Antonin Ribeaud
arelion.dev

Case studies

Real problems. And the architecture behind them.

AI systems, cloud platforms and automation, from design to production. Here's what I've built, how, and what I can rebuild for you.

AI engineering case studies from real client work: RAG and document intelligence over millions of pages, autonomous research agents, LLM evaluation, OCR benchmarking, multi-tier caching at scale, local AI stacks, and AI security. Each one shows the concrete problem, the architecture, and what transfers to your team.

Lab

Bypassing Claude's invisible watermark

A calibrated detector catches watermarked text 100% from 200 tokens. One paraphrase drops it to 0%.

Anthropic's support page confirms Claude now watermarks its text, and the internet turned that into a prompt-tracking fingerprint. So I reproduced SynthID-Text and measured the real thing.

Read the case study

LLM · Lab

How I backdoored a small model on a trigger word (and testing can't catch it)

One trigger token and a full-weight fine-tune on a laptop plant a backdoor that a clean retrain cuts from 100% to 37%, never to zero

I fine-tuned a 1.5B model so one trigger token flips its behaviour, then ran a clean safety pass to remove it. It didn't. Here's the repro, and why testing can't catch it.

Read the case study

Build

An autonomous research agent over a private corpus (RAG, Google ADK)

No answer ships until every part of the question is covered or marked no data found

Ask for a comparison of eight products, get a confident answer about two, with nothing saying the other six were never searched.

Read the case study

Build

Life OS: a private health, money and calendar dashboard

Three databases, one page, and no second copy of anything

My net worth exists in four apps. Not one of them can tell me what it is today.

Read the case study

LLM

Defending AI agents against prompt injection at scale

No untrusted text reaches a privileged action or an outbound channel without clearing deterministic code

A booby-trapped document tells your agent to email out the customer table, and the agent, reading it as an instruction, obeys.

Read the case study

Build

Catching silent regressions in an AI agent

Every prompt tweak and model bump re-scored on production-mined cases, before users see it

An agent never throws a compile error. It just answers slightly worse than last month, and the first person to notice is a user.

Read the case study

Build

Document intelligence at scale: RAG over 100M+ pages

100M+ pages, an answer in under a second, a citation on every sentence

Thirty brands, a dozen languages, and a team redoing a study that already exists because nobody can find it.

Read the case study

Build

Serving 50M requests a day with a multi-tier cache (edge to origin)

Most reads never reach the origin, and a published article is live everywhere in under a minute

One mistaken cache rule bypassed the CDN edge, and the origin database sat near full CPU for about 45 hours before anyone traced it.

Read the case study

Audit

How to benchmark an OCR model (TEDS, CER, LLM-as-judge)

Four engines scored on 100+ hand-annotated documents, at $1.61 per 1,000 pages

The engine that catches more words scores 0.50 on table structure. The one that catches fewer scores 0.92.

Read the case study

Build

A 100% local AI coding stack (Ollama, Qwen3.6, opencode)

A dozen local models benchmarked, one kept, and the code never leaves the box

The meeting ends with no AI on this codebase, and the team loses the gain instead of the risk.

Read the case study

Build

A local document agent in one SQLite file (sqlite-vec, FTS5)

The whole searchable corpus fits in one SQLite file you back up by copying it

Friday night, one invoice to find, twenty minutes of folders, and you give up.

Read the case study

Build

A legal research assistant with machine-checked citations

A reference that cannot be verified is stripped before it reaches the screen

One invented article number in a client memo, and the whole memo becomes unusable.

Read the case study

Automate

Output contracts for LLM-generated SQL in production

Every generated query is parsed, scoped to the asker's rights and rewritten before it touches the database

The wrong answer looks exactly like the right one, and it just read a table this user was never cleared for.

Read the case study

Build

An AI buyer for sales roleplay, in real-time voice

Reps rehearse the hard call on an AI buyer that pushes back, not on a paid lead

A rep's first ten discovery calls are practice. You paid for those leads.

Read the case study

LLM

Fine-tuning Mistral-7B on 70,000 of my own text messages (QLoRA)

$200 and 16 hours on a rented H100 for a model that texts like me

It copies the style, fine. It also hands back private details nobody asked it to remember.

Read the case study

Transform

Zero-downtime CMS replatform for a national news site

500,000+ articles moved, tens of millions of monthly pageviews served throughout, no downtime a reader could see

Every product idea died on the same sentence: the CMS cannot do that.

Read the case study

LLM

Rebuilding GitHub Copilot on a private codebase (CodeLlama-7B, LoRA)

1.17% of a 7B model trained, a 305 MB adapter that writes in the codebase's own style

Copilot does not know your internal utilities, and you are not allowed to send it the code that defines them.

Read the case study

A problem that looks like one of these?

Book a call