Articles
Web, React and Next.js, AI and RAG, and mobile. Written from production work, not tutorials about tutorials.
Mem0, Zep, Letta, LangMem and GoodMem on the same 419-turn conversation, graded only on the text they retrieved. Vendors publish 80-93% on LoCoMo. The best here scored 57.6%. Open harness, every number reproducible.
Agents in ProductionDuring a model outage, a fallback on a review-reply product I lead posted generic English replies to Swedish businesses' Google profiles for days, and every dashboard stayed green. This is the publish gate and the per-path fail-closed rule I use now.
Growth EngineeringA 30-mailbox Smartlead fleet quoted at 900 emails a day could reach about 250 new people. How I found 11,556 leads queued and unsent, the per-mailbox setting capping every box, and the Outlook accounts that kept pausing everything.
AI EngineeringI priced a custom AI support agent on Claude Sonnet 5 against Intercom Fin's $0.99 per outcome at 1,000, 5,000 and 20,000 conversations a month, with every assumption shown and the costs most quotes leave out.
RAG in PracticeCheaper long context in 2026 broke the old always-RAG advice. Here is the decision tree I use: when full-context prompting beats a pipeline, when RAG is mandatory, when fine-tuning earns its keep, plus the hybrid stack and a cost and latency comparison.
AI Product EngineeringApple's 3rd-gen on-device Foundation Models let mobile teams run classification, routing, and structured extraction for free, on-device, and offline. Here is the decision framework I use to choose on-device vs a cloud LLM, where on-device tops out, and the hybrid fallback pattern that sends only the hard cases to a paid API.
Freelance & CareerWhen an autonomous agent you built takes a wrong action in production, the law often attributes that action to whoever deployed it, and the contract decides who actually pays. Here are the six clauses I put in every AI agent engagement.
Web PerformanceOn 18 March 2026 Google made INP an equal ranking signal and dropped the "Good" LCP bar to 2.0s. Here is the React-specific triage list I run: transitions and deferred updates, long-task breaking, the React Compiler, streaming SSR, and a hard third-party script budget.
MCP & Agent ProtocolsAn end-to-end guide to building your first MCP App: declaring a UI resource, the sandboxed-iframe security model, routing UI actions through the same audited tools/call path, and what changed when Apps became an official extension in the 2026-07-28 spec.
Freelance & CareerWhen AI cuts a 20-hour job to 5 hours, hourly billing quietly cuts your pay 75 percent for the same result. Here is the repeatability-vs-measurability matrix I use to decide whether to bill hourly, productize, or value-price in 2026.
Mobile EngineeringReanimated 4 is New-Architecture-only and moves worklets into a separate react-native-worklets package. Here is the real migration off Reanimated 3 for a gesture-heavy app: the package split, the new CSS-style API vs the imperative one, and what RN 0.85's Shared Animation Backend lets you stop hand-rolling.
RAGA production architecture for a docs chatbot that grounds every claim in retrieved sources, links each claim to a source, and refuses rather than fabricates when retrieval confidence is low. Retrieval, rerank, validation, and the abstention threshold that moves correct deflection.
Mobile EngineeringGoogle Play flagged your AAB for the 16 KB page size requirement? Here is how I diagnose which native .so libraries are misaligned with check_elf_alignment.sh, bump the NDK and AGP to fix my own code, and handle the React Native dependencies that are the usual culprits before the 31 May 2026 enforcement.
AI EngineeringFinal-answer assertions keep your CI green while your agent calls the wrong tool, loops extra steps, and quietly burns budget. Here is the trajectory-scoring eval harness I run: deterministic tool mocks, six scored dimensions, and an LLM judge kept on a rubric and audit.
Next.jsNext.js 16.2 makes AI agents first-class users: create-next-app writes a version-matched AGENTS.md (100% vs 79% eval pass rate), @vercel/next-browser hands an LLM screenshots, network, and console in one call, and experimental Agent DevTools exposes framework state. Here is how I set it all up.
Agents in ProductionLong-running agents do not fail because they run out of tokens. They fail because accuracy drops as the context grows. Here is the compaction playbook I use to keep agents reliable past 100k tokens: trigger-based compaction, anchored iterative summarization, per-tool context feedback, and offloading state to external memory.
