Local AI uses
Everything local AI can already do in your industry
120 concrete uses of artificial intelligence with open models, organised by industry and by level of maturity: what is already in production, what is being piloted and what is starting to work. Everything listed here can run on your own servers, in Barcelona or anywhere else, without sending data to the cloud.
- ideas
- 120
- Industries
- 20
- We have built it
- 14
- In productionIn production: used today in companies and organisations
- In pilotsPilot: real organisations are testing it with measurable results
- ExperimentalExperimental: emerging technology that already works in the lab and is maturing fast
- We have built itWe have built it and it works in our Barcelona lab or for clients
Law firms and advisory firms
Professional confidentiality and the GDPR rule out pasting client files into a cloud chatbot. With open models on the firm's own server, AI works on the documents without them ever leaving the firm.
-
Search with citations across the firm's case files and documents
A RAG system indexes pleadings, contracts and rulings with embeddings in Qdrant and answers by citing the exact document and passage. It works with open models such as Qwen or Mistral on a server at the firm, without exposing client data to third parties.
-
Summaries and timelines of long case files
A long-context model served with vLLM reads hundreds of pages of proceedings and produces a summary, a timeline of events and a list of parties and deadlines. The lawyer reviews and corrects it; the documents never leave the internal network.
-
Anonymising judgments and pleadings before sharing them
Named-entity recognition models and a local LLM detect names, ID numbers, addresses and amounts and replace them with consistent placeholders. Useful for publishing rulings, training junior staff or sending documents to expert witnesses in compliance with the GDPR.
-
Contract review against the firm's own standards
The model compares each clause with the firm's standard positions, flags deviations, risks and missing clauses, and suggests alternative wording. It combines RAG over the internal playbook with an open LLM such as Qwen or GLM, without breaching professional confidentiality.
-
Draft pleadings in the firm's house style
An open model fine-tuned with LoRA on the firm's own pleadings drafts claims, formal demands or defences with its structure and tone. Always as a draft reviewed by a lawyer, who keeps professional and ethical responsibility for whatever they sign.
-
Verifying citations of case law and legislation
An agent checks that every judgment, article or regulation cited in a document exists and says what is claimed, comparing it against a local database of case law and legislation. It reduces the risk of invented citations, the most serious failure of generative AI in legal work.
Clinics and healthcare
Health data is a special category under the GDPR, and Catalan public healthcare already prioritises open models in Catalan such as those from the AINA project. Local AI helps with documentation and administration; clinical judgement always rests with the professional.
-
Transcribing consultations into a structured report
Whisper transcribes the consultation in Catalan or Spanish and a local LLM turns it into a draft report with patient history, examination and plan. The doctor reviews, corrects and signs it; neither the audio nor the text leaves the centre.
-
Internal protocol assistant for clinical staff
Doctors and nurses query the centre's protocols, guidelines and procedures in natural language and get answers that cite the current document. RAG with multilingual embeddings and a local model; clinical decisions always rest with the professional.
-
Plain-language instructions and letters for patients
Starting from a technical report, the model drafts discharge instructions, test preparation guidance or reminders in plain language and in the patient's own language. A professional approves every text before it is handed over.
-
Extracting data from lab results and scanned reports
OCR and open vision models read laboratory reports or reports from other centres and transfer values, units and dates into a structured table that staff review. It saves manual transcription, and the documents are not uploaded to any external service.
-
Medical record summary before each visit
Before the appointment, a model summarises previous reports, test results and medication in a short overview with a link to each original document. It is a reading aid for the professional, not a diagnosis, and it runs inside the clinic's network.
-
Assisted ICD-10 coding of discharge reports
The model suggests ICD-10 codes from the report text and highlights the sentence each one is based on; the coder accepts or corrects them. It uses an open LLM with RAG over the official classification, without sending medical records to third parties.
Industry and engineering
Drawings, formulas, processes and machine data are trade secrets. Across Catalonia's industrial base, local AI lets companies put that information to work on the shop floor and in the engineering office without depending on the internet or on third parties.
-
Technical assistant for plant manuals and procedures
Operators and maintenance staff ask how to calibrate a machine or what an error code means, and the system answers citing the right manual. RAG with Qdrant and an open model on a plant server that works even without an internet connection.
-
Visual quality inspection with in-line cameras
Detection models such as YOLO, trained on the company's own images, spot defects, badly assembled parts or incorrect labels on the production line. Inference runs on a device next to the camera, with millisecond latency and without sending images outside.
