Questions, answered

TwoPoros is a software engineering studio based in Bucharest, Romania. We build and run the software businesses depend on: AI systems and automation, web and mobile applications, and the infrastructure underneath, with private and local AI as our specialty.

TwoPoros is a software studio based in Bucharest, Romania. We work fully remote, mainly across European time zones.

Yes. TwoPoros works with clients across Romania and internationally, fully remote.

TwoPoros's specialty is private and local AI: we deploy language models on infrastructure you control, so your data never has to be sent to an external AI vendor. We also cover the full lifecycle, from design and development through deployment and maintenance, so one team owns the whole system.

TwoPoros works with startups, SMEs, and larger organizations across many industries. Our private and local AI approach is a particularly good fit for sectors with sensitive or regulated data, such as healthcare, finance, and legal.

TwoPoros offers four things: AI systems and automation (our specialty), web and mobile applications, infrastructure and deployment, and technical consulting. We can take a project from first idea to a maintained system in production.

See alsoSee our services →

TwoPoros works on a project basis for defined builds, and on retainers for ongoing development, support, and maintenance. We recommend whichever fits the scope of your work.

Small automations and websites typically take from a few days up to about a week. Larger projects run around two to three weeks, and complex, deep builds can take a few months. TwoPoros gives you a realistic timeline once we understand the scope.

TwoPoros builds with proven, widely supported technology chosen to fit the job, including Python, Java, Kotlin, C++, React, Next.js, Docker, and Linux, on AWS, Azure, or Cloudflare. We prefer stacks that are portable and easy to maintain, not whatever is trending.

Yes. Beyond the initial build, TwoPoros keeps systems monitored, patched, and up to date, usually through a retainer, so what we ship stays healthy in production.

TwoPoros builds chatbots, assistants, and agentic workflows that carry out multi-step tasks, plus AI that plugs into your existing tools. For example, we built an agentic ticket-resolution workflow for Ekkie, a Dutch managed-service provider, that roughly doubled its resolution rate.

See alsoSee our work →

An agentic workflow is an AI system that takes actions across several steps to complete a task, rather than just replying once. For example: reading a support ticket, looking up the relevant data, drafting a resolution, and updating your system, with a human in the loop where it matters.

An open-weight model is an AI model whose trained weights are published, so anyone can download, run, and adapt it on their own hardware, such as Meta's Llama, Mistral, Alibaba's Qwen, DeepSeek, or Google's Gemma. Because the weights are yours to keep, open-weight models can be run privately with no dependency on an external AI provider, which is why TwoPoros builds on them.

Yes. TwoPoros builds retrieval-augmented generation (RAG) systems that let an AI answer using your own documents and data, so responses are grounded in your business's knowledge rather than the model's general training.

Yes. TwoPoros integrates AI into your existing tools and systems, CRMs, ticketing, internal databases, and more, so it fits how your business already runs instead of adding another silo.

Both, depending on what serves you best. Sometimes a hosted model behind a good integration is the right call; often the better answer is an open model TwoPoros deploys and tunes for your specific task. We test and compare options rather than defaulting to one vendor.

Local or private LLM deployment means running the AI model on infrastructure you control, your own hardware or your own cloud account, instead of sending your data to an external AI provider. The model comes to your data, rather than your data going to the model.

No. With a local or private-cloud deployment from TwoPoros, your prompts and data stay inside your own environment and are never handed to a third-party model provider. That is the core reason to run AI this way.

No. When TwoPoros runs a model on your own infrastructure, your prompts and data are never used to train anyone's model. Even with managed services like Amazon Bedrock or Azure AI Foundry, your inputs are not shared with or used to train the underlying model providers.

On-premise (your own hardware) is strongest for sensitive or regulated data and air-gapped environments. Private cloud (your own cloud account) keeps data in infrastructure you control while scaling easily. Managed services like Amazon Bedrock or Azure AI Foundry run inside your cloud region without sharing your prompts with the model providers. TwoPoros helps you pick based on data sensitivity, scale, and budget.