Freelance & Career2026 is the year the AI-built app bills come due. Here is how I audit a vibe-coded codebase in a day, price the rescue, and decide whether to refactor in place or stage a rebuild without freezing the business.
AI AgentsA SKILL.md Agent Skill is a portable folder that loads the same workflow into Claude Code, Cursor, Copilot, Codex, OpenCode, and Kiro. I show how progressive disclosure keeps activation cheap, walk through a real reusable client skill, and explain why one open standard beats rebuilding the same workflow per tool.
ReactAfter enabling React Compiler 1.0, my React Hook Form fields stopped updating. Here is why RHF's mutation-based internals fight auto-memoization, the precise 'use no memo' scoping that fixes it, the eslint lint plus autofix that catches it, and a clear rule on whether to patch now or wait for RHF v8.
RAG in PracticeIf your SaaS already runs on Postgres, you can add real semantic search in a weekend with pgvector and one Next.js route, no second database. Here is the embed-on-write migration, the HNSW query, a semantic cache, and the vector count where Pinecone or Milvus actually starts to make sense.
SecurityMCP adoption is racing ahead of its security. Around 43% of public MCP servers carry an exploitable flaw and tool-poisoning hides in metadata that auto-approve never shows you. Here is the pre-integration audit I actually run before I let any third-party server near a real agent.
Mobile EngineeringReact Native 0.85 fully removed the bridge on 7 April 2026, so the New Architecture is the only option. Here is how I audit dependencies against the Directory, which libraries still crash, and the realistic effort to rescue a 0.7x Paper app.
Next.jsNext.js 16 renamed middleware.ts to proxy.ts and moved it to a Node-only runtime, so Edge-dependent auth can silently break on upgrade. Here is the exact fix: the codemod, the renamed config flags, the next-intl gotcha, and when to deliberately stay on the old file.
GEO & AI VisibilityGoogle AI Mode is now the default, and it splits one query into many sub-queries before retrieving per sub-query. Here is what query fan-out actually does under the hood, from someone who builds fan-out RAG pipelines, plus a concrete playbook for restructuring your pages to get surfaced.
Toolingtsgo is fast, but it has no JS plugin or transformer support and no stable compiler API until 7.1. Here is the dual-compiler tsconfig and CI setup I use to type-check with tsgo while keeping tsc for emit, plus a checklist of the tooling that breaks quietly.
AI Product EngineeringI benchmarked GPT-5.2 Strict Mode against Claude tool use on 200 messy extraction documents. GPT-5.2's CFG token masking guarantees valid JSON and hit ~92% semantic correctness; Claude won on tool orchestration and clean refusals. Here's how to pick, with the refusal-as-error pattern that keeps either one reliable.
Next.jsMigrating off unstable_cache to Next.js 16 'use cache' is easy to get wrong in one specific way: a cached function that reads the tenant from request context instead of taking it as an argument will serve one tenant's data to another. Here is the function-scope-first playbook and audit checklist I use on production multi-tenant SaaS.
GEO & AI VisibilityIn 2026 GPTBot, OAI-SearchBot, ClaudeBot, and PerplexityBot still fetch initial HTML only and do not run your JavaScript. Here is a curl-with-bot-UA test harness to prove what they see, plus the App Router fix that puts your content back in the server-rendered HTML.
Agents in ProductionOpus 4.8 makes hundreds of parallel subagents practical for codebase-scale work. Here is how I scope the fan-out, why the verification gate matters more than the model, and a real token and cost model with the fast mode that runs roughly 3x cheaper, plus where it overspends.
AI SecurityPrompt injection is OWASP's number one LLM risk and no input filter fixes it. The real defense is architectural: detect when an agent holds all three lethal trifecta conditions at once, then break one of them. Here is how I do it with capability scoping, egress allow-lists, and human approval on privileged calls.
SecurityCVE-2025-55182 is a CVSS 10.0 remote code execution flaw in the React Server Components deserialization path. If your app renders RSC payloads, you are exposed even without Server Functions. Here is the exact patched version for every react-server-dom line and Next.js release, plus WAF rules to buy time.
Mobile EngineeringGoogle Play makes target API 36 (Android 16) mandatory on August 31, 2026. The version bump is easy. The forced edge-to-edge display and predictive back changes that silently break your React Native layouts are not. Here is the triage checklist and EAS verification routine I run before the cutoff.
