Jyri GromykoLintunen

Six years of work on an ERP for a contractor doing building-facade renovation and apartment reconstruction. The system covers the whole chain: project setup, visual facade and wall editing on an interactive canvas, element tracking and history, configurable status workflows, ticketing, materials supply chain, personnel and equipment, and finance. Three separate applications share one REST API.

• A canvas editor for facades and walls (Konva / react-konva): panels, seams, holes, insulation tapes, mouldings, support frames and vertical profiling — each one a per-project preset that can be tuned without breaking existing projects.
• Export straight to the factory floor: a 99-column machining schema for the CNC line (Cut Rite / Schmid B300) — drilling coordinates X1–X10 and Y1–Y10, pockets, frames, diagonal cuts, ventilation holes, edge processing, material codes (HPL, fibre cement, aluminium composite) and per-pallet destacking order. A facade drawn in the editor becomes a program the machine can run, with nobody in between.
• A configurable state machine: status groups, status logic and status types, plus a change history for every entity (element, apartment, room, ticket, company). The customer defines their own workflows without a code change.
• Materials and inventory accounting: nomenclature with categories, stock levels, movements, material requests with line items, suppliers with price history, orders.
• Ticketing for crews and machinery: ticket types and groups, teams and equipment, phase-scoped statuses, dependencies between tickets and a full status history.
• Permissions on two levels: spatie/laravel-permission on the back end and CASL on the front end, with per-project roles — the same user can be a foreman on one project and a viewer on another.
• Excel import and export (maatwebsite/excel, xlsx), in-browser PDF generation (jspdf + html2canvas), multi-language support (i18next) and real-time notifications.
• A reproducible dev environment: Docker Compose (MySQL 8, Redis 7, Nginx), a one-command launcher and a `php artisan app:selftest` health check, with architectural decisions recorded as ADRs.

A GDPR-compliant sales-outreach platform that combines an email log, sequence-based campaign automation, lead enrichment and a React CRM. Two back ends live side by side: the Flask production app and a Django rewrite, sharing one PostgreSQL schema and matching each other endpoint for endpoint.

• A campaign engine: multi-step sequences, message variants with A/B testing (multi-armed bandit), condition-based branches, automatic actions and recipient enrollments.
• Orchestration of 17 lead sources under one scheduler: Finder.fi, PRH/YTJ, Jooble, LinkedIn Jobs and Messages, People Data Labs, Hunter, ZeroBounce, QuickEmail, Apify, Apilayer, Abstract, tech-stack detection and PageSpeed scoring.
• PageSpeed and ad-gap scoring: Google PageSpeed Insights, SSL detection, Google Ads Transparency and the Meta Ads Library are combined into a 0–100 score and a hot / warm / cold priority that refreshes automatically when a company's website changes.
• An enrichment chain where every contact passes through paid and free sources (PDL → Hunter → ZeroBounce → Gemini) until it has a verified, deliverable email address and a reasoned company context.
• A sales pipeline: deals, stages and stage history, tasks and task templates, an activity feed, saved filters and per-user column configurations.
• A GDPR layer: per-project blacklists, a compliance module, unsubscribe handling, bounce events and OAuth rotation for email accounts.
• A proxy layer with sticky sessions: one exit IP is pinned for a block of requests, scoped by a ContextVar, and a dropped session is retried on a fresh session id. Measured: 12 sequential requests gave 11 successes on 11 distinct Finnish IPs, median 1.6 s.
• Container-free production deployment: systemd + Nginx + native PostgreSQL. Development runs on Podman Compose with local overrides.

End-to-end bookkeeping automation for a sole trader, a limited company and a personal account. The system pulls bank statements, fetches receipts from email, reconciles them against transactions, computes VAT and pension expenses, publishes the ledger into Firefly III and generates a finished set of financial statements.

• Bank-statement retrieval and parsing: Nordea (CSV / JSON) and Holvi (XLSX / JSON). The limited company has two Holvi accounts — main and tax reserve — of which only the main one is published to the ledger while the reserve is tracked as a budget.
• Automatic receipt collection through the Gmail API (OAuth2), email-to-PDF conversion, and matching each receipt to the right transaction by amount, date and counterparty.
• A VAT engine: rates and tax classes, reverse-charge VAT, pension-scheme expenses, a counterparty map and rules that live in configuration rather than in code.
• Idempotent publishing into Firefly III: transactions are keyed by external_id, so re-running never creates duplicates. The Finnish chart of accounts is mapped to Firefly categories, budgets are read from a Google Sheets plan, PDF receipts are attached to their transactions.
• A financial-statement generator: a balance sheet and income statement compliant with Finnish accounting law for both company forms, output as XLSX and PDF. Plus a Vero.fi-ready HTML VAT return.
• A planning module: budget versus actuals across three accounts, the source plan coming from Google Sheets and the report published back into a second sheet.
• 20+ runner scripts (.sh and .bat) that drive the whole pipeline with one command; Chrome and fonts are auto-detected on Linux, macOS and Windows.

Replaces the client's Visma Nova based quote calculation with a web application. The pipeline scrapes Oikotie listings, geocodes the buildings, fetches street-level imagery, runs Claude Vision over it, computes man-hours and produces a proposal. It is being extended into per-postal-code prospecting of housing companies and export back to Nova.