-
Incident and 8D reports from voice notes
The technician dictates what happened at the machine, Whisper transcribes it locally and an LLM organises it into the company's incident or 8D report format. Less paperwork at the end of the shift and a searchable record of breakdowns.
-
Natural-language questions about production data
An agent turns questions such as ‘which line had the most stoppages this month?’ into SQL or time-series queries and returns the answer with a chart. MES and SCADA data stays on the plant network.
-
Requirements matrices from tenders and technical standards
A long-context LLM reads tender documents, customer specifications or standards and builds a requirements matrix that references the source clause. It speeds up bids and product approvals without exposing confidential customer documents.
-
Predictive maintenance with an explanation for every alert
Anomaly detection models on vibration, temperature or power consumption flag deviations, and a local LLM cross-checks each alert against the breakdown history to suggest a likely cause and action. The technician decides, and how they close the alert feeds back into the history.
Education and universities
Schools, secondary schools and universities handle data about minors and student work that should not end up in commercial services. More and more institutions, in Catalonia too, are buying servers to run language models on their own campus.
-
Institution-wide AI chat with accounts and usage controls
A ChatGPT-style interface on open models served with vLLM, with access through institutional accounts, limits per profile and usage logs. Students and teachers use AI without conversations or personal data leaving the campus.
-
Exercises and learning materials adapted to each level
From a topic or a text, a local LLM creates exercises, quizzes and easy-read versions for students with special educational needs. It works in Catalan and Spanish with open models, and the teacher selects and adjusts the output.
-
Transcription and subtitles for lectures and talks
Whisper transcribes and subtitles recorded classes, seminars and thesis defences in Catalan, Spanish or English, with timestamps for searching within the video. It improves accessibility without uploading recordings to external services.
-
Virtual tutor for a vocational training course on Discord
A bot connected to a local model answers vocational students' questions about the course syllabus at any time of day, inside the class's Discord server. Teachers control the content and no student data is handed over to third parties.
-
Synthetic data and LoRA fine-tuning to teach music and maths
We generate synthetic datasets and fine-tune an open model with LoRA so that it explains music and maths concepts using the teaching approach of an educational project. A small, specialised model that runs on modest hardware at almost zero cost per query.
-
Assisted marking with rubrics and feedback for students
The model assesses essays or exercises against the teacher's rubric and drafts personalised feedback that the teacher reviews before returning it. Student work is processed on the school's server, not on an external platform.
Public administration
Town councils, provincial councils and public bodies must comply with Spain's National Security Framework (ENS), serve citizens in Catalan and Spanish and reduce their dependence on clouds outside the EU. There are already tenders to buy in-house LLM servers and model platforms with per-request traceability.
-
Transcripts and draft minutes of council and committee meetings
WhisperX transcribes the session, distinguishing each speaker, and a local LLM produces draft minutes with resolutions and votes. The council secretary's office reviews and approves the document; the recording never leaves the institution.
-
Classifying incoming documents at the official registry
OCR and a local LLM read the applications and documents that arrive at the registry, extract the sender, subject and case file, and suggest which department should receive them. It cuts repetitive work while keeping personal data inside the administration.
-
Citizen assistant on procedures and by-laws
A multichannel assistant answers citizens' questions about procedures, fees and by-laws, citing the current official document, and hands over to a person when there is no clear answer. RAG over the body's documentation with an open model hosted on its own servers.
-
Model platform with traceability and cost control
A single gateway gives every department access to several open models, with authentication, logging of every request and usage limits. It is the pattern some public tenders already require for using AI with auditing and without being tied to a single vendor.
-
Support for reviewing procurement tender documents
The model compares tender documents with public sector procurement law and with previous tenders, flags inconsistent clauses and prepares summaries for the contracting committee. It supports the officer in charge, who is the one who signs and decides.
-
Faster Catalan generation with an extended vocabulary
We extend the model's vocabulary with frequent Catalan word fragments so that it needs fewer tokens per sentence and generates Catalan text faster and at lower cost. Especially useful for administrations and organisations that work mainly in Catalan.
Agriculture, vineyards and wineries
Out in the fields the signal drops, and farm data belongs to the farmer. From the Penedès to Lleida, local AI runs on a computer at the winery or the cooperative, even without a connection.
