Solar VAT Pipeline
A solar installer was filing hundreds of VAT returns by hand. We built an end-to-end pipeline: document reading, AI extraction, automated filing. The result: zero manual data entry.
Zero manual entry30+ projects sorted by domain. No buzzwords, no vague spec sheet — each piece of work explained in a single card, with what we did and the result it produced.
Some clients remain under NDA; every fact can be verified on request.
A solar installer was filing hundreds of VAT returns by hand. We built an end-to-end pipeline: document reading, AI extraction, automated filing. The result: zero manual data entry.
Zero manual entryAn operator was receiving thousands of energy-efficiency subsidy files to review. We built a detection engine: document fingerprinting, pixel-level analysis of signatures and stamps, AI extraction by document type. Reviewers now validate alerts instead of reading everything.
Thousands of files / monthA publisher needed to mask personal data (names, addresses, signatures) at the pixel level, within a few days. We tested Florence (Microsoft) and shipped a tool that automatically finds and masks those zones. Turnaround cut by 5x.
Turnaround /5Thousands of addresses to turn into land-registry parcels. 9 pipeline versions to cut false positives. V9: ~53% automated.
~53% auto (V9)12 control blocks for energy-efficiency subsidy files. Checks on the incentive, eligibility, the obligated party, products, quotes and filing. All orchestrated into one pipeline and tested end to end.
12 control blocksAutomatic contact discovery from company lists, centralized in a single interface, then synced automatically to the INES CRM. Humans only step in on the doubtful cases.
From list to CRM, no re-keyingThe team needed a purpose-built AI memory. We tested Graphiti, Cognee and others: none fit. So we built a custom architecture. 50 agents running in parallel processed 16,726 entries in 45 min.
16,726 entries in 45 minA dedicated platform for managing teams of AI agents. Complete architecture, 18 founding files, documented V2 spec.
18 founding filesAn AI agent that builds quotes by talking with the client. Multi-agent orchestrator, catalogue-driven extraction, conversation replay.
NDA — client anonymizedAn agent that watches YouTube videos for you and pulls out the essentials: summary, topics covered, keywords and a usable brief. Enough to follow a channel or a competitor without spending hours on it.
A video summed up in one briefWhile the sales rep is on the phone, the AI listens, pulls context from the CRM and feeds the right talking points in real time. Several models orchestrated together.
Real-time multi-LLMAn assistant that answers from the company's internal documents, with per-team access rights, connected to the existing Microsoft directory. Every answer cites its sources.
Sourced, traceable answersA fully automatic daily run: reading the news feeds, ranking by relevance with AI, then writing a bilingual digest with in-depth analysis. Runs locally, with no server or infrastructure to manage.
An hour of monitoring in a 5-min readA complete CRM for a craft workshop: one-off pieces, 5 roles, multi-directional statuses, beneficiary verification. Custom WordPress + Alpine.js.
5 roles, multi-directional statusesA complete map of how a complex company actually runs. Flows, automations, roles — all documented.
Complete mapping200 Zoho automations, 102 workflows, 367 triggers checked under real conditions over 60 days. 43% of dead workflows identified.
43% dead workflowsA web app for small-business owners: account creation, secure sign-in and a personal workspace. Verified end to end with automated tests and shipped on Railway.
Tested end to endDoes one of these look like your case?
Let's talk about your project →Building a company database, with a Rust architecture for performance-critical work on large volumes.
High-performance RustHeterogeneous data streams turned into usable dashboards through Airbyte.
Multi-source, unified58 Terraform resources under management. A multi-service cloud architecture (Lambda, SQS, DynamoDB). Shared across 3 projects.
58 Terraform resourcesStanding up a database fast from several heterogeneous sources. AWS infrastructure sized to the time constraint.
Multi-source, consolidatedA business platform: a catalogue of 1012 bilingual products, land-registry lookup, quotes, file assembly, verification. 6 Railway micro-services.
1012 products · 6 micro-servicesMigrating WordPress sites to a headless Astro architecture. Fast sites, better SEO, the team still on WordPress behind the scenes.
SEO 100/100Contacts structured by department, internal routing, multi-channel communication. Monday.com integrated.
Multi-department, routingA deep technical analysis of a production system. Diagnosis, prioritized action plan, strategy document delivered to the client.
Full audit + strategyAn admin interface for 1012 FR/HE products: full management, photos, assisted data entry and cascading dependencies.
1012 FR/HE productsReverse-engineering of a signature tool locked to a single PC. Exposed as a service with a queue for the whole company.
1 locked PC → N seatsA native desktop app: Tauri + Rust + whisper-rs. Multilingual transcription, local database, distributable.
Desktop app · MultilingualAutomating a business process thought impossible to automate, worked around by reconstructing how a locked tool operated. Built under a tight time constraint. Precise details under NDA.
Confidential detailsProprietary software delivered as an .exe, with no source code. Extracting the PyInstaller binary, disassembly, then a readable reconstruction of the three-phase power calculation in documented Python. Every formula recovered and checked against the original tool.
A black box turned back into source codeA sensitive-data processing solution for a European fintech. Secure architecture, compliance pipeline.
A business management system for an international hotel group. Multi-site integration, real-time synchronization.
A multi-LLM AI copilot orchestrated in real time for a sales coaching platform.