Agents draft the parts that repeat: documentation, data models, scaffolding, first-pass review. A senior engineer reads and owns every line that reaches your repository. That split is where MVPs in about four weeks under $10K come from, and you get it in writing before you sign.

we run this way on purpose

Bookings, records and integrations with systems that predate the website: the plumbing under a healthcare site is a full software project. We plan, build and deliver the whole thing, interface to infrastructure.

the automation is the boring, valuable half

Specs before swarms, a human who can explain the diff, and an AI disclosure you read before signing rather than discover in a postmortem. Our delivery page says where agents work, where they don't, and who is accountable when they are wrong.

what this looks like on a real engagement

We work out what your site has to say, design it around that, and ship it. Wijntransport, the redesign this article ends on, is one of ours.

a website built around your customer

Web products, mobile apps, websites, and senior engineers placed inside teams that already have all three. 100+ products delivered, 11+ years in business.

want to talk about what you're building?

Positioning first, then design, then the build, from one team. No templates, and no handover gap where a finished design meets somebody else's implementation.

this process, pointed at your website

Luna's price moved 80% in three weeks. We prefer open-weight models where they do the job and publish our own frameworks as open source, so nothing we build for you is welded to one vendor's price sheet. Our delivery page spells out where models sit in a build.

portability is a design decision

Most of the failures above are set in motion before anyone writes code: fuzzy objectives, and the wrong team. Our first call is a diagnosis, not a pitch, and you leave with a plan whether you hire us or not.

the 15 minutes that prevent most of this

Serving, observability, a fallback path, a security boundary and someone who carries the pager: that is the system around a self-hosted model, and it is the same system we plan, build and deliver for the rest of a product. Architecture, backend, infrastructure, handover.

the model is the easy half

The five gates are the cheap part. The expensive part is finding out two quarters in that retrieval, not training, was the missing prerequisite. Our first call is 15 minutes and diagnostic: you leave with a concrete action plan whether you hire us or not.

want a second opinion on the go or no-go?

If the constraint is one senior drowning in everyone else's diffs, another agent will not help. We place senior developers inside existing teams, hired for hands-on production work with coding agents and for spotting agent-written code that reads well and fails under scrutiny. You keep the roadmap; a manager on our side stays accountable for their work.

a review queue is a staffing problem

how software actually gets built

Keep reading

What we learned building and shipping software that month.

One email a month.

Check your inbox. The next issue goes out at the start of the month.

You're in, {{name}}.

What we learned building and shipping software that month. No drip sequence, no sales pitch.

A senior engineer reads this, not a sales rep. You'll have an answer within one working day.

let's talk about your project

A few lines is enough.

tell us what you're building

0 - How to Fail an IT Project - Hero

Defining project failure | The 5 fatal mistakes to avoid | Staying focused on project objectives | Hiring the right development team

How to fail an IT project

UX Hero

Double diamond design thinking | Planning and discovery methods | Research methods | Design methods | Testing and validation methods

How we use the design thinking process

Hero-Image-Medical

Leveraging automation in the fight against COVID-19 | Why automate? | Technology and implementations | Keep up with trends | Novomed: a case study

Automating healthcare websites

Hero-Image

Organization-centric versus user-centric design | Web design components that don't match | Pain points | Building customer-centric websites requires customization | Strategies and functionality to increase user experience | Wijntransport case study

Increase sales with customer-centric design

Agentic Engineering in 2026 - Hero

The vibe coding hangover | Context engineering ate prompt engineering | RAG vs. fine-tuning, mostly resolved | Multi-agent systems: nobody agrees yet | Code got cheap. Judgment didn't. | The hype, the KPIs, and the paycheck | The job title arrived before the job description

Agentic Engineering in 2026

As of 14 August 2026, twelve days after the EU AI Act transparency rules applied, the watermarks-remover repository has 6,017 GitHub stars and c2patxt, an implementation of the marking the law asks for, has three.

What the industry built for Article 50

Nobody was watching the right deadline | Which half of Article 50 applies to you | Anthropic shipped half of it on the day | Your source code was never in scope | The mark comes off in one paraphrase | Six thousand stars for the eraser, three for the marker | What the law already knew

The Deadline That Didn't Move

Across 1,267 solutions, 13 models picked Ruby zero times, while functionally equivalent libraries differ by up to 84% in the quality of the code models write with them. Popularity predicts that gap at 0.574 concordance.