-
Tasting notes, labels and sales copy for each wine
Using the data for each vintage, a local model drafts tasting notes, copy for the online shop and translations into English, French or German in the winery's own voice. No cost per text, and it respects the terminology of the designation of origin.
-
Vine pests: diagnosis from photos and voice notes in Catalan
The grower photographs the leaf or the bunch and dictates what they see in Catalan; an open vision model suggests the likely pest or disease and the note is transcribed into the field log. It runs on a family-owned estate in the Penedès, and the agronomist confirms before any treatment.
-
Weather station history in natural language
Questions such as ‘how many hours of leaf wetness were there this week?’ are answered by querying sensor data stored in InfluxDB and integrated with Home Assistant. A local model interprets the question, queries the history and explains the result.
-
Field log dictated by voice
Treatments, irrigation and field work are dictated from a mobile phone and a local model turns them into structured entries in the farm's field log. Less paperwork at the end of the day and data ready for inspections and certifications.
-
Assistant on CAP aid and DO regulations
Farmers and cooperative advisers ask about aid requirements, the designation of origin's product specification or authorised products, and get the answer with a citation to the official document. RAG over regulations that are updated periodically.
-
Yield estimation with computer vision
Detection models count bunches, fruit or plants in photos and videos taken in the field or by drone to estimate the yield of each plot. Processing happens locally with open models fine-tuned on images from the farm itself.
Architecture, engineering and design studios
Barcelona is full of studios that have bought a Mac Studio or a powerful workstation and barely use it for AI. With open models, that machine can generate images, look up regulations and organise projects without uploading client drawings to the cloud.
-
Local image generation and concept variations
Open diffusion models such as Stable Diffusion or FLUX, guided with ControlNet, generate facade variations, interior design options or moodboards from a sketch or a scale model. They run on the studio's workstation, with no per-image fees and no handing over of designs.
-
Querying the building code (CTE) and planning rules with citations
A RAG system over Spain's Technical Building Code (CTE), municipal by-laws and planning documents answers questions by citing the article and document. Useful at the preliminary design stage to check parameters without searching through hundreds of PDFs.
-
Site meeting minutes with tasks and owners
Meetings with the client and the contractor are recorded, transcribed with Whisper and turned into minutes with agreements, owners and deadlines. Everything is processed on the studio's server, without sending client conversations to third parties.
-
Project reports drafted from templates
The model produces drafts of descriptive reports, construction reports or specifications by combining the studio's templates with each project's data. The architect reviews and signs; the client's documents stay on the studio's server.
-
Visual search across the project and reference archive
Image embeddings such as CLIP index renders, site photos and references so they can be searched by description (‘brick facade with lattice screen’) or by similarity to another image. The studio's historical archive becomes searchable without tagging it by hand.
-
Reading drawings and quantity take-offs with vision models
Open vision-language models are starting to interpret PDF drawings to extract floor areas, element counts or legends. Today they are useful as a first pass for the technician to check, not as a final measurement.
Retail and e-commerce
Physical and online shops generate thousands of product listings, reviews and enquiries every month. With local AI, the cost per generated text is close to zero and customer data does not pass through external services.
-
Product listings and translations at scale
An open LLM generates descriptions, SEO titles and attributes from supplier data and translates them into Catalan, English or French in the brand's voice. With thousands of products, not paying per token to an external provider makes all the difference.
-
Semantic search for the online shop
Customers search the way they talk (‘trail running shoes for wet weather’) and a search engine with embeddings and Qdrant finds the right products even when the words do not match. It runs on the shop's own infrastructure.
-
Analysis of reviews and return reasons
The model classifies thousands of reviews, tickets and return reasons by topic and sentiment, and summarises what is going wrong with each product. Customer information is analysed without sending it to a third-party platform.
-
Product photography: backgrounds, cut-outs and lifestyle scenes
Open segmentation and diffusion models remove backgrounds, generate lifestyle scenes and adapt each image to every channel's formats, starting from a basic photo. No per-image fees and full control over the result.
-
Shopping assistant connected to the catalogue and stock
A chat on the website answers questions about sizes, compatibility, availability and delivery times by checking the catalogue and the ERP in real time. The model runs locally and only accesses the data it has been authorised to use.
-
Demand forecasts explained in natural language
Open time-series models such as Chronos forecast sales for each product and a local LLM explains the forecasts and suggests orders to the purchasing manager. Sales history, which is commercially sensitive, never leaves the company.
