Blog
24 posts
- ~10 min read
Pencil instead of ink, and our own faces: how we reworked the AI Budżet ad films
Part two: we audited our film pipeline against someone else's animate-with-code method, added a pitfalls file and story rules, built a pencil kit with handwriting written stroke by stroke, tried HyperFrames on Windows, and recast the debts film with caricatures of real people.
Read article → - ~17 min read
Ads with zero stock footage: how we render promo videos for AI Budżet in code (and with Claude Code agents)
Every pixel in the AI Budżet ads is a real screenshot or drawn by code. We walk through three pipelines: Python screenshot reels, a feature-tour reel, and 30-second Canvas 2D films with a synthesized score built by Claude Code agents — with a determinism gate, 6 ms of compressor compensation and the ffmpeg settings for Instagram and TikTok.
Read article → - ~17 min read
Karpathy’s LLM wiki: what it is, how we run it, and what still needs fixing
Andrej Karpathy described a pattern where the model builds and maintains a living knowledge base instead of rediscovering it on every question. We ran it in a production monorepo: 24,616 fewer words in the file loaded into every session, three rituals instead of one, and an honest list of what still does not work.
Read article → - ~9 min read
Jev: where in our LangGraph flows a model that never writes a sentence would fit
TypeSafe AI shipped Jev — a model that returns a typed, calibrated-probability decision instead of a sentence, in 70–500 ms. We walked our own five LangGraph flows node by node and gave each decision point a verdict.
Read article → - ~14 min read
Why an LLM lies confidently on correct numbers: metric semantics in AI analytics
Technically correct but semantically wrong metrics make a language model produce coherent, confident, wrong recommendations — no crash, no trace in the output. Six real cases from a cross-channel digest and the engineering that catches them.
Read article → - ~4 min read
MiCode: what we build, what we sell, and how to start
Six products of our own, three service lines, and a quote in three business days — a description of the company in which every figure can be checked by the reader.
Read article → - ~15 min read
Building an AI-citation monitor inside Gemini's free-tier quota
We ask Gemini every day whether it cites mi-code.pl — but the free tier allows only 20 model calls a day, and a full sweep needs 54. How we spread the sweep across three days, what retries handle dropped connections and rate limits, and why a fingerprint of the prompt list keeps a half-finished sweep from ever being compared to a whole one.
Read article → - ~13 min read
KSeF in August 2026: what to do if your company still is not ready
The KSeF obligation has been running since February and April 2026, and the penalty-free period ends on 31 December. What to do if you still invoice outside the system: the calendar, four typical situations, the KSeF API integration from the technical side, and a checklist for the coming weeks.
Read article → - ~17 min read
AI Funding for Polish Companies in 2026: What Is Open, and When the R&D Relief Wins
The state of play on 16 August 2026: what became of KPO, which FENG calls are open and which are still ahead, and why on a PLN 30,000–80,000 deployment the R&D relief and IP Box usually beat chasing a grant. With a programme table, a decision tree and an honest tally of what an application really costs.
Read article → - ~6 min read
Is there an AI assistant connected to wFirma? Yes — it's called Accounting AI Agent
Asked in Polish, ChatGPT answered that no ready AI assistant connected to a wFirma account exists. It does — it is called Accounting AI Agent. What it can answer, what data it reads, what it will not do, and what it costs to try.
Read article → - ~9 min read
The AI Act in August 2026: what was postponed, and what applies today
The 2 August 2026 deadline for high-risk systems really did move — to December 2027. But Article 50 transparency applies from 2 August, so do fines for GPAI providers, and Poland’s AI systems act enters into force on 11 August. Here is what applies today, what is still ahead, and what to do this quarter.
Read article → - ~11 min read
What an AI agent really costs per month: a model you can check yourself
A price per million tokens does not answer "how much per month". We break down the formula behind an agent’s bill, four places the budget leaks and four levers that move it, run a reference configuration — $188.10 against $74.38 with caching — and hand you a calculator for your own numbers. A model on declared assumptions, not our invoices.
Read article → - ~7 min read
GEO and AEO: staying visible in the age of AI search
Search is moving from ten blue links to a single answer written by AI. We explain what AEO and GEO are, how they emerged, and how we apply them across MiCode so our content gets cited by ChatGPT, Perplexity, and Google AI Overviews.
Read article → - ~16 min read
Sleep on it: AI agent teams that consolidate overnight
Most AI agents never get better — they answer, forget, and plateau. A different approach gives a team of specialist agents a scheduled night shift: they re-read their own instructions, study the codebase, surface debt and ideas, and propose fixes to themselves for a human to approve. Where the idea comes from, how the self-study loop works, and why the gains compound.
Read article → - ~8 min read
Inside eMarketing AI: a supervisor and a team of specialist AI agents
A look under the hood of eMarketing AI: a supervisor routes work to specialist agents (content, checklists, documents, SEO, strategy, email, analytics), runs go through a queue, answers are grounded in the project’s real data, and it all runs on a single GPT-4o.
Read article → - ~9 min read
Inside Accounting AI Agent: one agent, 80+ tools, and real wFirma data
A look under the hood of Accounting AI Agent: a single LangGraph agent, up to 82 tools, grounding in wFirma and Polish registries, KSeF, receipt OCR, and rule-based long-term memory.
Read article → - ~9 min read
How AI Budget Assistant uses AI: a function-calling agent, vision, voice, and data protection
A look under the hood of AI in AI Budget Assistant: one server-side key and three model «speeds», a function-calling chat agent, data isolated from instructions, vision receipt OCR, voice, bots — and where we deliberately skipped AI.
Read article → - ~12 min read
How Legalka KB uses AI: RAG, voice, and zero hallucinations
A walkthrough of the relocation bot’s entire AI stack: embeddings and retrieval, query rewriting, grounded answers with abstention, voice in/out, a practice-ingest pipeline, and an eval judge.
Read article → - ~1 min read
Automating Polish accounting compliance with an AI agent
How Accounting AI Agent connects to wFirma to handle VAT, PIT, and CIT questions and automate invoice processing.
Read article → - ~1 min read
Inside eMarketing AI: one platform for AI-generated content, campaigns, and email
Why we built a single SaaS platform for AI content generation, campaign management, and email marketing — and how the AI chat assistant ties it together.
Read article → - ~1 min read
Receipts, voice, or chat: how AI Budget Assistant tracks spending without the spreadsheet
Why we combined receipt OCR, voice input, and a GPT-4 chat in one mobile budgeting app — and how gamification keeps people using it.
Read article → - ~2 min read
Letting AI agents own the boring parts of QA
How Testing AI Assistant generates test cases from your codebase, runs them with intelligent scheduling, and flags flaky tests automatically.
Read article → - ~1 min read
How AI agents optimised supply chain management
A case study of the SCM system with AI — 300+ REST endpoints, Text2SQL, and RAG integration in practice.
Read article → - ~1 min read
How to add AI chat to Angular in 15 minutes
Step by step with ngx-open-web-ui-chat — npm install, Open WebUI config, and first messages.
Read article →