A Polish MFA analyst preparing a position paper has hours, not days. Source material runs to thousands of pages — government statements, think-tank reports, GDELT-indexed news in 65 languages, market data. Standard tools either summarise without showing their work or hand back something the analyst can't defend in front of a director. 'AI said so' is not an answer to 'where did this conclusion come from?' The MSZ hackathon brief made the constraints explicit: explainability and resistance to data poisoning, both non-negotiable.

Five-phase pipeline — discovery → scraping → extraction → aggregation → synthesis — with the LLM touched once, at the end, against an ~8K-character payload of pre-weighted facts. Discovery layers RSS over a curated 85-source whitelist on top of GDELT keyword queries across 113 domains. Extraction is local: spaCy + regex + FAISS over multilingual-e5-large embeddings. Hierarchical bottom-up summarisation keeps the LLM payload constant whether the corpus is fifty million words or five billion. ~$0.10 per analysis, regardless of input size. Every scenario is a chain of FAKT → MECHANIZM → SKUTEK → WNIOSEK steps with numbered citations back to source. The chain exists in the data, not in the model — explainability is a structure, not a UI feature. Anti-poisoning runs four layers: triangulation (≥3 independent sources from ≥2 categories), anomaly detection, consistency checks, dynamic source-trust scoring with state-controlled domains penalised. What this gives an analyst: any conclusion clicks back to the documents it came from.

The hackathon weekend: twenty-four hours straight, except a 1:30 a.m. run around Bydgoszcz with Dawid and a 7:30 a.m. break in a massage chair. Final pitch in front of an MSZ panel, fifty-five hours into the streak; some of what I said made sense, some of it didn't. Won. Deputy PM Radosław Sikorski handed the award. Five-hour drive home, sixty-seven hours awake. Five hours of sleep, then 8 a.m. EduChat lessons in Wodzisław Śląski with fifth-graders — forming complete sentences took noticeable effort; the kids were patient. Three months later, the protocol the ministry reserves for foreign ministers and heads of state: a private cabinet next to Sikorski's office, three high-ranking officials, a continuation conversation. v2 is what's being built now.

EU treaty law promises a single market for capital, but in practice savings rates remain trapped behind national banking apps and national-language interfaces. A Polish saver who notices that a Spanish or Luxembourgish bank pays meaningfully more on a fixed deposit hits a wall: the bank's signup is in Spanish or French, comparison sites (Raisin, deposit-solutions, the national portals) terminate at a referral link, and any portfolio view has to be reconstructed by hand in a spreadsheet across multiple currencies and inflation regimes. Nobody is doing the planning layer — the part where a user models a portfolio across countries, with inflation and FX baked in, before committing to opening anything.

Two product calls separate Blo from everything else in the space. First: simulate first, commit later. The MVP doesn't connect to any real bank. Users hand-enter accounts as 'Planned,' 'Active,' or 'Matured.' Removing the integration layer collapses the regulatory surface and lets the actual product (cross-border comparison + portfolio math + maturity calendar) ship in days rather than months. Second: real-return-first analytics. Every nominal rate is paired with the country's inflation, so a 5.75% PLN deposit and a 2.98% EUR deposit are compared on the axis that matters. Stack: Flutter 3.10, Provider for state, sqflite for local SQLite, ECB Data Portal API for live key rates, six ChangeNotifier providers above a DAO/SQLite layer. Local-first, no backend. The compound-interest projection runs on-device in `UserAccount` so portfolio totals, projected values, and per-account interest derive from one source of truth.

Day 1 of Blo was the same day the Harvard admissions decision arrived (rejection). The concept came together at Econverse V4 in Prague, where Marek Ditz signed on as mentor and stayed. We won the Prague stage, then the Lodz finals; the first version on the App Store was Swift, written by Dawid; I rewrote the whole stack in Flutter — sixteen thousand lines across two days, November 24–25, 2025. The planning layer is shipped. The integration layer waits on EUDI Wallet (mandated 2026), PSD3 (2026–2027), and FIDA (2027) — the regulatory window that collapses the cost of real cross-border bank integration. The bet is the calendar, not the pivot.

Polish managers facing daily problems — a difficult conversation about lateness, a suspected mobbing case, a budget cut to communicate, a hybrid team losing cohesion — have three options today, all bad: generic AI that hallucinates Kodeks Pracy article numbers; HR consultants at 800–2,000 PLN per session, days to schedule; or 90 minutes of PDF-from-Google methodology hunting. None of them produce a printable script with structured output. The harder problem under the surface: a manager who is overwhelmed doesn't know they need the Eisenhower Matrix; generic chat answers the question that was asked, not the one that should have been.