Hospitality and tourism
Barcelona welcomes visitors from all over the world, and every hotel, restaurant or agency answers the same questions in many languages. Local AI handles enquiries and translates without sharing guest data with external platforms.
-
Multilingual replies to bookings and guest enquiries
A local model drafts replies to emails and messages in the customer's language, with the establishment's availability, opening hours and policies. The team reviews and sends them; guest and booking data stays in-house.
-
Review replies in the establishment's own voice
The model suggests personalised replies to reviews in any language, spots recurring complaints and flags those that need management's attention. It runs on an open LLM, with no per-message fees.
-
Audio guides and tourist content in many languages
Open translation and speech synthesis models produce texts and audio guides for visitors in around ten languages from a script in Catalan or Spanish. Produced locally, with no cost per minute of audio.
-
Multilingual menus and allergen sheets for each dish
Starting from the recipes, the system produces the menu in several languages and, for each dish, a sheet covering the 14 allergens that must be declared by law, for the kitchen to review. When an ingredient changes, every version is updated.
-
Virtual concierge for guests
An assistant on the website or in the room answers questions about services, opening times, transport and local recommendations using the hotel's own information. RAG over internal documents and a local model, without sending conversations to third parties.
-
Staff rotas and purchasing based on forecast occupancy
An agent cross-references forecast occupancy, city events and the establishment's history to propose staff rotas and supplier orders. The manager approves the proposal, and staff data never leaves the business.
Logistics and transport
Delivery notes, CMRs, customs and incidents: logistics runs on repetitive paperwork. Local AI processes it in the warehouse or office itself, keeping customer and rate data safe.
-
Automatic reading of delivery notes, CMRs and freight invoices
OCR and open vision-language models extract the sender, consignee, packages, weights and amounts from scanned or photographed documents and enter them into the ERP or TMS. The operator only reviews doubtful cases.
-
Sorting incoming emails and orders
A local LLM reads the operations inbox, tells orders, changes, complaints and status enquiries apart, extracts the key data and creates the corresponding task. It integrates with n8n or the company's own system without sending emails to the cloud.
-
Handling delivery incidents with a suggested reply
For each incident, the system gathers the shipment tracking, the driver's photos and the customer's history, summarises the case and suggests a reply and compensation in line with company policy. A person approves before anything is sent.
-
Suggested TARIC codes for customs documentation
The model suggests the tariff heading for a product from its description and data sheet, referencing the explanatory notes of the nomenclature. The customs broker approves it, and the customer's commercial information never leaves the company.
-
Driver voice notes turned into incident reports
The driver dictates the incident in their own language, Whisper transcribes it locally and a model turns it into a report with time, place, type and actions taken. Especially useful for fleets with drivers of several nationalities.
-
Agents that replan routes when the unexpected happens
An agent combines a classic route optimiser with an LLM that interprets customer notices, road closures or staff absences to propose a new plan to the transport manager. The final decision is always made by a person.
Energy, water and utilities
Energy retailers, water companies and installers work with critical infrastructure subject to NIS2. Running AI locally avoids exposing operational data and keeps the service going even if the external connection fails.
-
Assistant for operating and safety procedures
Technical teams query procedures, diagrams and safety rules from the internal network and get answers that cite the current document. RAG with an open model running locally, available even if the internet connection goes down.
-
Classifying fault reports and complaints
A local LLM classifies fault reports, meter readings and complaints by type and urgency, extracts the address and contract number and routes them to the right team. Customers' supply data stays within the company.
-
Analysing installation inspection images
Vision models detect corrosion, hot spots in thermal images, or dirty or damaged solar panels in drone or inspection photos. The technician confirms each finding and the system generates the report.
-
Natural-language queries on meters and telemetry
Technicians and managers ask about consumption, peaks or losses, and an agent turns the question into queries on the time-series database, returning figures and charts. Network data never leaves the company's infrastructure.
-
Equipment status and power usage dashboard built by agents
Coding agents built a dashboard that shows, in real time, the status of the machines, the power they draw and the electricity cost, based on their own data. It is maintained and extended by asking for changes in natural language, with human review.
-
Building energy management with agents
An agent cross-references the weather forecast, hourly electricity prices and actual consumption to suggest when to charge batteries, switch on the heating or air conditioning, or shift processes. It runs locally alongside systems such as Home Assistant.
