubbornly manual when it comes to basic social etiquette. Every day, millions of people walk into mosques for prayer, classrooms for lectures, or medical offices for consultations, and every day, a percentage of them forget to silence their devices. This isn't just a minor annoyance; it is a persistent source of social friction. Before I started building Muffle, I looked for existing solutions. Most apps were either bloated with unnecessary permissions, required invasive cloud accounts, or simply failed to trigger at the right time. The fundamental problem wasn't just the lack of features like GPS-based prayer times or calendar-specific automation; it was the lack of reliability. If an automation app fails once, the user loses trust in it forever. If I am in a meeting, I cannot afford for the app to 'sleep' because the system decided to save battery at the expense of my configured routine. I needed something that could handle these state changes consistently, regardless of whether the phone was in my pocket, sitting on a desk, or buried in a bag. Architecting for Reliability When I began writing the core logic for Muffle, I immediately hit the wall that every Android developer eventually faces: Doze Mode. Android’s aggressive power management is designed to preserve battery by

Environment Variables the Safe Way Environment variables are the standard way to configure applications without hardcoding secrets or environment-specific details. But they're easy to misuse. I've seen API keys committed to repos, configs that crash when a variable is missing, and defaults that silently override production settings. Here's how I handle them safely. Never Commit Secrets The most important rule: never put real secrets in your code or commit them to version control. That includes .env files. Add .env to your .gitignore immediately. If you're using a framework like Laravel or a tool like Vite, the default .env.example is your friend. Commit that, but never the real one. For local development, you can generate a .env from the example and fill in your own values. For production, set variables through your hosting provider's dashboard or a secrets manager like AWS Secrets Manager or HashiCorp Vault. Read Variables Explicitly Don't access process.env directly all over your codebase. Instead, centralize your configuration. Create a config.js (or config.ts ) that reads and validates all the variables you need. // config.js const required = [ ' DATABASE_URL ' , ' JWT_SECRET ' , ' PORT ' ]; const missing = required . filter ( key => ! process . env [ key ]); if ( missing . length ) { throw new Error ( `Missing required environment variables: ${ missing . join ( ' , ' )} ` ); } module . exports = { databaseUrl : process . env . DATABASE_URL , jwtSecret : process . env . JWT_SECRET , port : parseInt ( process . env . PORT , 10 ) || 3000 , }; Now your app imports config and uses config.port . This has several benefits: Fail fast: if a required variable is missing, the app crashes at startup, not later when you try to use it. Type safety: you can parse and validate values once. Easy to mock in tests. Use Defaults Carefully Defaults are convenient, but they can hide problems. For example, if you default PORT to 3000 in production, you might accidentally run on the w

This is a submission for DEV's Summer Bug Smash: Smash Stories powered by Sentry. This happened a...

Lucid's new CEO Silvio Napoli listed four must-win priorities, including the successful launch of its midsize EV, finishing a factory in Saudi Arabia, cutting expenses, and robotaxis.

Anthropic shipped OS-level sandboxing for Claude Code. If you run an agent against a repo you care about, it's worth understanding precisely what moved — because a fair amount of the commentary treats it as "agents are contained now," and that's not what the documentation says. I read the docs carefully, partly because I build a tool in adjacent territory and needed to know whether I'd just been made redundant. Short answer: no. The longer answer is more interesting, and it starts with a compliment: the docs are unusually honest about their own limits. Most of what follows isn't something I discovered — it's something Anthropic wrote down, and more people should read it. What it actually does The sandbox uses OS primitives — Seatbelt on macOS, bubblewrap on Linux and WSL2. By default, sandboxed commands can write only to your working directory and the session temp directory. No network domains are pre-allowed: the first time a command needs a new host you're prompted, and approving it lasts the session. Crucially, this is enforced by the operating system on the running process , not by the model correctly interpreting a command. The docs put it well: the boundary holds regardless of what the model chose to run, and even if an allowed command does more than its name suggests. That's a real improvement over asking an agent nicely, and it's the right layer for what it solves. The motivation named in the docs is the same one I keep seeing in the wild: reducing the permission prompts that people stop reading. Approval fatigue is the disease; this is a real treatment for part of it. Five things worth knowing before you rely on it It's Bash-only. The sandbox constrains Bash commands and their child processes. Claude Code's own Read, Edit and Write tools don't run through it — they go through the permission system instead. "The sandbox is on" means shell commands are contained, not that every file operation is. Your working directory is inside the boundary by design. The de