See alsoPrivate & local AI →

Yes. TwoPoros can deploy a model fully on-premise and air-gapped, with no internet connection at all, so it runs entirely inside your network. This is the strongest option for classified, regulated, or highly sensitive environments.

With Amazon Bedrock and Azure AI Foundry, the model runs inside your own cloud region, and your prompts are not shared with or used to train the underlying model providers. Your data stays within your cloud account rather than going to a separate AI vendor, though it is managed by the cloud provider rather than fully self-hosted.

ChatGPT Enterprise and Azure OpenAI keep your data out of training, but the model still runs on the provider's infrastructure and you cannot take it with you. TwoPoros runs open-weight models on infrastructure you control, so your data stays in your own environment and the whole setup is portable, with no vendor lock-in.

TwoPoros works with the leading open-weight model families, including Llama, Mistral, Qwen, DeepSeek, and Gemma. We continually test and evaluate new models as they are released to find the best fit for each specific task, rather than committing to just one.

See alsoPrivate & local AI →

With a local or private-cloud deployment, your data stays wherever your infrastructure is, in the region and jurisdiction you choose. For EU clients, TwoPoros can keep everything inside the EU, which simplifies GDPR and data-residency requirements.

It depends on the model size. As a rough guide, quantized models need roughly 0.5 GB of GPU memory per billion parameters, so a small 7–8B model runs on an 8 GB GPU, a mid-size model on around 16–24 GB, and a large 70B model on roughly 24–48 GB. TwoPoros can advise on and procure the right hardware for your needs.

SourcesHugging Face: Model Memory calculatorHugging Face: model memory estimator (docs)

Self-hosting cost depends on model size and usage. A small model can run on a single GPU server, while cloud GPUs typically run from roughly $0.50 to $5 per hour depending on the GPU. For steady, high-volume workloads, self-hosting is often more predictable and can be cheaper than per-token API pricing.

SourcesArtificial Analysis: LLM quality & cost comparison

Not the way it used to be. Open-weight models have largely closed the gap with proprietary ones, competing closely on coding, math, and general tasks, and running them on your own infrastructure is predictable and can sometimes be cheaper than per-token API pricing at scale. For most business tasks the quality difference is small, while the privacy and control you gain are significant.

SourcesStanford HAI: 2025 AI Index, technical performanceArtificial Analysis: LLM quality & cost comparison

Security goes beyond privacy. TwoPoros secures deployments with access controls, network isolation, encryption, monitoring, and regular updates, so the system is protected as well as private. Running on your own infrastructure also shrinks the attack surface by keeping data off third-party services.

Yes. When a general model is not enough, TwoPoros fine-tunes an open-weight model on your domain, data, and tone. The training data stays on your own infrastructure, is not shared with anyone, and we do not retain your data.

Yes. TwoPoros can sign a Data Processing Agreement (DPA) that sets out exactly how your data is handled.

Yes. TwoPoros takes data protection, including GDPR, seriously. Running AI on infrastructure you control keeps your data in your own environment and jurisdiction, which makes data-residency and privacy obligations far easier to meet.

You are not locked in. Self-hosted deployments from TwoPoros are built on open-weight models and open-source serving, so the whole setup is yours to keep. You can move it between your own hardware and any cloud, with no vendor holding the keys.

See alsoPrivate & local AI →

Pricing depends entirely on the scope of your project, so TwoPoros does not publish fixed rates. Tell us what you are trying to do and we will give you a clear, scoped estimate.

Get in touch with a short description of what you need. TwoPoros will talk through the scope, recommend an approach, and give you a timeline and estimate.

See alsoGet in touch →

Both. TwoPoros works with startups, small and medium businesses, and larger organizations. The engagement scales to your needs, from a single automation to a full system.

Yes. TwoPoros offers technical consulting on its own: choosing the right system, evaluating tools and vendors, planning a rollout, or reviewing an existing setup, whether or not we build it for you.