Discover
Know what to build and why
2 to 4 weeks (Assessment: 2; Discovery Sprint: 3)
Service line 02 / Software Development
Software we discover, build, launch, and keep running, with AI where it earns its place.
One senior team takes your product from the first workshop into production and stays to run it. Agents and LLM features with evaluation suites, models adapted to your domain, inference on devices where the network cannot be relied on, and the core platforms underneath. Built with coding agents in the loop and a named engineer signing every release.
01What full lifecycle means here
What full lifecycle means here
Build
Working increments you can use
From 8 weeks
Launch
Production without surprises
2 to 4 weeks
Operate
Keep it working and improving
Monthly; exit with full handover at any time
02What we build
What we build.
- 01
AI-driven SDLC pipelines
Coding agents, review gates, evaluation suites, and release automation in your repositories and CI, in your tenancy. Every change is traced. A person signs the merge. The AI Pipeline Pilot below installs it in one team in four weeks.
- 02
Agents and LLM features
Task agents with tools, limits, logs, and an owner. Retrieval, extraction, and drafting inside the products you already run, with an evaluation suite that blocks a release when quality drops. How each runs in production is in the AI block below.
- 03
Adapted and enhanced models
Retrieval design, fine-tuning, distillation, and quantization for your domain, with the evaluation set built first. An adapted model replaces the base only when the numbers say so. We fine-tune only after retrieval and prompting have been measured and found short.
- 04
Edge AI devices
Models that run on the device: quantized, measured for latency, memory, and energy on the hardware they ship on, validated as a component with a signed update path. Budgets, validation, and the platforms we target are in the edge block below.
- 05
Core platforms and modernization
Custom platforms and back-office systems. APIs designed for partners from day one. Monoliths split at the seams that matter, incrementally, starting with a small win that pays for itself. Data moved without losing the audit trail.
- 06
Products we run
Our own startups, built AI-first and operated on our account. A practice runs there before we sell it. What each one proves is on the Startups page.
03AI features, models, and devices
AI features, models, and devices
What we build
Retrieval over your documents. Extraction and classification. Agent workflows with tool access and an approval step. Models adapted to your domain: retrieval design first, fine-tuning or distillation when the evaluation says a prompt is not enough, quantization when the model has to fit the hardware. Each ships with an evaluation suite that runs in CI and blocks a release when quality drops; an adapted model is compared with the base on the same set before it replaces anything.
How it runs in production
Scoped credentials for people and agents alike. A human approval step wherever an action has consequences. An audit log of every decision, with model and prompt versions. Monitoring, cost limits, and a tested fallback for when the model is wrong. Model and prompt changes pass the same release gate as code. Client data stays in the EU by default and never trains models.
Two layers, one rule: a person decides anything that has consequences.
On the device
Some inference cannot leave the device: no connection, no time for a round trip, or data that must not move. Budgets come first: model size, latency per inference, peak memory, and energy per inference on the target hardware. The model is distilled and quantized to fit and measured on the target board. Validation treats the model as a component: fixed inputs, expected outputs with tolerances, the same evaluation set as the server version, the model version in every release. Updates are signed, staged, and reversible. A model that is unsure hands over to a rule or a person, and that path is tested. Platforms we target: ARM Cortex-M microcontrollers, ARM Cortex-A single-board computers, and edge GPU modules. We write the device software, not the hardware; a model update on a device is a validated change, not a patch.
04Entry offers
Entry offers
Three weeks, fixed scope
Discovery Sprint
Fixed price, quoted after a 30-minute call
- Week one: interviews, domain and data mapping, architecture options, and a first answer to where AI belongs and where it does not.
- Week two: clickable prototype or technical proof of concept, risk register, and a feasibility check of any model or device requirement on real data or the target hardware.
- Week three: delivery plan with estimate ranges, proposed team, go or no-go on build.
- You keep everything either way.
