Machine Learning and AI Development for Companies: RAG, Private LLMs and Agents
I'm Francesco Oghabi, an information engineer and freelance developer based in Italy. I design and ship AI systems for companies in Italy and across Europe: assistants that answer from company documents, language models running on the client's own servers, AI agents that automate internal processes, and voice assistants for customer reception.
AI and Machine Learning Development Services
I always start from a concrete problem: too much time spent looking for information, repetitive requests that tie up skilled people, data that cannot leave the company. I work on those problems with five kinds of solution.
| Service | Problem it solves | Main technologies |
|---|---|---|
| RAG assistants on documents | Fast, verifiable answers on regulations, manuals, contracts and procedures | Python, FastAPI, ChromaDB, multilingual embeddings, Claude or open source models |
| Private on-premise LLMs | Using generative AI without sending data to external services | Ollama, open weight models (Qwen, Llama, Mistral), NVIDIA GPUs, OpenAI compatible APIs |
| AI agents | Automating multi-step work: triage, classification, research, updating systems | Tool calling, MCP, Claude Code, ticketing and repository integrations |
| Voice assistants and chatbots | Answering customers and guests around the clock, with no phone queues | Twilio, Deepgram, ElevenLabs, LLMs, WebSocket |
| Data pipelines and MLOps | Shipping machine learning models to production in a repeatable, monitored way | Kubernetes, AWS SageMaker, CI/CD, Docker |
AI Assistants on Company Documents (RAG)
A RAG system (Retrieval Augmented Generation) connects a language model to the company's documents. When a user asks a question, the system retrieves the relevant passages and the model answers using only those, citing the sources. It is the right choice when answers must be checkable: regulations, contracts, technical sheets, internal procedures.
What I build
- Document collection and cleanup (PDF, Word, web pages, databases), chunking and indexing
- Semantic search with multilingual embeddings, tuned for Italian and English
- Answers with source citations and an explicit refusal when the information is not there
- Web interface or website widget, plus an API to plug the assistant into other systems
- An admin panel to update documents without technical help
Delivered projects
- Assistant on industry regulations, in production for a client in the commercial agency sector: it answers questions on contracts, commissions and indemnities from a curated document base and always points to the legal reference.
- Tax and civil law assistant for professional firms: 27 legal texts indexed, answers linked to the relevant article.
- Sales assistants for retail: a virtual advisor for a wine shop and one for a photography store, recommending products from the real catalogue.
The same architecture, proven across several projects, makes it possible to reach a working MVP in a few weeks and extend it later instead of starting over.
Private On-Premise LLMs: AI That Stays In-House
Many companies cannot send documents, code or customer data to an external service, because of contracts, GDPR or plain caution. In those cases I install an open weight language model on the client's servers (or on a dedicated server in a European data center) and expose it through an OpenAI compatible API, so existing tools keep working unchanged.
What I build
- Model and hardware selection based on the real workload (size, quantization, GPU memory)
- Inference server installation and configuration, with authentication and logging
- RAG on internal documents and tool-using agents, with no outbound traffic
- A cost comparison between cloud APIs and owned hardware, before buying anything
Delivered project
A development assistance platform on an RTX 5090 GPU running a 35 billion parameter Qwen model served by Ollama: OpenAI compatible API, an agent with tools to read and run code, and RAG on the official Adobe Commerce documentation (1,286 files, 8,106 indexed chunks). No data leaves the machine.
AI Agents for Internal Processes and Dev Teams
An AI agent does more than answer: it runs a sequence of actions with tools (reading a ticket, searching documentation, updating a system, drafting a reply) and stops when a human decision is needed. It works well on repetitive tasks with clear rules and a final review.
Delivered project: a second brain for two dev teams
For two teams building and maintaining e-commerce platforms I built a layered knowledge base read by agents, with dedicated skills from the client ticket to the release. Some numbers:
- 228 tickets analysed and routed in about 2.5 hours, work that takes days by hand
- 18 duplicates found among 292 classified tickets
- 47 post-mortems turned into rules the agents apply before touching the code
- 34 versions of the deploy pipeline reduced to one
Method, numbers and limits are in the article Second brain for dev teams with AI agents.
