Writing
Guides, release breakdowns and lessons from shipping AI products and agents at Modelia: Claude, MCP, agent workflows and the engineering underneath. For the week's AI news with my take, read AI Pulse.
Latest
An AI Product Engineer owns an AI feature from the user's problem to the production system: picking the model, writing the prompts and evals, building the product around it, and answering for its quality and cost after launch. Here's how the role differs from ML and software engineering, the skills it needs, and what the work looks like day to day.
8 min
AI spend usually grows from a few habits: one default model for every task, the same long prompt sent thousands of times, and nobody knowing which feature costs what. Here's the five-step audit I run on a codebase, with TypeScript examples for model routing, response caching and per-feature metering.
8 min
AI features fail in ways CI never catches: a quality drop on real inputs, a cost spike from long prompts, one bad output that ends up in a screenshot. Here's the rollout playbook I use instead of a big-bang launch: blast radius, flags and a kill switch, guardrails, a staged exposure table and the metrics that decide each step.
Archive
39 articles shown
An AI Product Engineer owns an AI feature from the user's problem to the production system: picking the model, writing the prompts and evals, building the product around it, and answering for its quality and cost after launch. Here's how the role differs from ML and software engineering, the skills it needs, and what the work looks like day to day.
AI spend usually grows from a few habits: one default model for every task, the same long prompt sent thousands of times, and nobody knowing which feature costs what. Here's the five-step audit I run on a codebase, with TypeScript examples for model routing, response caching and per-feature metering.
AI features fail in ways CI never catches: a quality drop on real inputs, a cost spike from long prompts, one bad output that ends up in a screenshot. Here's the rollout playbook I use instead of a big-bang launch: blast radius, flags and a kill switch, guardrails, a staged exposure table and the metrics that decide each step.
Since September 2026 the Claude app works on Apple CarPlay, but it's a voice chat, not Claude Code. Here's what Claude on CarPlay can do, every way to reach Claude Code away from your desk, and how to talk to Claude Code hands-free while you drive.
A step-by-step guide to running Claude Code with Siri: one install command, one click to add the shortcut and three iPhone settings, then "Hey Siri, Jarvis" works from your iPhone, AirPods or CarPlay. Plus how the shortcut works and how to fix common problems.
A lot of engineering work is just asking questions. Here are 25 things to ask Claude Code by voice on your commute, grouped by job (PRs, CI, the codebase, small fixes, Slack, email and your calendar), with what actually happens for each and how to phrase requests so they work.
The engineering behind talking to Claude Code through Siri: one warm Claude Code process per device with stream-json (median 5 seconds per answer), an 8-second reply budget with background jobs, joining half-sentences, and the bug where every answer came one question late. With code and numbers from 109 real turns.
In two weeks, OpenAI, Atlassian, Oracle and SAP each shipped agents with named owners, scoped permissions, approval rules and audit logs, and America.gov tested whether citizens will let AI act for them. Here's what launched, why only 23% of consumers trust agents with payments, and a working TypeScript agent envelope you can copy.
OpenAI Dots are persistent ChatGPT agents launched at DevDay on September 29, 2026. Each one runs on GPT-6 Astra with its own cloud computer, connects to 4,000+ apps and keeps working after you log off. Here's how they work, what they cost, what they can't do yet, and an undo-ledger pattern for long-running agents.
Atlassian announced AMP, the Agentic Multiplayer Protocol, at Team '26 Europe on October 7, 2026. Every agent gets an owner and profile, works from the 250-billion-object Teamwork Graph and logs its actions. Here's how it works, what's shipped vs coming soon, an admin checklist, and a permission model you can copy.
Oracle announced Fusion Claw on September 29, 2026: a governed agent runtime inspired by OpenClaw, with an Enterprise Operating Envelope, an isolated runtime that keeps the model away from business records, and an Outcome Receipt for every task. Here's how it works, the 25 launch apps, the open questions, and a propose-validate-apply pattern you can copy.
Shoppers welcome AI that helps them decide, but few will let it pay: 23% trust GenAI with payments, and only 5% want an agent buying on its own. Caps, returns and instant revocation change that. Here's the 2026 survey data from Visa, Checkout.com, Cover Genius and RTB House, and a spend mandate you can enforce in code.
Anthropic released Claude Haiku 5.5 on October 7, 2026: $0.10 input and $0.50 output per million tokens, a 1M-token context window, adjustable effort for the first time on a Haiku model, and benchmark scores that beat GPT-6 Luna. Here's the pricing math, the benchmarks, working code, and how to use it as a subagent under Opus 5.5 and Sonnet 5.5.
SpaceXAI released an experimental official TypeScript SDK, @xai-official/sdk: Grok text, voice, image and video in one zero-dependency package, plus server-side X search, web search, code execution and remote MCP. Here's working code for each piece and the production gotchas hidden in the README.