Single Flask app. Each of 145 tools is a Python class with a server-side ui_schema across 35 reusable components rendered identically on web and Flutter. The agentic loop streams Claude Sonnet 4.5, intercepts tool_use events, executes server-side, sends tool_result back, re-streams in a single user turn. Tool routing at scale: hybrid retrieval — cosine + BM25 over the tool descriptions themselves, with negative-signal blocks ('DO NOT USE WHEN…') stripped before vectorisation, since cosine has no concept of negation. Then a Claude Haiku re-rank top-10 → top-3, dynamic injection into the conversation's tool set with LRU eviction at 15. Token cost per turn dropped from ~25k (naive 'send all tools') to 700–2,800 — a 10–22× reduction at ~50–150ms added latency. Module 17, 'Prawne ABC menedżera', is the moat: generic Claude hallucinates KP article numbers; ManagAI doesn't.

Cezary had twenty-five years of management consulting and a thesis. I had the stack and the conviction the thesis was buildable. Three months and ten days from first commit to 145 tools. Single contributor in the git log; mobile (Flutter) and web shipped in parallel. The legal-tone calibration is the model learning by contrast — the system prompt holds both wrong versions, labeled, alongside the correct one. The wrong versions cost two iterations to find: a STOP-emergency mode managers rejected as alarming, and a therapy-speak version they rejected as performance. Reading actual transcripts out loud is what surfaced both.

Polish SMBs need software they actually own — bespoke ERPs, mobile apps, AI integrations — but SAP and Comarch are expensive and wrong-sized for the specific operational realities of, say, a paint shop or a service business. The Polish enterprise-IT default is corporate consultancy with three layers between the buyer and whoever is writing code.

Two of the most useful client builds — an automotive-industry ERP with bidirectional WooCommerce sync (orders, products, customers, stock — replacing manual Excel transcription with a mobile app for production-floor workers) and an AI voice assistant on FreePBX/Asterisk with sub-2-second streaming response and per-caller memory keyed to phone number, capable of inbound and outbound calls. EduChat is one of Assistron's products; covered above.

Origin: Econverse 2023 in Wrocław. Five-person student team, project called ZwrotMatch, won the eliminations, lost the finals. The team of five didn't survive the competition; Kuba and I kept the working relationship. Voicebots were one of the unbuilt ideas from the brainstorm; we started building them after. That's the seed. The rest: a software house run from Wodzisław Śląski, ~50k population, a region better known for coal mining than for AI agencies. VAT-invoice, code-warranty contracts mean carrying every piece of the compliance overhead a Polish sp. z o.o. requires.

Polish school principals are required by the Karta Nauczyciela and a MEN regulation to produce a written evaluation for every teacher. The grading itself is fast — a principal already knows their staff. The slow part is the paperwork: twelve numbered criteria, each needing formal multi-sentence justification, the right grade band selected (`wyróżniająca` / `bardzo dobra` / `dobra` / `negatywna`), supporting opinions attached. Twenty-plus hours per cycle, per principal — a week behind a desk. On top of that, MEN published a new regulation in 2025 that took effect on November 27th — the structure changed from eleven criteria to twelve, the points scale changed, existing templates and habits don't carry over.

Two product calls the existing tooling didn't make. First: don't ask the LLM to evaluate the teacher. Ask it to write the paperwork. The principal supplies observations and a target grade; the model's job is formal Polish prose against the criterion structure, not judgment. The system prompt tells the model that most teachers are competent and to default to 'bardzo dobra' (75–85%) unless the principal volunteers a problem. Second: keep teacher names out of the database entirely. Names live in browser localStorage keyed by UUID; the server stores only the UUID, the score JSON, and the documentation text. The LLM call sees observations, never an identified person — a concrete RODO answer instead of a privacy-policy paragraph. Stack: Flask 3, MySQL in prod / SQLite in dev, OpenAI GPT-4o with Anthropic fallback, WeasyPrint for PDF, Alpine.js on the conversation page.

A complete product, shipped by one engineer in three weeks against a non-negotiable November 27, 2025 regulatory deadline that the existing templates didn't carry across. Distributed free to several principals through their first cycle.