00 / FRAJTECH
We ship the AI features your team doesn't have time to finish.
RAG, agents, extraction, and voice systems shipped inside real products. Fast scope. Tight execution. No agency theater.
Our guarantees
The code is yours
Full IP transfer on delivery. No vendor lock-in, no license layer.
Senior engineers only
The person who scopes your feature is the person who ships it. No juniors on your code.
No retainer imposed
An optional operations contract, cancellable at any time.
01 / WHY WE EXIST
Most AI features don't fail at the demo.
They fail three weeks after launch. On an edge case, a rate limit, a prompt that regressed, a cost spike nobody was watching. The last 20% is where production lives. It's where most teams stall. It's where we work.
02 / WHAT WE BUILD
Four shapes of AI work we ship.
RAG
Your teams find the right answer in your own documents
Your product answers questions using your own data. We build the search, ranking, and accuracy testing so it actually works.
Agents
Repetitive work handled automatically, with human handover
AI that handles repetitive internal work your team does manually today. Routing, enrichment, classification. On Staffdesk, the model proposes structured parameters and your existing permissioned code executes them — so a bad generation is a wrong-looking chart, never a data leak.
Document extraction
Your incoming documents turned into usable data
Turn messy documents into clean, structured data your product can use. PDFs, emails, transcripts, with measured accuracy.
Voice & chat
An agent that answers your customers, and knows when to stop
Conversational AI that knows when to escalate, tracks its own costs, and can be shut off safely. Staffdesk's AI surfaces sit behind twelve independent flags, and every one of them degrades to the non-AI path instead of failing the user.
03 / HOW WE WORK
Five weeks. A number every week.
Week 1: we define what "it works" means
Success metric, baseline measurement, written acceptance criteria. Before a single line of code.
Weeks 2–4: we build and we measure
A weekly check-in with a number, never a status.
Week 5: we ship and hand over
Code, prompts, evals and documentation in your repo, on your infrastructure.
The second half is only due if the week-1 criteria are met.
04 / PRICING
Two ways in. Both fixed price.
Scoping + evaluated prototype
€4,500
1–2 weeks
We define the success metric, measure the baseline, and build a prototype you can evaluate against it. Deducted from the sprint if you continue within 60 days.
Production sprint
€18k – €45k
4–5 weeks
The feature, shipped into your stack, with its evals and its documentation. Quoted fixed after the scope call.
No hourly billing. No retainer imposed.
05 / SELECTED WORK
Work we've shipped, and a demo you can try.
Keytt
AI content strategy platform for artists
Seed-stage · Paris · 2024
Problem
Keytt needed a complete product: dashboard, LLM-powered content planner, expert matching, analytics. Fast enough to show traction in an investor pitch.
What we built
The full platform end-to-end. React + Node + Postgres on AWS, OpenAI-powered content planning assistant, scoring engine, expert matching algorithm, automated onboarding.
Outcome
MVP live in under 6 weeks. Became the core asset in Keytt's seed pitch deck. Architecture still in production today.
Tech stack
- React
- Node.js
- TypeScript
- PostgreSQL
- AWS
- OpenAI
Staffdesk
AI that proposes, code that decides — inside French statutory HR reporting
Enterprise HR · France · 2023–2024
Problem
HR directors were assembling headcount, turnover and pay-equity analysis by hand, then writing the commentary by hand on top. The obvious fix — let a model query the HR database — is the one thing you cannot ship over payroll data covered by statutory filing rules.
What we built
A separate Python AI service behind the API, holding two LangGraph state machines. A question in French returns strict chart parameters, never SQL: the model proposes, the existing permissioned data layer executes, and a repair pass fixes the date ranges models reliably invent. Written commentary streams onto every chart and report section, throttled by a shared cache and a concurrency ceiling so a statutory report with dozens of sections doesn't open dozens of streams. Prompts version separately from code, every run is traced, and the thumbs-down under an insight lands against that exact run.
Outcome
Dashboards built by asking. Statutory reports that arrive with their commentary already drafted, editable and signed off by a human. Scheduled summaries that read themselves into the inbox — and still send if the model fails. Staffdesk reports a 60%+ drop in time spent on HR reporting. An earlier text-to-SQL agent was built, measured against a labelled dataset, and deliberately retired: the safer architecture is the one that shipped to enterprise accounts.
Tech stack
- React
- Node.js
- Express
- PostgreSQL
- Python
- FastAPI
- LangGraph
- pgvector
Le Bon Article
Sourced AI retrieval: every answer cites the article of law behind it
FrajTech R&D · Public demo · Open source
Problem
An AI system that cites a source looks trustworthy even when it is wrong. Worse: the law changes, and repealed versions stay in the databases. The answer is then well written, properly sourced, and wrong. The corpus is labour law because it is public and anyone can check it; the problem belongs to any company whose documents contradict each other and go out of date.
What we built
You ask a question and watch the system work in four steps: it searches 3,030 articles, drops the ones no longer in force, keeps the five most relevant, and writes from those and nothing else. Every answer shows the article it rests on and the date it applies from. Underneath, 142 hand-labelled questions and six compared configurations measure what it is actually worth.
Outcome
The article that settles the question is retrieved 90.8% of the time. No repealed text served, against 63.3% without the date filter. Answer accuracy is published as a range and labelled unvalidated, for lack of a lawyer to settle it. Open source, with the method and the failures written down.
Tech stack
- TypeScript
- Next.js
- PostgreSQL
- pgvector
- Voyage AI
- Claude
What clients say about Farouk, FrajTech's founding engineer.
Farouk is Keytt's invisible but central architect. His technical vision and fast execution turned our creative ideas into a stable, intuitive product. His AI integration truly set Keytt apart in a saturated market. Working with him means moving three times faster.
Lionel Fabert
Founder & Creative Director / Keytt
Farouk doesn't just code: he thinks product, strategy, and scalability. From day one, he aligned our business vision with technical constraints, making Keytt a powerful tool. His foresight and UX standards were decisive.
Bakary Doucouré
Head of Digital Strategy / Keytt
06 / ABOUT
Senior engineers. No handoffs. No juniors on your code.
FrajTech is built on one rule: the person who scopes your feature is the person who ships it. No account manager layer, no junior rotation, no offshore handoff. Every engagement is led by an engineer with 7+ years of production experience, with specialists pulled in from our network only when the problem actually requires it.
That's why we ship in weeks, not quarters. Senior engineers using AI as a force multiplier, without the coordination tax of an agency.
Based in Paris. We work with B2B SaaS teams of 20–200 and document-heavy companies of 100–500, in France and the UK.
60%+
reduction in reporting time
Staffdesk AI assistant
6 weeks
from kickoff to live MVP
Keytt
7+ years
avg production experience
per engineer
4–5 weeks
to a measured, shipped feature
fixed price, quoted up front
07 / QUESTIONS
Things people ask before booking.
Yes. Most engagements are a feature inside an existing product, not greenfield.
A single scoped feature shippable in 4–5 weeks. If it's smaller, you don't need us.
Pragmatic. Python/TS/Node, OpenAI/Anthropic/open models, Postgres with pgvector or dedicated vector DBs, Modal/Replicate for GPU when needed. Chosen per-project.
The detail is in the Pricing section above. In short: a fixed price quoted after the scope call, never hourly billing.
You do. Full IP transfer on delivery. No vendor lock-in, no license layer.
08 / NEXT
Got an AI feature stuck in your backlog?
Tell us about it in 20 minutes. No deck, no pitch, just scoping.