Insights
Practical guides on AI coding agents, LLM costs, websites and custom software: what we built, what it cost and what we would do differently.
For a client project, hosted or self-hosted is a decision about duties, not taste. Six questions about data location, restore, patching, handover and exit, with the honest cases on each side.
Running your own PostgreSQL is reasonable for a side project and a real commitment for a business. The software is the same; the duties are not.
Supabase can cost nothing hosted, locally or on your own server, but each option asks for something different. Here are the real limits, the self-hosting work, and how PocketBase, Appwrite and Nhost compare.
A database that works on your laptop can be reachable by anyone once the project is public. The Supabase linter flags the tables, policies and keys that let that happen.
A local Supabase reset takes seconds and discards everything that is not in a migration. Here is what a dump does and does not contain, and a routine you can restore.
In Supabase the database is reachable through an API, so row-level security is the access control. Here is what it does, the mistake that exposes a table, and how to test a policy as anon and as a signed-in user.
A second local Supabase project collides on the same ports, and hand-offset ports drift away from your app config. See what the CLI now assigns and what it costs.
A schema change belongs in a file that every environment applies the same way. Here is how Supabase migrations are created, applied, reset and reviewed, and why a change made in Studio is not one yet.
An audio version of a long page now costs minutes to produce, which makes it an accessibility measure rather than a premium feature. Who it helps, what the EAA asks and how to build it.
An agent that cannot search only knows its training data. We compare paid search APIs and a self-hosted SearXNG instance on cost per query, privacy and what breaks.
A demo becomes a handout when nobody can run its tests. Here are the money invariants worth asserting, and the scaffolding that lets a reviewer clone the repository and watch the ledger prove itself.
Publishing tools is distribution that compounds: what a public repository really earns a services business, what it never will, and how to run it deliberately.
A voice can be cloned from five seconds of audio. The hard part is consent: what a donator must agree to, what the AI Act makes you disclose, and which records answer the question later.
EU copyright protects works of human authorship, so a machine-only manuscript may have no owner at all. The practical question is what record proves your contribution, and what Article 50 asks you to disclose.
An autonomous security agent is an architecture problem before it is a model problem. Written scope, proof of authorisation, allowlists, evidence and human sign-off decide whether the work is lawful.
You cannot judge prose you watched being generated. Remove the identifying marks, write the failure rubric before reading, use two readers, and keep the verdict in a file.
Cloud TTS bills per character and sends your text away. Local models now clear the bar for narration, so the real choice is hardware, quality and convenience. We compare F5-TTS, GPT-SoVITS, XTTS, Piper and Kokoro, and test what Romanian actually needs.
"Write in my style" is an instruction a model cannot check. Measure voice from the author's own prose, then score drafts against the profile.
Every business that adds a chatbot or an agent now owns an input-validation problem it did not have. What prompt injection is, how it differs from a jailbreak, and which architectural fixes limit the damage.
Security reading lists rot faster than any other kind. We publish the method we use to judge, prune and date one, plus fifteen entries checked in October 2026.
You do not need a bug bounty to hear about a bug before the internet does. One page, one security.txt, a response window and a safe-harbour sentence cover most of it.
A shared browser profile puts every client session in one cookie jar, under extensions that can read them all. Here is what a stolen session grants, and what real isolation means.
A config file has callers, a contract and a failure mode. Validate it before the write, replace it atomically, control concurrent writers, keep history and explain every key.
A balance column drifts the moment two writers touch it. Store the entries, derive the balance, lock the row inside the transaction, and correct mistakes with reversals instead of edits.
Bank transfer, Romanian card processors, Stripe and SEPA for a small SRL, with the fees we could verify today and the reconciliation work that decides whether the month closes.
Long-form AI fiction does not fail because of a bad prompt. It fails when the context is compacted and the story's state disappears. Files, gates and a small script fix it.
Claude Max weekly limit, measured: our Max 5x account, capped every week, used 48% to 82% less after Opus 5.5 arrived. One account only. See the data, the method and what Anthropic has published.
AI news September 2026 in one page: new OpenAI, Anthropic and Google models, lower API prices, two big acquisitions and a safety scare after AI agents left test environments. Every item is sourced.
Chinese AI models vs US frontier labs on 4 October 2026: dated releases, independent and lab-reported benchmarks, token prices, licences and what to choose.
A table of AI coding subscriptions compared as of 4 October 2026: monthly price, models, usage limits as each vendor states them, supported tools and terms for third-party use.
Zero data retention means a provider stores neither your prompt nor the answer. Through a router the promise must hold at every hop. Here is what OpenRouter, OpenCode and Command Code document as of 4 October 2026.
The best LLM for each task as of 4 October 2026: writing, coding, research, long documents, translation, voice and local models, with leaderboard evidence, prices and trade-offs.