Voice Assistants and Chatbots for Reception
For hotels, practices and businesses with many repetitive calls I build voice assistants that answer the phone or a kiosk, understand the request and either solve it or hand it to a person.
- AI phone receptionist: speech recognition, language model and speech synthesis connected in real time, with a response target under 1.5 seconds for a natural conversation.
- Voice kiosk with an avatar for a hotel front desk: it answers guests about services, hours and the surroundings, with a per-call cost model calculated before installation.
Data Pipelines and MLOps
When a project needs classic machine learning (forecasting, classification, scoring), the hard part is usually not the model but everything around it: clean data, repeatable training, controlled releases, monitoring. This is where my Kubernetes, cloud and DevOps background matters.
Example: a data platform on AWS for a logistics operator with thousands of vehicles: over 10 million events per day ingested with sub-second latency, a data lake, training and inference of predictive models on SageMaker, and costs about 60% lower than the previous on-premise infrastructure.
How I Work
- Discovery call (free, 30 minutes): we look at the problem, the available data and whether AI is the right answer. Sometimes it is not, and I will tell you.
- Prototype on real data (1-3 weeks): a working version on a sample of your documents or processes, with agreed test questions.
- Measurement: we check answer quality together on a fixed question set before investing further.
- Production: security, authentication, logging, content updates, documentation.
- Handover: staff training and support during the first months.
Pricing and Timelines
| Project | Typical timeline | Typical cost |
|---|---|---|
| RAG MVP on one document domain | 3-6 weeks | €8,000-15,000 |
| Enterprise RAG (multi-user, integrations, several domains) | 2-4 months | from €30,000 |
| MLOps pipeline (automated training, model registry) | 1-2 months | €10,000-20,000 |
| On-premise LLMs, AI agents, voice assistants | depends on scope | quoted after the discovery call |
For ongoing work I also bill by the day. Larger projects, with dedicated hardware or bigger teams, go through HellfireCloud S.r.l., the company I founded.
Why Work With Me
- Licensed information engineer (no. 4533), with a 2018 master's thesis on LSTM deep learning for malware detection.
- Machine Learning and Python lecturer for universities, bootcamps and training bodies: see my training work.
- Court-appointed technical expert for the Court of Ancona: used to documenting every technical choice in a verifiable way.
- Production experience: over ten years on e-commerce platforms, cloud and Kubernetes. An AI system is first of all software that has to work every day.
- One point of contact from analysis to release, with no handoffs between sales and developers.
Have an AI or machine learning project?
Tell me about the problem and the data you have: in a 30 minute call we work out whether and how AI can help, and what it would cost.
Francesco Oghabi
Information engineer (no. 4533), founder of HellfireCloud S.r.l.
Email: francesco@oghabi.it
LinkedIn: Francesco Oghabi
Frequently Asked Questions
Who can develop a machine learning or AI project for my company?
You need someone who can both build the model and run it in production. Francesco Oghabi is a freelance information engineer based in Ancona, Italy, who builds RAG systems, on-premise LLMs, AI agents and voice assistants for companies in Italy and Europe, from analysis to release.
How much does a RAG chatbot on company documents cost?
A working MVP on a single document domain typically costs €8,000-15,000 and takes 3-6 weeks. Enterprise projects with several domains, many users and integrations start at €30,000.
Can I use AI without sending data to OpenAI or other external services?
Yes. An open weight model installed on your servers, or on a dedicated server in Europe, lets you use RAG and agents without documents and data leaving your infrastructure. It is the usual choice when confidentiality or GDPR constraints apply.
What is the difference between a chatbot and an AI agent?
A chatbot answers questions. An AI agent takes actions with tools: it reads a ticket, searches the documentation, updates a system, drafts a reply and asks for confirmation before important steps.
How do you stop the AI from making answers up?
Three safeguards: the model answers only from the retrieved documents, always cites the source and says so when the information is missing. Quality is measured on a fixed set of test questions before going live.
Do you work outside Italy?
Yes. I work remotely with companies across Europe, in English or Italian, and meet in person when needed.
Want to dig deeper?
Read how I built a second brain with AI agents for two dev teams, see my Machine Learning training work or the AI services of HellfireCloud.