• A 10-phase pipeline: Oikotie scraper (cards, details and photos) → Digitransit geocoding → Google Street View (URL-signed) and Mapillary → Claude Sonnet Vision → man-hour formula v2.0 with a v3.0 refinement → a proposal from a single command.
• LLM extraction of property-manager details from free-form text, import of a list of 584 property-management offices and matching against the existing quote database.
• Cost tracking for every LLM call with its own self-test — the price of vision analysis stays visible before a run is scaled to hundreds of targets.
• A 5-page Streamlit dashboard over the results, an 11-table database and importers for the client's Excel dumps.
• The work is split into six parallel stages (S0–S5) that can run in up to nine simultaneous sessions — each with its own file, ownership and acceptance criteria.

The platform ingests bank statements (XLSX), Obsidian vault notes and GitHub repo READMEs, runs eight specialised agent personas in parallel against the user's question, peer-reviews each response and synthesises a chairman's verdict with an antifragility score (0–100) and a top-three action list.

• A five-stage workflow: ingest → an architect agent proposes the council → interactive editing → parallel execution → peer review and synthesis.
• Multi-user support with session isolation, URL-based routing, shareable session links and caching that prevents accidental reruns of expensive operations.
• Real-time token and cost tracking, configurable reasoning depth and live progress tracking during a run.
• A Docker Compose stack: FastAPI back end, React front end, a LiteLLM proxy, ChromaDB for vector search and an optional Telegram bot.

A two-year effort aimed at measuring what part of successful investors' results is actually repeatable. Public top-trader profiles are read programmatically, their open positions are mirrored onto a demo account, and every trade is logged for analysis. Around that grew news collection, headline sentiment analysis and strategy comparison.

• A copy-trading bot for three markets: forex (cTrader / IC Markets), stocks and crypto (Binance). Profiles are polled every ten seconds; a new open position is mirrored and closed when the original closes.
• A parser for traders and strategies from public performance profiles (FXBlue): extracting the users, collecting each strategy's history and comparing them statistically — which results survive scrutiny and which are noise.
• A news pipeline with sentiment analysis: collecting foreign and crypto news (Cointelegraph among others), classifying headlines and mapping the sources behind media indices — signal was separated from noise before anything was traded on it.
• Strategy modelling and testing: trade-history export and PnL analysis, pattern search, demo trades written to a log and results compared across periods.
• Deployment to a dedicated server as an endless loop so the watching continues around the clock unattended; every trade is stored with its date, direction, instrument and size.

• XML-based product and stock synchronisation with the POS every five minutes — millions of SQL statements a day without slowing the storefront down.
• Conversion lifted from 0.2 % to 1 %: SEO, page speed, Redis object cache, image optimisation and a redesigned purchase funnel.
• 14 separate background services: product sync, order export, stock updates, image repair and optimisation, SEO scripts, a Shopping feed, database cleaning and cache warming.

• Two Python microservices connect the store to the ERP: an importer unpacks the Nova archive and upserts the catalogue through the WooCommerce REST API (stock, price, status, categories, brand, facade attributes); an order exporter runs the other way.
• Multilingual with WPML, SEO with Rank Math and WooCommerce HPOS order storage. Product JSON-LD was fixed on a single-source principle so search engines never see two competing Product entities.
• A full local Docker replica of the production store so changes can be tested without touching sales — including cache purging and URL rewriting.

• The same Nova / Loocos integration pair as westface, adapted to a different attribute model (brand, colour, size) and a different order-storage mode.
• A runtime translation layer with TranslatePress — products are stored in Finnish only, so no duplicate post is created per language.
• A local development stack with its own shims (REST authentication without HTTPS) clearly marked local-only so they can never reach production.

• Wall beds and murphy beds as product categories, brochure downloads and a financing application straight from the site.
• Built to last: the site has been in uninterrupted production since 2018 and still runs in 2026 without a rebuild.

• Audience segmentation and per-channel messaging on LinkedIn Ads and Google Ads, with conversion optimisation at every step of the ticket-sales funnel.

• Recurring Excel imports of premises lists are cleaned and normalised; the space-group data model was designed as a written spec before implementation.
• Spam protection on contact forms with Cloudflare Turnstile, with the plan and spec written down before any code.

• A cluster tree of roughly 400 keywords covering the whole industry: bookkeeping, accounting firms, payroll, financial statements, VAT and company formation — each branch mapped to its own page intent, from service pages to the blog.
• A competitor analysis of some 50 industry domains (Accountor, Azets, Rantalainen, Talenom, Netvisor, Visma, Zervant among them): organic traffic, referring domains and backlink depth — what is worth attacking and what is not.
• Results tracked against Search Console data rather than estimates — ranking and click development was recorded and the keyword list revised accordingly.

• Keyword analysis and campaign structure for both niches, technical SEO fixes and expanding the product pages to hundreds of items.
• A walk-through of competitors' backlink profiles on the expansion-joint side: where the industry's links actually come from and which of them are reproducible.

• 2019–2020: the first parsers and a draft lead CRM, target-group analysis, call scripts and lead-tracking sheets; a Helsinki events parser and a niche survey of the Nordic company registries (Proff.no/dk/se).
• 2021–2022: Fonecta and Finder parsers and a full construction-industry company database — 400+ companies with names, domains, decision-makers, revenue brackets and business IDs. On top of that, automated vacancy mailing.
• 2026: the same need, productised — 17 lead sources, an enrichment chain, campaign sequences and a sales pipeline in one platform. Six years of handwork crystallised into a system.

• A local Docker clone of the production site with database import, so changes can be made safely before release.