I built carcode, an open-source way to code from your car. Say "Hey Siri, Jarvis" and Claude Code reviews PRs, reads Slack and ships code changes on your Mac while CarPlay reads the answer aloud. Here's how it works, what broke on real drives, and how to set it up in five minutes.
On June 30, 2026 the US lifted the export controls it placed on Claude Fable 5 and Mythos 5, and Fable 5 came back online globally on July 1. Here's the full timeline, what changed under the hood — a 99%+ bypass classifier, refusals, and fallback billing — and what developers need to do to ship on claude-fable-5.
A widespread outage hit Facebook, Instagram, and Messenger on June 12, 2026, with users across the US, Europe, Asia, and the Middle East reporting blank feeds, forced logouts, and 'query error' messages. Here's what's confirmed, what isn't, and what engineers can learn from Meta's outage history.
Anthropic shipped Claude Opus 4.8 on May 28, 2026. SWE-Bench Pro jumps from 64.3% to 69.2%, Fast Mode runs 2.5× faster at 3× lower cost, and Dynamic Workflows lets Claude Code orchestrate hundreds of subagents. Here's the full developer breakdown.
Google's Gemini CLI shuts down June 18, 2026. A practical, step-by-step migration guide to Antigravity CLI with before/after code, env vars, auth, and CI fixes — done in 15 minutes.
Anthropic's unreleased Claude Mythos model uncovered 10,000+ zero-day vulnerabilities in a month through Project Glasswing. What AI-powered vulnerability discovery means for developers, dependency security, and the future of cybersecurity in 2026.
GitHub disclosed a supply chain breach this week exposing internal code and customer data. I audited Modelia's dependency tree right after — here's what I found, what you should check, and the patching playbook for 2026.
Meta, Amazon, Microsoft and Google are spending record amounts on AI infrastructure while cutting 170,000 jobs in 2026. A grounded, honest read on what this actually means for engineers 1–3 years into their career.
Model Context Protocol hit 97 million installs in 2026 but most tutorials stop at hello-world. A hands-on guide to building production MCP servers with real code, auth, tool registration, and observability — from a production engineer at Modelia.
Gartner says 60% of code will be AI-written by end of 2026. Stack Overflow says 84% of developers use AI but only 29% trust it. I've shipped production systems with AI daily for 18 months — here's what's real, what's hype, and what changes for your job.
Shopify just shipped an AI agent that reviews your app for App Store compliance before submission. How the new self-review tool, Partner Dashboard tracking, and automated pre-submission checks change Shopify app review in 2026.
Anthropic released Claude Opus 4.7 on April 16, 2026. A hands-on look at what changed, how it compares to Opus 4.6, the new xhigh effort level and task budgets, and what the Project Glasswing safeguards mean for production AI systems.
How we built autonomous AI agents at Asynq.ai and Modelia.ai that reason, plan, and execute multi-step tasks. Practical patterns for tool use, memory, orchestration, and guardrails in production agentic systems.
How we use Anthropic's Claude API to power agentic workflows at Modelia.ai — from tool use and extended thinking to structured outputs and multi-turn agents. Practical patterns for building reliable AI agents with Claude.
Lessons from running agentic AI systems at scale — how we handle failures, monitor agent behavior, manage costs, and scale from single agents to multi-agent pipelines at Modelia.ai and Asynq.ai.
A comprehensive guide to architecting type-safe microservices using tRPC, Prisma ORM, and Node.js. Learn patterns from Asynq.ai and Modelia.ai's studio platform to achieve 40% performance improvements.
Learn techniques to improve database performance by 40%. Real-world examples from production systems at Asynq.ai and Modelia.ai handling millions of requests daily.
Modern strategies to build blazing-fast React applications. From code splitting to memo patterns, real techniques from building EduFly and GyanSathi LMS.
From development to production: containerization strategies that ensure consistency, security, and scalability. Lessons from deploying EduFly on AWS.
Master the new App Router, Server Components, and data fetching patterns. Everything you need to build production-ready Next.js applications.
Insights from building frontend interfaces at Bharat Electronics Limited (BEL) for the Indian Airforce — rigorous processes, reliability, documentation, and the weight of critical systems.
Implementing effective caching patterns that reduced our API response times by 60% at Asynq.ai. Learn cache invalidation, TTL strategies, and common pitfalls.
Go beyond basics with discriminated unions, template literals, and conditional types. Real patterns from enterprise codebases at Modelia.ai.
Cost-effective cloud architecture that grows with your startup. How I architected EduFly to serve 15K+ users on AWS with 99.9% uptime.
Automate your deployment pipeline from testing to production. Real workflows from EduFly, Asynq.ai, and Modelia.ai deployments.
Keep reading
One email when a new article or AI Pulse edition goes out. Nothing else.