Apple shipped Liquid Glass across iOS 26 and macOS, and suddenly every product I look at has a frosted panel floating over something. I spent a few weeks rebuilding the effect properly for a project, and most of what I found online stops at one line: backdrop-filter : blur ( 16 px ); Which gives you a gray rectangle. That's not what makes Apple's version look like glass, and figuring out the difference took me longer than it should have. So here are the six techniques I ended up with, roughly in order of how well they're supported, along with the things that wasted my time. 1. The plain glassmorphism card Everyone knows this one, but there are three parts to it and most implementations ship only the first. .glass-card { position : absolute ; inset : 20% ; border-radius : 16px ; backdrop-filter : blur ( 16px ) saturate ( 180% ); -webkit-backdrop-filter : blur ( 16px ) saturate ( 180% ); background-color : rgba ( 255 , 255 , 255 , 0.08 ); border : 1px solid rgba ( 255 , 255 , 255 , 0.12 ); box-shadow : 0 8px 32px rgba ( 0 , 0 , 0 , 0.2 ); pointer-events : none ; } The saturate(180%) is the part I kept forgetting, and it turns out to be the whole trick. Blurring averages colors together, and averaging colors drains saturation out of them — so a pure blur comes out looking like dirty plastic rather than glass. Pushing saturation back up compensates. Drag it down to 100% in the pen above and you'll see the effect just die. The background tint matters for a similar reason. With a fully transparent background you get a blur but no surface — nothing reads as a physical pane sitting there. Something around 8% white is enough to suggest one without washing out whatever is behind it. Wrapped in React, so the numbers are adjustable: " use client " ; type GlassCardProps = { blur ?: number ; saturate ?: number ; opacity ?: number ; radius ?: number ; }; export default function GlassCard ({ blur = 16 , saturate = 180 , opacity = 0.08 , radius = 16 , }: GlassCardProps ) { return (

A status page has a blind spot. It can tell you that your API is returning 502s. It can tell you that a TCP port stopped accepting connections. It can tell you when the incident started. It usually cannot tell you why . Was the application host out of memory? Was disk I/O saturated? Did load climb for 40 minutes before users noticed? Was the server completely healthy and the real problem somewhere else? Those answers often live in a separate monitoring product, disconnected from the incident timeline and disconnected from the status page. That is why I built Servers for StatusPage.me. It is a small, customer-installed host metrics agent and dashboard. You install it on a machine you operate, and it reports CPU, memory, swap, load, disk, and network metrics back to your account. The important part is not “now there are more graphs.” The important part is seeing an outage and the host evidence around it on the same timeline. External checks answer one question. Host metrics answer another. Regular uptime monitoring is still the right tool for the outside-in view: Can users reach the website? Is the API returning the expected response? Does DNS resolve correctly? Is the database port open? Did a scheduled job run? But those checks do not run inside your infrastructure. A healthy HTTP response does not prove that a background worker is about to run out of memory. A timeout does not prove that the app server is overloaded. And an incident can start with a slow disk or growing swap usage long before an endpoint is fully unavailable. The distinction is simple: External monitoring tells you what users can see. Host metrics help explain what the machine was doing when they saw it. You need both. What Servers includes Each registered host gets a dedicated dashboard page with: CPU user, system, and I/O wait utilization Memory use Swap use Load averages Disk use and read/write throughput Network inbound and outbound throughput A human-readable OS description for account owners

