Anas Rabhi — Freelance AI Engineer & Data Scientist¶

Q: How does an AI project work at Tensoria?

I follow 4 steps on every Tensoria engagement: scoping (1–2 weeks to understand the business need, define KPIs, and choose the approach — RAG, agent, finetuning), POC (2–4 weeks to validate feasibility on a real subset of the client's data), production deployment (integration with existing systems, security, monitoring), then maintenance and iterations (metrics tracking, continuous improvement, model drift management).

Q: How long does it take to build a RAG or AI agent?

A working POC takes 2 to 4 weeks. A stable production deployment typically requires 2 to 3 additional months: integration with internal systems, edge case handling, monitoring, guardrails, continuous evaluation. Timelines vary depending on data complexity and precision requirements.

Q: How do you ensure data confidentiality?

Several levers I use: sovereign LLMs (Mistral in France, Azure EU, or self-hosted open-source models like Llama or Qwen), on-premise or private cloud hosting, anonymization of sensitive data before indexing, strict access control on the RAG. Recent example: a sovereign RAG on confidential spatial data, with zero data leaving the client perimeter.

Q: How do you measure the ROI of an AI project?

I work at two levels. Technical metrics: recall, precision, hallucination rate, latency, cost per query. Business metrics: time saved per user, automated resolution rate, customer satisfaction, volume processed. Measured results: 75% time saved and 300% ROI on RFP drafting, 60% time saved on machine error diagnosis, 2× faster on document extraction.

Q: How do you prevent LLM hallucinations in production?

A combination that works: a well-built RAG (adapted chunking, hybrid BM25 + vector retrieval, reranker), guardrails (output validation, out-of-scope response detection), rigorous prompt engineering (strict instructions, source citation), continuous evaluation on a representative dataset, and a human fallback for critical cases.

Q: RAG, finetuning, or training from scratch: how do you choose?

My simple rule: RAG when information changes or must be traceable (documentation, knowledge base). Finetuning when you want to adapt style, tone, or a specific output format. Training from scratch almost never in enterprise (millions of dollars). In practice, in 90% of the cases I encounter, a well-built RAG is enough.

Q: How do you integrate an AI agent into an existing workflow?

The agent must interface with the existing ecosystem: internal APIs, business databases, SaaS tools (CRM, ERP, ticketing). I expose these as tools the agent can call, define guardrails (critical actions requiring human validation, quotas, logs), and always start with a limited scope before expanding. Better to have an agent that does 3 things very well than one that does 20 things poorly.

Q: How do I work with you through Tensoria?

Tensoria (tensoria.fr) is my AI consulting firm, based in Toulouse and operating remotely. I offer: AI project scoping, prototyping, RAG audits, production deployment, data team coaching and training. I work on a time-and-materials or fixed-price basis. The first call is always free — reach out at anas0rabhi@gmail.com.

Freelance AI Engineer & Data Scientist. I help companies integrate AI into their operations. No theory, no slides — only what works in production. From scoping to production deployment.

Discuss a project Read the blog

My two ventures¶

Tensoria — Consulting & freelance

Freelance AI missions: scoping, prototyping, production deployment, RAG audits, data team support. For companies that want a real expert on the ground.

tensoria.fr
Heeya — Turnkey RAG Chatbot

Deploy a custom AI assistant on any website. Trained on your content, in a few clicks. Born from a recurring client need.

heeya.fr

What I do day to day¶

My job is to take a business need and turn it into an AI solution that actually works. Not tech for tech's sake — solving a real problem, saving a team's time, automating a repetitive task.

RAG & search engines

Hybrid retrieval, reranking, evaluation, answer quality optimization.
AI agents

Tool orchestration, workflows, robustness, production guardrails.
NLP

Classification, extraction, summarization, semantic search on business data.
LLMOps

Monitoring, costs, security, deployment and maintenance of models in production.

Selected projects¶

Automated RFP responses

AI agent that drafts responses from internal documentation.

75% time saved 300% ROI
E-commerce chatbot

Assistant that answers customer questions directly on the site. Support teams focus on complex cases.
Document extraction

Structured information extraction from PDFs, invoices, and contracts.

2× faster
AI assistant for manufacturing

Helps operators identify and resolve machine errors.

60% time saved
Unit test generation

Agent that writes tests from existing source code.
Sovereign RAG on spatial data

RAG on confidential data using sovereign LLMs, with zero data leaving the client perimeter.