What the models actually reach for

The essay came back | Rails is as boring as it gets | Popularity barely beats a coin | The model's taste turns over faster than yours | Documentation is the only lever anyone measured | What the token actually bought | Boring may become irrelevant | Nobody ran the experiment

Boring Is Not In-Distribution

A verified DeepSeek invoice of $8.79 for 1.28 billion input tokens at a 98.582% cache hit rate costs $19.21 off-peak and $38.42 at peak under the table taking effect at 16:00 UTC on August 16, 2026.

The same 1.28 billion tokens, billed three ways

Three days' notice | The cache line took the worst of it | Peak hours are somebody else's business day | Buy a GPU, they said | Where the break-even actually sits | The line item you can't rent | The strongest case for doing nothing | What to do before Sunday

The Break-Even Moved

Across 247,694 instruction lifetimes in 1,867 repositories, agentic prompt files grew 226% over their lifetime, while 58% of agent config files carry a single commit; recording why a rule exists removed 99.3% of the excess.

The spec file nobody edits and nobody deletes

The file that only grows | It just has to be slightly behind the code | Nothing in the repo notices | A stale spec is not ignored, it is obeyed | Nobody ships a conflict checker | The steelman wins on the scoreboard | What measured better was compiling it | English is the new code and has no comments

Markdown Has No Compiler

Anthropic ran sixteen parallel Claude instances across nearly 2,000 sessions for just under $20,000, and the resulting C compiler emits slower code than GCC with optimizations disabled. February 2026.

What sixteen parallel Claudes bought

The day the fleet shipped | Twenty thousand dollars for a worse compiler | Fifty-four percent of the commits were paperwork | The benchmark calls it bloat | The failure mode is exponential | Anthropic quietly voted against its own feature | Where fleets actually win | You are the bottleneck, and you always were

Fleets Need a Grader

The claim that 9 in 10 startups fail traced to no primary source: the citation chain dead-ends in an uncited 2011 PDF, while BLS data through March 2025 shows 51.4% of new establishments still alive at five years.

The most quoted statistic in tech

Not even Paul Graham knows | The paper trail ends in a PDF | One article, four failure rates | Nine in ten of what, exactly | Failure stopped being a binary | Watch a zombie statistic being born | Why the number refuses to die

The Number Nobody Checked

Claude Code sends roughly 33,000 tokens before reading your prompt; OpenCode sends about 6,900 pointed at the same Anthropic model. Systima, July 2026.

Tokens sent before your first word

The consultant who bills for reading your email | Same model, same tests, four times the bill | Where the tokens actually go | The cache is the business model | Fan-out is a force multiplier on waste | The steelman: the harness also sets the score | The bill has reached the CFO | Audit the wrapper

The Harness Tax

A real DeepSeek invoice of $6.36 for 1,015,102,000 tokens at a 97.5% cache-hit rate, against $142.11 for the same tokens had every one been a cache miss.

The same billion tokens, billed two ways

The receipt and the warning | One cut, one retraction, one warning | The fuse got swapped for a bigger one | You do not have a token bill, you have a cache-hit rate | A chip story wearing a pricing costume | The strongest case that nothing is ending | Flash has an exit, Pro does not | Plan your usage accordingly

A Billion Tokens for $6.36

Google reported fixing 1,072 security bugs across Chrome 149 and 150 in the same week JFrog found 54 of 55 published advisories were fabricated.

Same technology, same week, opposite results

A 9.8 for code that never existed | The machines actually got good | Five findings became one | The harness matters more than the model | What it looks like when someone pays | Four organizations and nobody ran the code | Fixing is the part nobody automated | What the bliss taught everyone

Somebody Has to Run the Exploit

A causal study of Germany’s Sovereign Tech Fund measured a 143.8% rise in commits with no detectable increase in the number of contributors.

Public funding bought commits, not contributors

The sign only robots can read | Everybody patched it differently | Except the machines got good | The ladder went missing | You cannot mint a maintainer | Meanwhile, open got captured | The market started pricing signatures | Remember which half got expensive

Your Code Got Cheap. Your Signature Didn’t.