I'm a brand strategist, not a developer — but every campaign I run eventually turns into a conversation with someone's engineering team. Over the past year, that conversation has shifted. It used to be about meta tags and sitemaps. Now it's about whether a site is even readable by the models powering AI answers. Here's the technical breakdown I actually walk dev teams through. SEO (Search Engine Optimization) This is the one most engineers already know. Crawlable HTML, clean URL structure, fast Core Web Vitals, valid schema.org markup, an accurate sitemap.xml and robots.txt. The mechanics haven't changed much — what's changed is how much weight structured data carries now, because it's the same markup that AEO and GEO systems lean on. AEO (Answer Engine Optimization) This is about formatting content so it can be lifted directly into a featured snippet or a voice/chat answer. Practically, that means: a direct, self-contained answer to the implied question within the first 1-2 sentences of a section, genuine FAQ schema ( FAQPage in JSON-LD, not just visually-styled accordions), and heading structure that maps to actual questions people ask, not just keyword strings. If a section can't be understood correctly when read on its own, out of context, it won't get picked up. GEO (Generative Engine Optimization) This is the newest layer, and it's aimed at large language models rather than traditional crawlers — think AI Overviews, Perplexity, ChatGPT's browsing mode. A few things I've seen actually move the needle here: an llms.txt file at the root (still informal, not a ratified standard, but increasingly respected), consistent factual claims about an entity across every page and every third-party mention (NAP consistency isn't just a local-SEO thing anymore, it's an entity-recognition thing), and content that states things plainly rather than burying them in marketing language — generative models tend to extract and cite the most unambiguous sentence in a block, so ambigui

I have been a huge Storybook and Chromatic fan for years. But at some point the bill got my attention, and when I looked into why, the fix turned out to be simple. This is the write-up of what I changed. It works on any per-snapshot tool, not just Chromatic. First, some backstory on how I got here, because it explains why the cost crept up in the first place. How I ended up paying for a lot of snapshots In the past I would build a gigantic end-to-end pipeline that was flaky as hell and made me spend time every week fixing it. It took 40 minutes to run, and when it went red someone would assume it was just flaky, merge the change anyway, and then find out it truly did break the system. So I stopped writing lots of E2Es and moved to Storybook for interaction and visual testing. Much better. But because I was rendering every state of every component as its own story to get the screenshots in place, I was generating a lot of screenshots. And every snapshot tool, Chromatic, Percy, Playwright screenshots, UI Verify, renders and bills per story. So the number of stories is the cost, and it is also the noise surface: more stories means more places for a diff to flake. I ended up paying a lot, which made me think about whether there were ways to optimise it. There were. Here they are. The core idea: combine states into one story The naive pattern is one story per variant times state times theme. A component with 5 sizes, 3 states, and 2 themes is 30 snapshots the naive way. The whole idea below is to collapse that matrix into a handful of stories while keeping full coverage. Move 1: one gallery story, not N stories For something like a Button, there is no need to have separate Primary, Secondary, and Tertiary stories. I prefer one AllVariants story that maps through the prop combinations and renders them in a grid. One snapshot then covers the entire matrix. As a bonus you get a nice grid that shows every permutation at a glance, with no extra clicks to see the variations. /

Following up on a common claim in the AI avatar space — that voice/video avatars convert better than plain text chat — there's surprisingly little rigorous testing behind it. If you're building or embedding one of these widgets, here's a practical way to actually measure it instead of trusting vendor case studies. Why This Is Harder Than a Normal A/B Test Standard A/B testing swaps one variable (a button color, a headline) while holding everything else constant. Avatar vs. text chat isn't that clean — you're changing interaction modality, response latency expectations, and visual real estate simultaneously. You need to isolate the variable that actually matters: does voice/video presence drive conversion, independent of the underlying conversation quality? A Cleaner Experimental Setup javascript // Pseudocode for variant assignment function assignVariant(sessionId) { const hash = hashSessionId(sessionId); return hash % 2 === 0 ? 'avatar' : 'text'; } Key controls to hold constant across both variants: Same LLM backend and prompt/knowledge base — the conversation logic shouldn't differ, only the presentation layer Same lead capture form and CTA placement — don't let UI differences beyond avatar-vs-text confound the result Same traffic source — segment by acquisition channel if traffic mix varies, since paid vs. organic visitors convert differently regardless of chat UI Minimum sample size before evaluating — novelty effects are real; running this for 3 days will overstate the avatar's lift. Run for at least 2-3 weeks to let novelty decay. Metrics to Track (Not Just Conversion Rate) Conversion rate alone hides why one variant wins or loses: session_start_to_first_message (engagement friction) message_count_per_session (depth of interaction) time_to_form_completion (avatar/video adds latency — does it cost or gain time?) bounce_rate_before_first_response lead_quality_score (if you can grade downstream — a lead isn't a conversion if it's junk) A common finding worth watc