MCP & Agent ProtocolsThe 2026-07-28 MCP spec makes remote servers stateless by default. Here is how I drop the initialize handshake, remove the Mcp-Session-Id header, move sticky-session servers behind a plain round-robin load balancer, and survive the Tier-1 SDK gotchas inside the validation window.
AI EngineeringStrict structured outputs hold ~99.9% schema compliance while plain JSON mode fails 8-15% of the time. I compare OpenAI, Claude, and Gemini with one Zod schema.
AI EngineeringMost AI agents don't need a memory vendor. Updated with measured results: Mem0, Letta, Zep, LangMem and GoodMem on one 419-turn conversation, where the best scored 57.6%.
AI EngineeringThe final MCP spec ships July 28, 2026 and removes sessions from the protocol. I migrated my production Node server; here is the exact diff and checklist.
AI EngineeringYes, you can ship private RAG on one 24GB GPU in 2026. I ran a 50-question eval: Gemma 4 26B MoE wins English corpora, Qwen3.6 27B wins multilingual.
AI EngineeringI audited my production MCP stack against the NSA's May 2026 guidance and the OX Security RCE disclosure. Here is the 12-point hardening checklist I use.
AI EngineeringPricing converged at $20/$200 and SWE-bench scores sit within a point, so workflow decides. Real cost-per-feature numbers from paid client projects.
AI EngineeringNo. I rebuilt my production SaaS pipeline as agentic retrieval: cost per query down 36%, accuracy up from 68% to 89%. Only naive top-k RAG died in 2026.
Web DevelopmentGEO means writing answer-first chunks AI engines can lift: 69% of Google searches are zero-click in 2026, but ChatGPT referrals convert at 15.9%.
AI EngineeringI cut my reputation SaaS's LLM bill 79%, from $41.60 to $8.90 per 1,000 AI replies, using routing, prompt caching, semantic caching, and batching.
Web DevelopmentSelf-hosting Next.js on Hetzner with Coolify costs me $6-17/mo vs $40-150 on Vercel Pro. The real math, ops hours included, after the April 2026 breach.
AI EngineeringOTel GenAI spans went stable in early 2026. Here is how I instrument a TypeScript agent in Node.js, track cost per trace, and alert on silent failures.
Web DevelopmentFor referential equality, no: React Compiler 1.0 handles it. I profiled a production dashboard to show what you still hand-memoize and what breaks.
Web DevelopmentNext.js 16 breaks middleware, revalidateTag, and caching with zero errors. Exact symptoms and fixes from migrating my production SaaS in 14 hours.
Web Developmenttsgo cut my monorepo's full type check from 71s to 9.2s, but plugins and compiler-API tools broke. Real before/after numbers and a switch-or-wait verdict.
Web DevelopmentWebMCP, announced at Google I/O 2026, lets your web app register typed tools AI agents can call in Chrome 149. Here is how I exposed mine, with code.
Mobile DevelopmentGoogle Play now blocks all React Native updates without 16KB page-size support. Here's the fix: RN 0.77+, AGP 8.5.1+, NDK r28, plus updated native deps.
Web DevelopmentRoughly 8,000 of 10,000 AI-built startup apps need rescue work in 2026. Here is the 60-minute triage and fix order I use to save them without a rewrite.
Web DevelopmentVibe code to validate, hire an engineer before money flows. A five-question framework with six-month cost math from a developer who rescues vibe-coded apps.
Mobile DevelopmentExpo Go for SDK 55/56 was never approved on the App Store. Here is my tested migration to development builds, plus SDK 56 gotchas and real costs.
Mobile DevelopmentWWDC 2026 made on-device AI the default for React Native. I built a private review summarizer: no API key, $0 per call, 28 tokens/sec on iPhone 16 Pro.
AI EngineeringAn MVP RAG chatbot costs $4k-$12k to build in 2026, production builds run $15k-$40k+, and running costs land at $5-$30 per 1,000 queries. Full breakdown.
Mobile DevelopmentFor most startup MVPs I pick React Native: it reuses React web talent, code, and hiring pipelines. Flutter wins for custom UI-heavy apps. My framework inside.
AI EngineeringYou do not need a rewrite to add AI to a SaaS. Add one endpoint beside your existing API, wire it to one workflow, stream the output, and cap token spend.
I take on a few projects at a time: web apps, AI features, and mobile. Tell me what you are working on.
Start a conversation