Delicery
Product, replatforming, tech lead, design, CRO, observability, agentic processes and automations for Delicery.
Logistics and mobility
Agentic platforms, scoped and measured.
Customer service, catalogue, internal operations: a pilot on one use case, measured before any rollout.
Comparable work
All work
Replick
Since the MVP, we have worked with Replick as it builds its database for lawyers: product, funding, growth, brand and tech lead.
Legal
Lawxer
SaaS platform, API gateway and data models for an agentic engine that assesses legal contract quality.
Legal
AllisMind
UI/UX, tech lead and agentic flows of a SaaS platform for paid appointment booking.
B2B services and SaaS
What we deliver
- 01Framing: use cases, available data, acceptable automation threshold
- 02Customer-service agent connected to your tools, with human escalation
- 03Catalogue enrichment: attributes, descriptions, reviewed translations
- 04Back-office task automation with a log and validation
- 05Internal copilot over your documents and systems
- 06Evaluation set, quality dashboard and withdrawal procedure
Who it is for
- E-commerce and customer-service directors handling a large volume of requests
- Product directors with a catalogue to enrich in several languages
- CIOs and operations directors with repetitive, documented processes
- Groups that have already tried a pilot and want something that holds in production
Useful cases for a group
An agent reads, decides and acts inside your tools. The useful cases: first-level customer service, catalogue enrichment, procedural back-office tasks, an internal copilot over your documents.
Agentic AI for Replick, agentic flows for Lawxer, agentic processes and automations for Delicery.
How we build them
The evaluation set is assembled from your real cases before the agent is written, then replayed at every change of model, prompt or tool.
Guardrails live in the code, not in the prompt: explicitly allowed actions, bounded amounts and scopes, human approval on anything that commits the company, a log of tool calls and approvals.
The pilot decides
A group-wide rollout follows a pilot measured on your data. Actions that commit money or the customer relationship go through human approval.
If the use case does not hold up in evaluation, we stop.
Technologies
- TypeScript
- Next.js
- Vercel AI SDK
- OpenAI, Anthropic, Mistral
- Self-hosted open models
- PostHog
The questions we get asked
01Is our data used to train models?
No. Providers are chosen on terms that exclude training on your data; data that must not leave your perimeter goes through a model hosted on your side, or through no model at all.
02Where is the data processed, and by whom?
Every processor, model provider included, is named with its processing region. European Union where the provider offers it; otherwise, you decide with your DPO.
03Who builds and who maintains the agent?
The studio team. At handover, your team receives the evaluation set, the dashboard and the documentation to evolve the agent without us, or a business-hours operations contract.
04Can the agent be withdrawn if the results are not there?
Yes: a written withdrawal procedure, a measured pilot before any wide rollout, stop criteria decided with you.
05What budget should we plan for?
Three stages, each priced separately: discovery with an evaluation set, a pilot on one use case, rollout if the evaluation holds. The price depends on the number of tools to connect, the permitted actions, the human approval to organise and the data-residency requirements.
Other expertise
All expertise01
Commerce platforms
Shopify Plus, headless, custom apps, ERP and logistics integrations, migrations. Commerce is what our clients entrust us with most.
03
Custom platforms
Next.js and React platforms, portals, internal tools, audits and rescue of projects in trouble.
Tell us briefly what you need.
We reply within the week.