Challenge
Whole stacks of invoices, tax letters, insurance policies, payslips and contracts end up scanned into one big PDF. Sorting that by hand means finding where each document starts and ends, naming it and filing it. It's worse when there are no separator pages, when pages were scanned out of order, or when the documents are in several languages.
Approach
Hosted models from the European provider Mistral read every page (OCR) and classify the documents. The real value is in the custom layer around them:
- Boundary detection: finds page counters like "Page 2 of 5" in five languages, uses reference numbers such as policy or invoice numbers, and groups pages by sender. It rebuilds multi-page documents, puts shuffled pages back in order and flags missing pages.
- Classification into 16 categories with multilingual descriptions. If none fits, the model suggests its own category.
- Sanitise mode: rebuilds a clean PDF from the recognised content only – no scripts, macros or embedded files survive.
- Combine mode: turns PDFs, images, Word and text files into a single PDF.
Result
You upload a PDF of up to 100 MB and get back a ZIP of folders named in your language, plus an Excel log. Uploaded files are deleted automatically after a set time.
A real 9-page scan with no separator pages was correctly split into its 4 documents – including a 5-page insurance policy that had previously come apart into five single pages.
What it proves
A production document pipeline with OCR, document splitting, classification, safe PDF rebuilding and five languages can be built and run by a lean team. The value lies in the layer around the model, not in the model itself.
On data location: PaperSweep uses hosted models because that's the better trade-off for a public service. For confidential company documents, we'll discuss a version with locally run models with you.
Challenge
Many piston aircraft carry a JPI engine monitor that records cylinder head and exhaust gas temperatures, oil data and more on every flight. Most pilots never review that data systematically. Early warning signs – a failing exhaust valve, uneven fuel delivery, shock cooling on descent, magneto faults – go unnoticed until they become a maintenance problem.
Approach: compute first, then explain
- Custom parser for the proprietary JPI file format, up to 200 flights per file.
- Rule-based checks in Python: exhaust-valve check, injector balance, induction leaks and carb ice, outlier cylinders by Z-score, shock-cooling rate, magneto check, and flight phases with per-phase statistics.
- Explanation by a language model: Llama 3.1 8B turns the computed findings – not raw data – into a plain-language debrief, with citations from a curated collection of published engine-management guidance.
- Two models, one cross-check: a second model assesses the same flight. Findings flagged by both are marked as consensus – without a third AI call.
Result
Upload a file and get an overall status, anomalies, trends across recent flights, recommended actions and maintenance flags – with sources. The application is about 5,900 lines of Python and JavaScript and uses no frameworks.
Runs on about €5 a month of shared web hosting.
What it proves
A small, inexpensive 8B model gives useful, cited answers when the findings are computed transparently first. For businesses, the same pattern applies to machine and sensor data: compute with transparent rules, then let the model explain in plain language.
Note: HobbsMD is advisory only. It does not replace a licensed mechanic and is not for airworthiness decisions.
Challenge
A cycling group rides twice a week and wants to post about five short clips a week to Instagram and TikTok. Every ride produces hours of GoPro footage. Finding the usable 10 to 40 seconds meant scrubbing through it all by hand.
Approach
Open vision models score the footage with no training at all: SigLIP as the main scorer, OpenCLIP as a cross-check. On top sits a custom layer:
- Plain-language scene categories such as "mountain lake" or "tree-lined road", described in configuration files – new categories need no retraining.
- Ranking within a category: testing showed SigLIP's rankings are reliable but its absolute scores aren't.
- Sensor data instead of guessing: clips only while riding (8 km/h or faster); climbs from the bike computer's barometric altitude, because vision can't judge gradient.
- Automatic place captions from Wikipedia and Wikidata using GPS coordinates, plus automatic editing in portrait or landscape.
Result
One ride's raw footage becomes ready-to-post highlight clips, theme montages, captions, and draft descriptions and hashtags.
Example ride on 11 October 2026: 160 minutes of footage in 32 files → 2,604 frames scored against 48 categories → 4 highlight clips and 5 montages, with no manual footage review.