On Quesma’s eight-puzzle Baba Is Bench suite, DeepSeek V4-Flash-0731 cost $1.16 against $42.21 for Claude Sonnet 5, both taking 22 turns.

Same score, one thirty-sixth the bill

The receipts | The body didn't change | The number that reproduced | Ask for high, get low | Where the amazement outran the evidence | It runs on hardware you can buy | What the celebration skipped | Remember the date

Same Score, 1/36th the Bill

Capable agent configurations lost 30 points of assertion pass rate from baseline to fully specified tasks, while 31% more pull requests merged without any review.

More specification, worse output

The spec that couldn't build itself | Two bets, opposite directions | More spec, worse output | The steelman, and it is a good one | Nobody reads the 200 lines | Nobody has the receipt | What survives | We build one of these

Nobody Reads the Spec

DeepSeek held V4-Flash at $0.28 per million output tokens while its self-reported Terminal Bench 2.1 score rose from 56.9 to 82.7 on unchanged weights.

DeepSeek V4-Flash-0731: same price, higher scores

Nobody shipped a model today | The cache line is the whole price war | Every one of those numbers is self-reported | Cheap tokens, expensive tasks | What the bill actually looks like | The 35-minute panic | The fine print has a fuse in it | What to do before the surcharge lands

DeepSeek Shipped a Post-Train, Not a Model

OpenAI cut GPT-5.6 Luna 80 percent while OpenAI, Google, and Anthropic’s combined OpenRouter token share fell from 70 to 30 percent in a year.

The Luna price cut versus the market-share collapse

A price cut three weeks after launch is a confession | Nobody cuts 80% out of kindness | The sticker price is the least honest number in AI | Who should switch this week | Who should stay put | The trust discount | Anthropic is fighting a different war | The playbook

The Day AI Pricing Fell Off a Cliff

The parallel-agent ceiling is three; practitioners report being wiped out for the day by 11am.

Agent fatigue: the ceiling and the clock

The tiredness got a name | Three agents, then the cliff | Manager schedule, IC paycheck | Review was never the fun part | The machine never idles, so neither do you | Budget attention like you budget tokens

Your Agents Aren’t Tired. You Are.

LinearB 2026 benchmarks: teams merge 98 percent more pull requests while review wait grows 4.6 times.

Coding agents in production: throughput up, review crunched

The control group defected | Review is where the speedup goes to die | Specs beat swarms | "The simplest fix" is a tell | The invoice got a CFO | MCP grew up and got a boring job | Blast radius, measured in seconds | An amplifier, not an equalizer

Coding Agents Hit Production. Production Hit Back.

A small compute core surrounded by data, verification, serving, and monitoring systems.

The operational system surrounding a fine-tuning run

The GPU bill leaves out most of the project | The scorer becomes product policy | Put every recurring cost in the model | Require five gates before training | Shadow traffic decides whether the pilot ships

The Real Cost of a $500 Fine-Tune

Kimi K3 accelerator clusters compared by hardware, provider speed, and monthly cost.

What it costs to self-host Kimi K3

The short answer starts at $40,000 a month | K3 fits on 8 B300s or 16 B200s | Provider speed currently varies by about 5× | Renting costs $70,000 to $467,000 a month | Buying starts at $575,245 | API break-even starts at 17.4 billion tokens | Self-host when control or steady load pays for it

What It Actually Costs to Self-Host Kimi K3

about banner

We are honest, careful, and dedicated with a proactive attitude and a lot of soul!

"A small team of passionate professionals will always outperform a giant team of second-rate developers."

Through this selection of our work we hope to inspire in you the confidence, trust, and enthusiasm our clients feel when starting a new project with us.

best practices and trends in the web development industry

be part of our team and enjoy life!

careers_hero_2

contact_image_2

we build digital products that drive real business results.

industries we excel at

We are defined by the work that we do, folded on the customer’s needs, doing our best to highlight and bring to life our client’s vision.

projects we are most proud of

We consult, design, develop & scale robust web & mobile solutions that fuel innovation & deliver digital success!

Theodor Diaconu

Published by Theodor Diaconu