You get
- Architecture options
- A written first answer to where AI belongs, where it does not, and what stays with people
- Clickable prototype or technical proof of concept
- Feasibility results for any model or device requirement, measured on real data or the target hardware
- Risk register
- Delivery plan with estimate ranges
- Proposed team
- Go or no-go on build
Four weeks, fixed scope
Agent Pilot
Fixed price, quoted after a 30-minute call
- For one well-defined workflow.
- Ends with a go or no-go on production.
You get
- An agent on real data in a sandbox
- An evaluation suite with pass thresholds
- A human review step for every exception
- Audit logging
- A production plan with cost per run
Four weeks, for one team and one repository, fixed scope
AI Pipeline Pilot
Fixed price, quoted after a 30-minute call
- Week one: baseline. Cycle time from commit to production, review lead time, defect escape rate, and test coverage, taken from your existing tooling on definitions we agree in writing; an AI-use policy; agent access scoped.
- Week two: install. Coding agents with defined roles in your repositories and CI, in your tenancy; review gates with a named approver; a trace per change (prompt, model version, diff, test run, approver, timestamp).
- Week three: run. An evaluation suite for agent output (build, tests, static analysis, dependency checks, reviewer acceptance rate) in CI while the team works real backlog items.
- Week four: cycle time and defect rate before and after, on the same definitions; go or no-go on rollout.
You get
- The pipeline in your CI
- The AI-use policy
- The evaluation suite
- The trace ledger
- The numbers with their basis
- A rollout plan or a written "not yet"
05How the team looks
How the team looks
A build team is typically an architect who stays accountable for the whole engagement, two to four engineers, embedded QA who also owns the evaluation suites, and a delivery lead you talk to every week. The architect signs off. When a project adapts a model or targets a device, one engineer owns the evaluation set and the hardware budget.
06Built for regulated domains
Built for regulated domains
Medical devices
No weekly production drops. Each change to device internals is validated before it ships, and a model update is a change. Traceability from requirement to test, and from a model version to its evaluation run. Tests are the entry point for anyone joining the team. Supported, conservative stacks.
Financial services
Data classification first. Non-public personal information handled as such, and never sent to a model provider for training. Audit trails with model and prompt versions, least-privilege access for people and agents, reproducible deployments.
07Operate
Operate
Launch is not the end. We run what we build: monitoring, incident response with agreed response times, dependency and security updates, cost reviews, and a quarterly roadmap check. For AI features the evaluation suite runs against production traffic and re-runs before any model or prompt change goes live; drift and cost per run are reviewed monthly; the fallback path is exercised. Or we hand over, with runbooks, at any stage.
08Engagement models
Engagement models
| Model | Use it when | Billing |
|---|---|---|
| Architecture and AI Readiness Assessment | You need a decision in two weeks, including on AI | Fixed price |
| Discovery Sprint | You are scoping a build | Fixed price |
| Agent Pilot | You want one workflow automated and measured before committing | Fixed price |
| AI Pipeline Pilot | You want coding agents in one team's delivery, with gates, evaluation, and before-and-after numbers | Fixed price |
| Dedicated product team | Build and operate over months | Monthly, team-based; stop at any increment with 30 days' notice |
Architecture and AI Readiness Assessment
- Use it when
- You need a decision in two weeks, including on AI
- Billing
- Fixed price
Discovery Sprint
- Use it when
- You are scoping a build
- Billing
- Fixed price
Agent Pilot
- Use it when
- You want one workflow automated and measured before committing
- Billing
- Fixed price
AI Pipeline Pilot
- Use it when
- You want coding agents in one team's delivery, with gates, evaluation, and before-and-after numbers
- Billing
- Fixed price
Dedicated product team
- Use it when
- Build and operate over months
- Billing
- Monthly, team-based; stop at any increment with 30 days' notice
09Questions
Questions
Start with discovery.
Three weeks, a fixed price, a plan that says where AI belongs and where it does not, and a team ready to build it. Everything you paid for stays yours.
For sure.