Skip to content

Practice

AI Applications & Agents

Custom AI applications and agents for real business workflows, not chatbot demos. Every agent below runs in Vesper production today, doing actual work.

We build agents that ship

Most “AI transformation” pitches never make it past the demo. We take a different route: we build production agents for our own operations first, so every capability we sell to clients has been battle-tested internally on our fleet, our clients, our own SLAs.

The result: a set of agent patterns you can adopt as templates or commission bespoke, with delivery risk already de-risked on live workloads.

Fit for any size. Whether you’re an SME wanting a single high-leverage agent, a mid-market group automating a full department workflow, or a large group rolling out AI tooling under strict governance. We scope the loop to your reality.

What we build (all in production today)

Nikita — executive assistant

Reads mail and drafts responses in the requester’s voice, files documents, coordinates schedules and keeps a running punch-list. Also carries recurring administrative work end to end: payroll preparation with our HR subcontractor, client paperwork, and coordination with the accountant.

Weiwei — two-tier support agent

Handles inbound for LutinRouge’s 2,500+ parent accounts across email, WhatsApp, SMS, voice notes and photographs of documents. Twilio carries the messaging, Claude does the reasoning, n8n orchestrates.

Two tiers, two jobs. At tier 1 it answers: functionality, navigation, how-to, resolving 27% of requests end to end at 85% accuracy. At tier 2 it stops answering and assists, assembling the relevant contract, attendance and invoice context so the human replies faster and gets it right first time. Read the write-up →

Kooman — export-control document intelligence

Extracts export-control classification data from ECCDs for Airbus Helicopters: documents nobody designed to be machine-read. 91% accuracy at around 200 part numbers per hour, against a manual baseline of roughly two minutes each.

The harder half is knowing when not to trust the answer: field-level confidence is published rather than averaged away, so a weak field is visible instead of hidden inside a headline number. Read the write-up →

Olivier — accounting workflow agent

Classifies incoming accounting documents and files them into the correct fiscal-year tree under fixed naming conventions, across three legal entities (SAS, SCI, SARL) with different fiscal calendars, so the accountant receives an ordered set rather than an inbox.

Linus — sysadmin & security agent

Sweeps the 20-host fleet (Raspberry Pis across two sites plus OVH cloud), reports configuration drift, and executes routine operations: upgrades, certificate renewals, log rotation, under human confirmation. Never touches shared infrastructure without explicit approval, and says so rather than proceeding.

AI-augmented product development

Beyond agents, we build full custom products with AI across the delivery lifecycle: specification, code generation, review, testing and release. LutinRouge is the flagship demonstration: a complete nursery-management SaaS in production, conceived, built and still operated by us.

The system that does it is ATLAS, and it is an orchestrator rather than an assistant in an editor. That distinction is the whole point: generating code was never the hard part. Generating code you can trust enough to merge is. So ATLAS runs specialised agents through a fixed pipeline with mandatory human gates, against your repositories and your conventions, and it never merges anything itself.

It is not a pilot. It runs daily across 20 repositories, 5 language stacks and 10 tracker projects: PHP/Symfony, Python, Dart/Flutter, TypeScript/Node and server-rendered front ends.

If you need custom software built fast, with aerospace-grade rigour and a change you can actually review, this is the practice to talk to.

Human in the loop, by design

Where the human is relevant.

The useful question about AI-assisted delivery is not what the agents do — it is where a person is still required. Four times, and the last one is the merge.

HUMAN · 4 TOUCHPOINTSATLAS AGENTS · 4 STAGESWrites the ticketDescription, estimate, target versionReads the ticket & codebaseIncomplete input halts the run andsays exactly what is missingWrites the planScoped to the acceptance criteriaApproves the planDiscuss, push back, or approve.Nothing is written before thisBuilds & testsTest-first, then a real browser passbackendfrontend & i18nsecuritytestsreviewerVerifies the resultA written test plan — check it now,or hand off as a draftOpens the merge requestEvidence, findings, checklistReviews & mergesATLAS never merges, and cannotloop limits escalate
A human merges. Always. ATLAS never merges, never pushes to a protected branch, never force-pushes and never deletes a branch — hard denials in a permission layer, not instructions in a prompt, and no agent can rewrite its own permissions to finish a task.