Before the motion filter, 3 of 4 auto-picked clips were shot while stopped; after it, 0 of 4.
What it proves
Open models plus sensor data the camera already records can replace manual footage review for a small content team – with no model training, no labelled data and no cloud. And where vision falls short, GPS and altitude data are what make the results usable.
Challenge
Student pilots, returning pilots and sim pilots get little chance to practise ATC radio calls before doing it for real. The usual options – an instructor, online networks, or recordings you can't talk back to – don't let you practise any flight you want, whenever you want, with feedback.
Approach
You pick a departure and an arrival airport, and the app builds a complete flight:
- Routes: IFR routes by graph search over real navigation data, VFR routes that avoid controlled airspace, taxi routes over the airport's ground network.
- Radio calls: a language model writes the calls for every phase of flight – grounded in the computed route so they stay realistic.
- Voice loop: the controller speaks through text-to-speech. You answer with push-to-talk, an ATC-fine-tuned Whisper model transcribes you, and a scorer grades your readback.
Result
A working prototype that runs end to end: a full IFR flight with synthesised ATC and spoken readbacks.
What it proves
Domain depth and AI in one product: route computation that follows real aviation rules, a real-time voice loop, and language-model output grounded in computed data. The AI and voice providers have already been swapped without a rewrite.
Note: WilcoRoger is a practice tool only. It is not approved flight training and not for real-world navigation.
Challenge
Little Backup Box is an open-source project that turns a Raspberry Pi into a portable photo and video backup device. Two gaps showed up in practice: it copied everything from a card, including sidecar files and thumbnail folders. And older network storage couldn't be used as a target.
Approach
- Media-only backups: an option to copy only files whose extensions are on an editable list.
- New network target: we made it work with older network storage that only supports a legacy login, without modifying the operating system – for trusted home or office networks.
- Robust on the road: if the network storage can't be reached, the local backup still runs.
- Web interface for the new target in English, German, French and Spanish.
Result
7 commits in about 5 days, +542 / −37 lines – in an existing project with about 1,600 commits.
What it proves
We extend the open-source tools you already rely on instead of rebuilding them – and make new software work with older hardware that's still in use.
Little Backup Box is a project by outdoorbits, licensed under GPL-3.0. Our work is a contribution, not an AleaNCore product.
Challenge
Weekly video campaigns on several Instagram accounts meant posting by hand at set times. Captions and hashtags sat in spreadsheets with nothing to carry them through to publishing.
Approach
Campaign spreadsheets with captions, hashtags and media are imported into one shared queue for all accounts. A scheduler publishes every post that's due – photos, videos, reels, carousels and stories. Failed posts are logged and can be retried individually.
Result
May to October 2026: 92 posts scheduled across 2 accounts, 83 of them published automatically (about 90%).
What it proves
A lean automation that has been in use for months – with no extra SaaS subscriptions. Failures are caught and logged instead of going unnoticed.
PostInsta is internal tooling, in use for the Instagram accounts of Peony Bloom Estates and PaperSweep – not a customer service.
Challenge
Another operator offered to sell Peony Bloom Estates two short-term rental units. The asking price rested on estimates, not on what a short-term rental actually costs to run month by month. Both sides needed one transparent model they could check and adjust themselves.
Approach
- Revenue model: occupancy with monthly overrides, seasonal nightly rates and platform commission.
- Cost model: 12 cost lines from real operations – from rent and utilities to cleaning and the local tourist tax; cleaning and tourist tax are derived from bookings and nights.
- Break-even projection: profits are accumulated month by month until the investment is recovered – with a clear message when that never happens or takes decades.
- Excel as the source: all default values come from a workbook and can be changed without touching code.
- Bilingual and shareable: the other side could enter their own numbers online and see profit and break-even recalculated instantly.
Result
About 1,000 lines of R, built from 2 to 15 April 2026 and deployed on 15 April 2026.
What it proves
Statistics and real operating data become a decision tool for a real investment. Instead of arguing over spreadsheets, both sides work from one model they can test themselves. The same approach fits any purchase, investment or pricing decision.