Banking, insurance and finance
Banks and insurers headquartered in Barcelona are setting up internal AI offices under DORA, the GDPR and the EU AI Act. Open models on their own infrastructure let them use generative AI without handing over customer data.
-
Reading claim forms and claim file documents
OCR and vision-language models extract data from joint accident reports, repair invoices and loss adjusters' reports and complete the claim file. The claims handler approves it, and the policyholder's documents never leave the insurer.
-
Internal assistant on products, policy terms and circulars
The branch network and the call centre look up policy terms, internal procedures and circulars, with answers that cite the source. RAG over the institution's documentation with an open model in its own data centre.
-
Call summaries with complaint detection
Calls are transcribed with Whisper and a local model produces a summary with the reason for the call, the sentiment and whether there is a formal complaint to log. Recordings and transcripts stay inside the institution.
-
Preparing documents for loan assessments
The model extracts income, debts and key data from payslips, tax returns and bank statements to prepare the assessment sheet. The decision belongs to the analyst: creditworthiness assessment is high-risk under the EU AI Act and requires human oversight.
-
Compliance review of marketing communications
An LLM checks brochures, emails and sales scripts against transparency rules and internal policies, and flags problematic wording before publication. It runs locally and leaves an audit trail for every review.
-
Explaining fraud alerts to the analyst
Classic fraud detection models flag suspicious transactions and a local LLM writes a clear explanation of the factors that weighed most. It speeds up the analyst's review without taking transaction data out of the institution.
Media, agencies and marketing
Agencies, media outlets and marketing departments produce content every day. With open models running locally, the cost per text, image or voice-over plummets, and client briefs do not pass through third-party services.
-
Daily industry news report generated locally
Every morning, a process gathers selected sources and websites, a local model filters what matters and writes a summary report with links to the originals. We use it every day to keep up with AI news, with no per-token cost even when it processes hundreds of articles.
-
Voice-overs with a synthetic voice cloned with consent
Open speech synthesis models produce voice-overs, audio guides or audio versions of articles with a voice cloned from samples of a person who has given explicit consent. Everything runs locally, without handing the voice over to any platform.
-
Transcription and subtitles for videos and podcasts
Whisper transcribes interviews, videos and podcasts in Catalan, Spanish or English and generates synchronised subtitles; a local LLM then extracts quotes, chapters and social media copy. No cost per minute, however large the volume.
-
Adapting content to each channel and to the brand voice
From a press release or an article, the model produces versions for the website, newsletter and social media that follow each client's style guide, supplied as context or through LoRA fine-tuning. Confidential briefs never leave the agency.
-
Content-based search in video and photo archives
Vision and transcription models index the audiovisual archive so it can be searched by what is seen or said (‘press conference at the port, 2019’). Archives of thousands of hours, processed locally with no cost per minute.
-
Agents that build WordPress blocks with visual checks
For a web agency, coding agents build Gutenberg blocks from a design, test them in a browser and compare screenshots with the mock-up until they match. A person reviews the result before it is published.
Human resources
CVs, payslips and appraisals are sensitive personal data, and the EU AI Act classes AI used to select or evaluate people as high-risk. Local AI helps with HR administration, with human oversight and without handing over staff data.
-
Employee assistant on collective agreements, leave and policies
Staff ask about holidays, leave, the collective agreement or expenses and get answers that cite the current internal policy. Local RAG over HR documentation, without exposing internal information to third parties.
-
Structured extraction from CVs
The model turns hundreds of CVs in different formats into a common profile with experience, education and languages so the recruitment team can compare them. It does not decide on or score candidates: that use is high-risk under the EU AI Act and requires human oversight.
-
Clear and inclusive job ads and job descriptions
From a few notes, an open model running locally drafts clear job ads and job descriptions in Catalan and Spanish and flags biased or exclusionary wording. The recruitment team reviews the final text.
-
Analysing engagement surveys and open-ended comments
A local LLM groups thousands of open-ended responses from internal surveys by theme and sentiment and summarises areas for improvement by department. Anonymous staff responses never leave the company.
-
Guided onboarding for new starters
An assistant supports new starters through their first days: it explains tools, processes and contacts for their role and answers questions using the company's documentation. Fewer interruptions for the team and answers that are always consistent.
-
Simulator for interviews and difficult conversations
A local model role-plays a candidate or a team member by voice so that managers can practise interviews, feedback conversations or sensitive situations, followed by an assessment. The practice sessions are not stored on external servers.