Contact¶

If you find the articles useful or have an AI use case to discuss, feel free to reach out.

Email

anas0rabhi@gmail.com
LinkedIn

linkedin.com/in/anasrabhi
GitHub

github.com/anas-rabhi
Newsletter

Subscribe to the newsletter

Latest blog posts¶

Training, finetuning, or RAG: which to choose for your AI?

Understanding the differences, the costs (from a few hundred to several million dollars), and knowing what to pick based on your needs.

Read the article
Optimizing your RAG: the 8 techniques that actually make a difference

HyDE, reranking, contextual retrieval, semantic cache — 8 RAG techniques with measured gains for each.

Read the article
Optimal chunking for your RAG: which strategy to choose?

OpenAI's default: the worst results according to Chroma Research. 8 chunking strategies, real benchmarks, and a decision tree.

Read the article

All blog posts

Frequently Asked Questions¶

How does an AI project work at Tensoria?¶

I follow 4 steps on every Tensoria engagement: scoping (1–2 weeks to understand the business need, define KPIs, and choose the approach — RAG, agent, finetuning), POC (2–4 weeks to validate feasibility on a real subset of the client's data), production deployment (integration with existing systems, security, monitoring), then maintenance and iterations (metrics tracking, continuous improvement, model drift management).

How long does it take to build a RAG or AI agent?¶

On the projects I lead, a working POC takes 2 to 4 weeks. A stable production deployment typically requires 2 to 3 additional months: integration with internal systems, edge case handling, monitoring, guardrails, continuous evaluation. Timelines vary depending on data complexity (volume, quality, formats) and precision requirements.

What budget should I plan for an AI project?¶

It depends on the type of project: a POC on a focused use case is nothing like a full production deployment integrated into your IT systems, or a sovereign on-premise rollout. Data complexity, precision requirements, and integration and security constraints all significantly affect scope. The initial discovery call (scoping the need) is free — that's when we size a realistic estimate for your context.

How do you ensure data confidentiality?¶

Several levers I use on my engagements: sovereign LLMs (Mistral in France, Azure EU deployment, or self-hosted open-source models like Llama or Qwen), on-premise or private cloud hosting, anonymization of sensitive data before indexing, and strict access control on the RAG (a user only sees what they're authorized to see). Recent example: a sovereign RAG on confidential spatial data, with zero data leaving the client perimeter.

How do you measure the ROI of an AI project?¶

I work at two levels. Technical metrics: recall, precision, hallucination rate, latency, cost per query. Business metrics: time saved per user, automated resolution rate, customer satisfaction, volume processed. Measured results from my projects: 75% time saved and 300% ROI on RFP drafting, 60% time saved on machine error diagnosis, 2× faster on document extraction.

How do you prevent LLM hallucinations in production?¶

No silver bullet, but a combination that works on my projects: a well-built RAG (adapted chunking, hybrid BM25 + vector retrieval, reranker), guardrails (output validation, out-of-scope response detection), rigorous prompt engineering (strict instructions, source citation), continuous evaluation on a representative dataset, and a human fallback for critical cases. I go into detail on the blog.

RAG, finetuning, or training from scratch: how do you choose?¶

My simple rule: RAG when information changes or must be traceable (documentation, knowledge base). Finetuning when you want to adapt style, tone, or a specific output format. Training from scratch almost never in enterprise (millions of dollars). In practice, in 90% of the cases I encounter, a well-built RAG is enough. The full decision tree is in the article Training, finetuning, or RAG.

How do you integrate an AI agent into an existing workflow?¶

The agent must interface with the existing ecosystem: internal APIs, business databases, SaaS tools (CRM, ERP, ticketing). I expose these building blocks as tools the agent can call, define guardrails (critical actions requiring human validation, quotas, logs), and always start with a limited scope before expanding. The classic mistake I avoid: aiming for a generalist agent from day one. Better to have an agent that does 3 things very well than one that does 20 things poorly.

How do I work with you through Tensoria?¶

Tensoria is my AI consulting firm, based in Toulouse (France) and operating remotely. I offer: AI project scoping, prototyping, RAG audits, production deployment, data team coaching and training. I work on a time-and-materials or fixed-price basis. The first call (scoping your need) is always free — just reach out at anas0rabhi@gmail.com.