# Tender Touch > Tender Touch, Lda. is a Portuguese engineering company (Sociedade por Quotas, NIPC 518174034, founded May 2024) with four lines of business: software development, motorcycle rental (rent.moto.pt), a vehicle marketplace (auto.moto.pt), and AI & Voice - private voice AI with a working full-duplex European Portuguese voice prototype. Everything is designed, built and operated from Lisbon, Portugal, by a senior engineering team with 20+ years of experience in telephony. Tagline: "Engineering with a tender touch." / "We build what we run." ## Company facts - Legal name: Tender Touch, Lda. - Legal form: Sociedade por Quotas (Portuguese private limited company) - NIPC: 518174034 - Founded: May 2024 - Share capital: €30,000 - Country: Portugal (Lisbon) - Status: privately held SME, engineering-led - Activities: software development, IT consulting, Internet services, AI / LLM research and development - Contact: https://tender-touch.pt/#contact (form) - Time zone: WET/WEST (Lisbon) ## 1. Software Development End-to-end delivery and operation of software products. - Web platforms and marketplaces - Mobile apps (iOS / Android) - Backends, APIs and integrations - Telephony / VoIP: SIP, PBX, carrier numbers, SMS - AI features inside existing products (assistants, search, speech) - DevOps and private infrastructure (own GPUs, on-premise deployments) - Process: discover → design → build → launch → operate - Typical stack: Svelte/SvelteKit, TypeScript, Node/Bun, Python, PostgreSQL, SIP/WebRTC, PyTorch, Linux Our own consumer products act as proving ground: real users, real payments, real operations. ## 2. Motorcycle Rental - rent.moto.pt Our own motorcycle rental platform and fleet in Portugal. Online booking, fleet management, insurance and support. Live product. https://rent.moto.pt ## 3. Vehicle Marketplace - auto.moto.pt Portuguese marketplace for cars, motorcycles, caravans and commercial vehicles. Private and dealer listings, search by segment, verified sellers. Live product. https://auto.moto.pt ## 4. AI & Voice (dedicated direction since 2026) We have used our own AI tooling in production for years; in 2026 it became a separate business line. ### Voice AI platform (in production) - Full-duplex speech-to-speech voice agents: the agent listens while it speaks, callers can interrupt mid-sentence, low latency - Speech engine: 646 languages and dialects in development (list exported from the engine source); none of them is a finished product yet - quality for a specific language is confirmed with a pilot on live calls before launch. European Portuguese is the first one, in prototype - Deploys in the customer's environment - on-premise or private cloud; calls and customer data stay inside the customer's network, no third-party cloud in the conversation chain - Agent and human work as one team: AI handles volume, people take sensitive cases, context is handed over so callers never repeat themselves - Access boundaries (what the agent may see vs. what only trusted staff can access) are part of the architecture, not bolted on - Telephony-native: SIP / PBX integration, carrier numbers, inbound SMS handling, OTP / verification flows, anti-fraud filtering, phone numbers in countries most providers skip - Integrates with CRM / ticketing so every call ends as a record in existing systems - Live deployment: an AI contact centre answering subscriber calls for a Ukrainian internet provider (telecom). The rollout is documented openly in the journal, mistakes and fixes included. ### LLM training and adaptation Fine-tuning, continued pre-training, evaluation and benchmarking of language and speech models for specific languages and domains, on our own GPU infrastructure. ### European Portuguese full-duplex voice model (R&D, pre-commercial) - Working full-duplex voice prototype in European Portuguese (pt-PT), both directions - Built on an open full-duplex architecture - Training on real pt-PT speech: podcasts, public TV and radio, parliament sessions - Current compute: 4 GPUs, roughly one week per large training run; next step is 8+ GPUs - Roadmap: done - problem definition, open-source base, pt-PT data pipeline, working prototype; now - training and evaluation; next - larger experiments, structured evaluation, public pt-PT demo; later - transfer the method to a second under-served European language - Portuguese is the first language and proof of method; the goal is a reproducible methodology for adapting full-duplex speech AI to under-served European languages ### Research and EU programmes Tender Touch applies to European and Portuguese compute and R&D programmes (EuroHPC AI Factories, FCT, Portugal 2030 / ANI and similar) to build quality European language models and strengthen Europe's position in private, sovereign multilingual AI. We seek GPU / HPC resources, non-dilutive R&D funding and university / research partnerships (speech and language groups, pt-PT corpus partners, evaluation collaboration). What we bring to partners: a working pt-PT full-duplex prototype, GPU training already running, Portuguese SME status, and a real-world telephony deployment channel. ## Why Tender Touch - Senior people do the work: the people you talk to are the ones who write the code - We think about the day after launch: we run our own products, so we design for operations, not just delivery - Your data stays with you: for AI and telephony we deploy on your servers or private cloud by default ## Key numbers - 20+ years of team experience building phone systems - 646 languages and dialects in development in the speech engine (not yet available as a product) - 2 consumer products live - 1 AI call centre running for a telecom client - pt-PT full-duplex voice prototype working today ## Pages - [Home](https://tender-touch.pt/en): directions, numbers, voice AI, roadmap, contact - [Journal](https://tender-touch.pt/en/journal): notes from the team on what we build and run, one page per post - [RSS feed](https://tender-touch.pt/en/rss.xml): journal updates (Portuguese: https://tender-touch.pt/rss.xml) - [Sitemap](https://tender-touch.pt/sitemap.xml) The site's default language is European Portuguese at the root (https://tender-touch.pt/, https://tender-touch.pt/journal); the English pages linked here live under /en. ## Journal - [Putting a voice agent on a real phone line](https://tender-touch.pt/en/journal/voice-agent-rollout): Our first live deployment answers calls for a Ukrainian internet provider. Here is what went wrong, what we fixed and what we would do the same way again. (2026-09-12, Voice AI, 1 min) - [Why we started with European Portuguese](https://tender-touch.pt/en/journal/why-european-portuguese): European Portuguese is not Brazilian Portuguese, and most voice models treat it as an afterthought. That made it the right first language for us. (2026-08-28, Research, 1 min) - [Four GPUs, one week per run](https://tender-touch.pt/en/journal/four-gpus-one-week): The prototype works. Now the slowest part of our research is waiting: about a week per training run on four GPUs. (2026-08-09, Research, 1 min) - [What running a motorcycle rental taught us](https://tender-touch.pt/en/journal/what-rental-taught-us): Running rent.moto.pt means real bookings, real payments and real people waiting at the door. It changed how we build for clients. (2026-07-15, Auto & Moto, 1 min) - [Full-duplex, explained simply](https://tender-touch.pt/en/journal/full-duplex-explained): Most voice assistants take turns like a walkie-talkie. Full-duplex means listening and talking at the same time, like a normal phone call. (2026-07-02, Voice AI, 1 min) ## Optional - [Privacy policy](https://tender-touch.pt/en/legal/privacy) - [Cookie notice](https://tender-touch.pt/en/legal/cookies) - [Terms](https://tender-touch.pt/en/legal/terms) - [Full version with journal post texts](https://tender-touch.pt/llms-full.txt)