Customer service
Switchboards, chats and support inboxes are full of conversations containing personal data. With local models, repetitive work is automated in Catalan, Spanish or English, with no per-conversation cost and no dependence on a single provider.
-
Support chatbot built on your knowledge base
An assistant answers frequently asked questions by citing the documentation, terms and conditions or product manual, and passes the conversation to a person when it does not know the answer. RAG with Qdrant and an open model served with vLLM on your own infrastructure.
-
Draft replies for the support team
For each ticket, the model suggests a reply using the customer's history and similar cases already resolved; the agent adjusts it and sends it. It cuts response times without taking control away from people.
-
Transcribing, summarising and tagging every call
Whisper transcribes every call and a local LLM summarises it, identifies the reason and sentiment and tags it in the CRM. It makes it possible to analyse one hundred per cent of calls without sending recordings to the cloud.
-
Voice agent that answers the switchboard and books appointments
Local speech recognition, an LLM and speech synthesis answer calls to give information, take messages or book appointments, and transfer to a person when needed. Latency is low because everything runs on your own server.
-
Detecting customers at risk of leaving
The model analyses tickets, emails and calls to detect frustration, repeated complaints or an intention to switch provider, and alerts the account manager. Sensitive commercial information, processed locally.
-
A knowledge base that grows with resolved tickets
An agent spots new questions in closed tickets, drafts the corresponding help article and submits it for review before publication. Support documentation grows at the pace of real cases.
Software, IT and technical teams
Development and IT teams were the first to adopt local AI: code is intellectual property, and the per-token cost of an assistant used all day soon soars. It is also what we use most every day in our Barcelona lab.
-
Private ChatGPT-style chat with accounts, budgets and usage tracking
A model gateway gives each team member their own account, a budget and access to several open models, with a per-user usage dashboard. Ours has already served about 11.15 billion tokens without paying any provider per token.
-
Code assistant in the editor with local models
Open models specialised in programming, such as Qwen Coder or DeepSeek Coder served with vLLM, provide autocompletion and chat inside the IDE. Source code never leaves the company and costs do not grow with each developer.
-
Questions about the code, the wiki and internal documentation
RAG over repositories, wikis and tickets lets you ask how a module works or why a decision was made, with links to the source code or page. Very useful for bringing people onto large projects.
-
Automatic merge request reviews
An agent reviews each change for bugs, security issues and deviations from the team's guidelines, and leaves comments on the merge request itself. The human reviewer decides, and the repository is not shared with external services.
-
Incident diagnosis from logs and metrics
When an alert fires, an agent gathers logs, metrics and recent changes, summarises what happened and suggests hypotheses and diagnostic commands to the on-call engineer. Logs, full of internal data, are analysed locally.
-
Virtual company of development agents with human approval
Several agents with different roles (product, development, review and testing) work on GitLab: they open issues, write code and propose merge requests, but nothing is merged without a person's approval. Everything runs on local models.
Cybersecurity
Security teams cannot send logs, vulnerabilities or incident reports to a cloud AI. Local models automate analysis while keeping the most sensitive information inside the perimeter, even on isolated networks.
-
Full AI stack on isolated networks with no internet connection
Open models, the inference engine and RAG are installed on air-gapped networks, with model updates delivered on verified physical media. For defence, critical infrastructure or laboratories where no data can leave.
-
Analysing suspicious emails and reported phishing
A local model examines the text, links and attachments of emails reported by staff, explains why they are suspicious and replies to the employee. Internal emails are not sent to external analysis services.
-
SIEM alert triage with summary and context
A local LLM enriches each alert with asset data, related events and threat intelligence, and suggests a priority and next steps to the analyst. It cuts the noise without taking logs out of the organisation.
-
Prompt injection testing for chatbots and agents
Before an assistant or an agent goes into production, it is attacked with batteries of prompt injection, data extraction and tool abuse tests generated with local models. The result is a risk report with recommended fixes.
-
Incident reports and draft NIS2 notifications
From the technical timeline, the model produces the incident report and the draft notifications required by NIS2, in language suited to senior management and to the authority. The security lead approves them before they are sent.
-
Agents that hunt for vulnerabilities in code and infrastructure
Agents running on local models analyse code and infrastructure as code for vulnerabilities, try to reproduce them in a test environment and propose a patch. They complement classic static analysis tools.