How EvalPort's Grader System Works When designing EvalPort, the grader system was the hardest part to get right. Every eval framework has its own way of scoring LLM outputs — DeepEval uses metric classes, Promptfoo uses assertion objects, Inspect AI uses solver functions. We needed a system expressive enough to cover 90%+ of real-world eval needs, but simple enough that any framework could implement it. The result: 11 grader types that carry their own semantics. A grader isn't just a name — it specifies its parameters, its model, its threshold. An eval suite is self-describing. The 11 Grader Types exact_match — Compare output to expected output, optionally ignoring case. contains — Check if the output contains a substring. regex — Match against a regular expression. semantic_similarity — Embed output and expected output, compare cosine similarity against a threshold. llm_judge — Use an LLM to evaluate the output against a prompt template. The most powerful grader. json_schema — Validate that the output is valid JSON matching a JSON Schema. json_path — Extract a value from JSON output using a JSONPath expression, then compare it. code — Run a function to evaluate the output. human — Defer to human review. model_graded — Compare the output to a reference answer using a model. custom — Escape hatch for graders not covered by built-in types. How Graders Connect to Test Cases A test case references graders by ID. Multiple graders can evaluate the same test case. The ResultSet records each grader's score separately. Why This Design Works Self-describing: An eval suite carries everything a framework needs to execute it. Framework-agnostic: Any framework can implement any subset of grader types. Extensible: The custom type lets frameworks bring their own graders. Comparable: Results from different frameworks use the same grader IDs. Try It pip install evalport-sdk npm install evalport-sdk Spec: https://github.com/adhabnr-ux/evalport/blob/main/spec/SPEC.md Repo: https://gith

I built a debugging-practice site where student code runs entirely in the browser . Python via Pyodide , JavaScript in a worker. No server executes anything. No execution bill, no queue, no sandbox to maintain. But four things bit me hard. 1. Your arguments aren't Python objects Pass a JS object into Python and you get this: TypeError: 'pyodide.ffi.JsProxy' object is not subscriptable It's not a dict . It's a live view of the JS object, and it supports neither obj[key] nor .get() . Convert explicitly: const pyArgs = input . map (( arg ) => pyodide . toPy ( arg )); const result = fn (... pyArgs ); 2. null is not None This one passed my entire test suite while being broken in production. pyodide . toPy ( null ) check result type(v) JsNull bool(v) False ✅ falsy, as expected v is None False ❌ the surprise It's falsy, so truthiness checks work fine. But is None fails — which was exactly what my code was checking. Why my tests missed it: the harness used json.loads . The app used toPy . Different conversion paths, different answers. If you need a real None , create it in Python. Don't pass one across. 3. sys.settrace is a free step debugger Want to show users their code running line by line? Python basically hands it to you: def _tracer ( frame , event , arg ): if frame . f_code . co_name != target : return None # skip library frames if event == " line " : steps . append ({ " line " : frame . f_lineno , " locals " : dict ( frame . f_locals ), }) return _tracer Two things this naive version gets wrong: Add a step cap. A tight loop generates steps faster than it burns a 5-second timeout. You need both guards. Handle exception . During unwinding, the return event still fires with arg=None . Miss it and your trace says "returned None" for code that crashed. 4. Your snapshots are lying A user screenshot exposed this one. Every step in the trace showed the final state of a list. Step 1 included mutations that hadn't happened yet. tracing: nums = []; nums.append(1); nums.append(

The USB port in your car is likely not powerful enough to charge your laptop, but there are other options.