Talk to an engineerIntro call
03 - Build & Modernize / Custom Software Development with AI

Custom Software Development with AI

End-to-end product engineering with AI at the centre of the process - production applications, agentic backends and AI-augmented internal tools built on a staged pipeline where every change passes automated gates and two independent code reviews before production, and an engineer reads risky changes line by line.

Talk to an engineerSee the package
Fixed-price entry packageProduction PoC SprintEUR 8,0001 use case · up to 4 weeks from scopingWhat is included
In short
01

Senior judgment at AI speed

Production software built with AI at the centre of the process, with an engineer accountable for what runs in production.

02

Every change through the same gates

Ticket, spec, plan, automated checks, two independent code reviews, PR with preview, verification in production.

03

Handed over, not held hostage

Code, pipeline, infrastructure and runbooks go to your team with pair sessions and a recorded walkthrough.

23Minds builds production software with AI at the centre of the engineering process.

Show the full summary

The practice delivers web applications, agentic backends, AI-augmented internal tools and AI-native products to teams that need senior engineering judgment at the velocity AI now enables. Every change moves through the same pipeline: ticket, spec, plan, implementation, automated gates (typecheck, lint, tests, build, a dependency and secrets check, two independent code reviews), PR with preview, line-by-line human review for risky changes, merge, verification in production. The pipeline accelerates delivery and holds stability, and the team measures both: throughput and defects found after merge. ISO 9001:2015 quality assurance backing.

Why 23Minds

What makes this service defensible.

  • This site is the proof

    This website is built and maintained in the same pipeline, by AI under engineering direction; the About page shows how.

  • Throughput and defects both measured

    The pipeline reports delivery speed next to defects found after merge, so speed never hides quality.

  • PoC in four weeks at a fixed price

    One use case on your data, deployed to staging with code, tests and a benchmark report.

  • ISO 9001:2015 quality system

    Certified by DNV, no. 10000439701-MSC-RvA-POL; delivery runs as a documented procedure.

How we work

The process, phase by phase.

  1. 1

    1. Intro call + scope

    Intro call, problem definition, technical scoping, fixed-fee phase proposal.

    Week 0-1
  2. 2

    2. Ticket, spec, plan

    Every change starts as a ticket, becomes a written spec, then a plan. Gate: you approve scope and plan before a line of code is written.

    Per change
  3. 3

    3. Implementation

    AI agents write the code and its unit tests under a senior engineer who sets the pattern, directs, corrects and rejects. Gate: every change in behaviour has a test that checks it, and without one the change does not move on.

    Per change
  4. 4

    4. Automated gates

    Typecheck, lint, tests, build and a dependency and secrets check run on every branch, then models from two different families review the diff independently. Gate: every critical or high finding is fixed before the change moves on, with the same threshold for AI-written and hand-written code.

    Every commit
  5. 5

    5. PR, preview + risk review

    Each PR gets a preview environment with the change running end to end, and both reviews are recorded in the PR. Gate: changes to migrations and data, authentication and permissions, payments or integrations with your systems are read line by line by an engineer before they ship.

    Per change
  6. 6

    6. Merge + production verification

    After deploy we run smoke checks on the live address and read the logs from the first minutes. Gate: if the check finds a problem, we first restore service using the rollback plan agreed before the deploy, then fix; the change is done when it works in production and you have accepted it there.

    Per release
  7. 7

    7. Hardening, launch + handover

    Production readiness review, load testing, observability completeness, on-call training, then pair sessions and documented handover. Optional maintenance under AI Application Maintenance.

    Last 2-4 weeks
How to start

Pick the way in that fits your budget and timing.

Every practice has a free first step, an entry package with a written deliverable and its price listed here, and a full project or retainer quoted after a first call. Once the scope is agreed, the price is fixed, with no hidden costs.

  1. 1
    Step 1 · Free

    Intro call or self-assessment

    60 minutes · 1 engineer · no obligation

    Free
    Show details
    • Intro call with an engineer, or the online self-assessment
    • A written next step within one business day
    • No obligation and no sales deck
    Talk to an engineer
  2. Recommended
    2
    Step 2 · Fixed price

    Production PoC Sprint

    1 use case · up to 4 weeks from scoping

    EUR 8,000 net

    Fixed price, no hidden costs; extras only from the Not included list.

    Show details
    • One defined use case built as a working PoC on your data
    • Deployed to staging
    • Source code and tests handed over
    • Benchmark report against the agreed metric
    • Not included: Production deployment and maintenance, integration with more than two source systems, infrastructure and inference costs, UI design work beyond a working sketch.
    Ask for this package
  3. 3
    Step 3 · Project or retainer

    Full scope, quoted after a first call

    Full scope · phased delivery · quoted per phase

    from EUR 28,000 net, quoted after a first call
    Show details
    • End-to-end product delivery to production
    • Infrastructure and observability included
    • Handover to your team
    Start the conversation
Sounds familiar?

The situations that bring teams to this service.