Real estate and property management
Estate agents and property managers handle contracts, deeds and data on owners and tenants. In a market as busy as Barcelona's, local AI speeds up paperwork without exposing personal data.
-
Property listings in several languages from photos and data
A vision model describes the property photos and a local LLM writes the listing with the technical details in Catalan, Spanish and English, adapted to each property portal. The agent reviews it before publishing.
-
Extracting data from deeds, land registry extracts and contracts
OCR and a local LLM extract owners, encumbrances, floor areas, cadastral references and dates from scanned documents into a structured record. Documents containing personal data, processed without leaving the office.
-
Minutes of owners' association meetings from the recording
The meeting is recorded, WhisperX transcribes it identifying each speaker and a local model drafts the minutes with resolutions and votes. The property manager reviews them before sending them to the owners.
-
Virtual home staging of empty rooms
Open diffusion models furnish and decorate photos of empty rooms or show possible renovations, and the listing states that the images are virtual. No per-image cost, on the agency's own workstation.
-
Assistant for tenants and owners to report issues
A chat collects faults and enquiries, asks for photos, classifies the urgency and creates a work order for the right tradesperson, checking leases and the building's rules when needed. Service at any hour without handing data over to third parties.
-
Indicative valuation explained with comparable properties
The system looks for comparable properties in the agency's own database, adjusts for floor area, condition and floor level, and an LLM explains the price range with the comparables used. It is only indicative and is always validated by a professional.
Third sector and NGOs
Foundations, social organisations and NGOs support vulnerable people on tight budgets. Catalonia has a very active third sector, and local AI makes automation possible without paying per use or exposing highly sensitive data.
-
Drafting project proposals and grant reports
Using previous reports, activity data and the rules of the funding call, the model produces draft proposals and grant reports with the required structure. The team reviews, completes and submits them.
-
Easy-read documents
The model adapts letters, regulations and leaflets into easy-read versions for people with intellectual disabilities, older people or newcomers, following easy-read writing guidelines. A person always validates the final text.
-
Annual report and impact indicators
A local LLM groups and summarises surveys, testimonials and activity data to prepare the annual report and the impact indicators that funders ask for, without exposing participants' data.
-
Volunteer assistant for protocols and resources
Volunteers and staff look up action protocols, local social resources and internal procedures, with answers that cite the source. A small RAG system that runs on a modest computer at the organisation itself.
-
Summaries of social care case files
The follow-up notes for each person supported are summarised in an up-to-date record for the intervention team, with a link to each original note. Information about vulnerable groups that never leaves the organisation.
-
Voice interpreting with people who do not speak the language
Speech recognition, translation and speech synthesis with open models help staff communicate with people who speak Arabic, Urdu, Romanian or Ukrainian. It is day-to-day support that runs on a laptop, not a substitute for interpreters in critical situations.
Freelancers and office work
For independent professionals and small offices, local AI is one more assistant on the team: it transcribes, searches, drafts and organises on your own computer or server, with no monthly per-user fees.
-
Meeting transcription that identifies each speaker
WhisperX transcribes the meeting, diarisation with pyannote separates each speaker and a local model produces the minutes with agreements and tasks. We use it every day, and the audio never leaves the computer.
-
Search with citations across your own documents
Contracts, emails, PDFs and notes are indexed with embeddings in Qdrant and queried in natural language; every answer cites the exact document and passage it comes from. Nothing is uploaded to the cloud.
-
Reading invoices and receipts for the accounts
OCR and open vision-language models extract the supplier, tax ID, taxable base, VAT and date from invoices and receipts, and export them to the accounting software or to your accountant. Tax documents do not pass through third-party platforms.
-
Email under control: sorting, summaries and drafts
A local model sorts incoming email, summarises long threads and suggests replies in your own tone. It integrates with your email client or with n8n, without handing your inbox over to an external service.
-
Private assistants for each person with their own memory
Each person has their own assistant, which remembers their preferences, projects and previous conversations in a separate memory that nobody else can see. It runs on local models and the memory is stored on your own server.
-
An agent that fills in forms and handles repetitive paperwork
An agent with an automated browser and a local model fills in web forms, downloads certificates or updates spreadsheets following your instructions, and asks for confirmation before each submission. It still needs supervision at every step.
Which of these fits your business?
In a 30-minute call we review your processes and your data, and tell you which uses make sense locally, what hardware you need and where to start. And if your case does not need local AI, we will tell you that too.