The EU AI Act small business rules in 2026: which dates moved after the omnibus, what tool users must do, when to label AI content, and the fines as written.
To use AI effectively, give it context and examples, ask for a format, refine the draft and check every fact. This guide covers prompts, limits, privacy and a checklist.
The best AI coding plan depends on your monthly calls. Six usage profiles priced at API list prices and matched to published plan allowances, with the arithmetic and when to switch.
A router gives you one key, one bill and many models. A direct API gives you one contract and fewer parties. See how OpenRouter, OpenCode, Command Code and a local gateway differ, checked 4 October 2026.
AI inference costs per token fall about 13× a year for a fixed capability, yet each task uses more tokens. Track cost per task, with a worked example, not price per token.
GLM Coding Plan starts at $18 a month. We compared it with Kimi Code, MiniMax M Plan, Alibaba's Coding Plan and DeepSeek on 4 October 2026, including Claude Code setup.
e-Factura Romania explained: who must send invoices through RO e-Factura, the 5 working-day deadline, how SPV works, common errors and the fines in the law. Checked on 4 October 2026.
OpenCode Go costs $10 a month, Command Code starts at $1 with a $10 GOAT plan, and Charm Hyper (called HyperCharm in some tools) sells credits. Prices, limits and caveats as of 4 October 2026.
The best open weight coding models on 4 October 2026, ranked with dated scores, licences, hardware needs and Claude Code setup for local and hosted use.
AI capex at four big tech firms is heading for about $720–745 billion in 2026. We lay out the revenue it chases, the bull and bear cases, and what it means for buyers.
AI hallucinations are confident answers that are false, because models predict likely words and are rewarded for guessing. Search, source texts and a verification routine lower the risk.
AI adoption in Romania: 5.21% of enterprises used AI in 2025, the lowest share in the EU. Eurostat data, first use cases, costs, GDPR, the AI Act and verified funding.
On personal plans, ChatGPT may train on your chats unless you opt out. Compare ChatGPT, Gemini, Claude, Copilot and others: training defaults, opt-outs, temporary chats and retention, checked 4 October 2026.
Cybersecurity for small business starts with ten habits: unique passwords, 2FA, payment checks, updates, tested 3-2-1 backups and an incident plan. Checklist and Romanian contacts included.
AI agents ROI comes down to time saved times loaded cost, minus tokens, review and maintenance. A formula, a worked example, a pilot method and when not to automate.
AI for students works as a tutor, not a ghostwriter. See what ASE and other Romanian universities allow, how to self-test, check sources, cite AI use and prompt it like a tutor.
Claude Code with other models works through ANTHROPIC_BASE_URL, either directly or via an LLM gateway. Here is the setup, tier routing, fallbacks and what you give up.
Website cost in Romania in 2026 runs from about €200 for a template to €10,000 for an online shop. See the ranges by site type and what drives the price.
Claude Code is free to install, but the model is what costs money. These are the real free routes, local models, free tiers and gateways, and what you give up.
Prompt caching lets an LLM reuse the part of your prompt that has not changed and bill it at a fraction of the input price. Here is how the cache, its 5-minute and 1-hour TTL, and keep-alive pings work.
Add Google Calendar booking on your own website with no booking SaaS: free slots from free/busy, double-booking checks, reminders and Meet links.
Claude subscription vs API comes down to how steady your usage is. See how each is billed, how the cache TTL differs, what drives Claude Code cost and how to track it.
A git-based CMS keeps content as files in Git, so your site needs no database. See how danubelabs.net works, what the limits are and who it suits.
Local Whisper transcription keeps your audio on your own computer. Convert to 16 kHz mono with FFmpeg, run faster-whisper on a GPU or CPU, and export TXT, SRT or VTT, with speaker labels from pyannote.
Claude Code keeps session history as JSONL files and deletes them after 30 days by default. Learn where they live, how to change cleanupPeriodDays, fix a broken resume and archive everything.
To translate a Word document with AI and keep the formatting, extract the text run by run, translate it in chunks with context, and write it back into the same runs. PDFs need extraction or OCR first.
Give every Vercel, Netlify, Cloudflare, Supabase or Gmail account its own isolated browser session and you can switch without logging out. Here are the options, the security trade-offs and the rules to check.
A step-by-step YOLOv8 custom training walkthrough with real results: 0.84 precision, 0.50 recall and 0.59 mAP@50. Learn what low recall means and why more varied data helped more than a bigger model.
Small teams do not lack automation ideas, they lack a way to rank them. A one-page scoring method we use in consulting engagements.
Embedding a client's live site in your portfolio is a great demo, until X-Frame-Options or a Content-Security-Policy quietly blanks the frame. A practical guide.
Most AI features are expensive because nobody measured them. Here is the routing, caching and prompt work that took one production workload from thousands to tens of dollars a month.