  • AI prototypes built by the team do not survive the move to production
  • A prompt-built prototype stalls at the demo: no tests, no review trail, no runbook, no owner in production - the delivered system has all four
  • In-house engineering is full; the roadmap keeps slipping
Show the other problems (2)
  • Agentic workflows demo well but the team has no production pattern for them
  • Custom development quotes from agencies do not account for AI-driven velocity
Who it's for
  • Product engineering teams shipping AI-native or AI-augmented products
  • Founders prototyping a new AI-first product who need senior engineering
  • Mid-stage companies replacing or extending a critical internal tool
  • Scale-ups whose product roadmap exceeds in-house engineering capacity
Failure modes

Where this goes wrong, and what we do about it.

  1. 1. The prototype reaches production, the review trail does notWhat stays with you: a repository where every merged change links back to its ticket, its spec and its recorded reviews
    What happens

    A prompt-built prototype gets to a demo in days. The code came from an assistant, nobody wrote a spec, there are no tests, and the first production incident lands on an engineer who has never read it. The question comes back on engineering forums in the same shape every month: the new project is almost entirely generated, and nobody knows how to take ownership of it.

    What we do about it

    Every change starts as a ticket and a written spec before any code exists. What the agent wrote passes the automated gates and two independent reviews by models from different families, with the same threshold for generated and hand-written code, and changes to migrations and data, authentication and permissions, payments or your integrations are also read by an engineer. The review trail lives in the repository, not in a chat window.

  2. 2. The agent adds branches to what it can seeWhat stays with you: duplication and reuse measurements at the start and after each phase, plus a repository context file the team keeps maintaining
    What happens

    An agent knows only the files in its context. When it meets an existing structure it extends it with one more branch instead of tidying it, and every extension makes the next one more likely. GitClear measured this pattern across 623 million changed lines: the share of refactoring falls while copy-paste and duplicated blocks rise. No single change looks wrong on the day it merges.

    What we do about it

    We keep the repository legible for agents and people alike: context files, a directory layout you can guess, tests that run with one command, module boundaries visible without reading everything. Duplication and cross-file reuse are measured at the start and after every phase, and refactoring is its own ticket with its own budget, not something expected to happen on the side.

  3. 3. An instruction to the agent is a suggestionWhat stays with you: a list of rules split into tool-enforced and prompt-enforced, with the enforcing mechanism named for each
    What happens

    A rule written in a prompt or a context file holds for a short task and weakens over a long chain of steps. When a validation fails, the agent is as likely to fix the validation as the code, because from where it stands both close the task. Practitioners who run agents daily reach the same conclusion: enforcement has to sit where the agent structurally cannot reach it.

    What we do about it

    Rules that must hold live outside the agent's reach: CI gates that run in a separate environment, branch protection, review by a person. We test that by trying to get around each rule the way an agent would, and we record next to every rule what enforces it.

What you get

Deliverables, not deliverable-shaped slides.

  • Working production application (web, agentic backend, internal tool)
  • Source code with documentation, tests, deployment configs, runbooks
  • Delivery pipeline handed over with the code: CI gates, PR template, review checklist, preview environments
  • Infrastructure-as-code for production, staging, preview environments
  • Observability and on-call setup (tracing, alerts, runbooks)
  • Knowledge transfer to your team - pair sessions, walkthroughs, recorded handover
FAQ

Questions procurement teams ask.

From the intro call to the first sprint typically 2-3 weeks. Faster for urgent engagements (1 week) when scope is already clear and we have capacity.
That is the trade we refuse, and the reason the pipeline exists. We track delivery throughput and defects found after merge side by side, report both, and treat rising post-merge defects as a reason to tighten gates, not as the price of speed. Industry data shows the trade-off: DORA's 2024 report estimated that a 25% increase in AI adoption goes with a 7.2% drop in delivery stability, and the 2025 report, based on nearly 5,000 respondents, still found stability falling while throughput rose. Results on speed differ: developers using GitHub Copilot completed 26.08% more tasks in three field experiments with 4,867 developers (Cui et al., Management Science, 2026), while experienced open-source developers in METR's 2025 trial took 19% longer with AI tools.
Fixed fee per phase. The first phase is a PoC sprint when the scope has to be tested against your data; its package price is in the price list on this page. Where the scope is already clear, we start from the first build phase at a fixed price. Later phases are priced one at a time, on what the previous one taught us.
You own the code. We retain rights to reusable internal libraries and patterns we wrote before the engagement. Standard mutual NDA on the engagement itself.
Yes. Engagement teams are stable from kickoff to handover. We do not swap engineers mid-engagement unless you ask us to.
Yes. Agentic backends (single-shot, multi-step, multi-agent), RAG systems, LLM-backed APIs, AI evaluation pipelines, embedding and vector search infrastructure. Systems with AI inside are handed over with an evaluation of answer quality as well.
Related services

Often shipped together with Custom Software Development with AI.

Build & Modernize

Legacy Modernization

AI-assisted refactoring and re-platforming for systems that block roadmap progress.

Learn more
Operate & Measure

AI Application Maintenance

Ongoing maintenance for AI-touched systems: provider migrations, prompt tuning, dependency upgrades.

Learn more
Sources

Where the dates and numbers come from.

Talk to an engineer.

Describe where you are with AI-assisted custom software development. A reply within one business day.

Talk to an engineer
Szczecin - ul. Wawrzyniaka 6WWarszawaZielona GóraKraków