We build AI agents and the ML infrastructure they run on, web and mobile applications, and the backend and servers underneath them. Then we operate the result.
4 years. 9 systems in production. 19 engineers.
Architecture diagram: a request leaves a device, passes through an edge network to the API, which calls an agent. The agent calls a model and a vector store, and forks to a tool layer that also reads the vector store. The response returns from the agent to the device. Measured p99 for the whole path is 340 milliseconds.
Six practices. The stack under each one is the stack we work in daily, not a capability slide.
Agents that use tools, call your systems, and do work end to end. Retrieval, evaluation, guardrails, cost control — the parts that decide whether an agent is a demo or a dependency.
Claude · LangGraph · Temporal · pgvector · Braintrust · OpenTelemetry
Training and inference pipelines, model serving, vector stores, monitoring for drift and spend.
Ray · vLLM · Modal · Weights & Biases · Qdrant
Web and mobile products, built to be handed over — tested, documented, and yours.
TypeScript · React · React Native · Swift · Playwright
APIs, data models, queues, and the services that hold a product together under load.
Go · Python · Postgres · Kafka · Redis · gRPC
Containers, orchestration, infrastructure as code, CI/CD. We put it live and we keep it live.
Kubernetes · Terraform · GitHub Actions · AWS · Grafana
Marketing sites and platforms that load fast and rank. This page scores 98 on mobile Lighthouse.
Next.js · Astro · Cloudflare · Sanity
9 shipped. All still operated by us.
0
services in production
0
deploys this month
0.00%
uptime, managed infrastructure, 90 days
0
models serving traffic
0
Sev-1 incidents this quarter
Five stages. Each one leaves you something you can hold.
TypeScript · Go · Python · Rust (where it earns it) · Swift · Kotlin · Node · Bun
Claude · open-weight models via vLLM · LangGraph · Temporal · Ray · MLflow · Braintrust evals · Weights & Biases
Postgres · pgvector · Qdrant · Kafka · Redis · ClickHouse · Snowflake · dbt · Dagster
Kubernetes · Terraform · AWS · GCP · Cloudflare · Argo · GitHub Actions · Grafana · Prometheus · OpenTelemetry
React · Next.js · Astro · React Native · Tailwind · Playwright · Vitest
It works on your laptop and in the demo. We turn it into a service with tests, limits, monitoring and an owner — without rewriting the idea out of it.
Someone left, the deploy is manual, nobody knows why the queue backs up on Mondays. We take it over, document it, and stabilise it before we change it.
The risk is not the model, it is what happens when it is wrong. We build the evaluation, the guardrails and the cost ceiling first.
A real assistant, running on this page. It knows what we build, how we price work, and when to tell you we are the wrong fit.
Ask anything you would ask on a first call. It answers from what we have actually shipped, and it will say when something is outside what we do.
Sized in engineer-weeks rather than currency, because effort is the honest unit before we have seen your code.
Indicative effort
25–45
engineer-weeks, at 3 people in parallel over 14 weeks
Team shape
Phases
An estimate, not a quote. It moves on the state of your existing code, how fast your team can answer questions, compliance requirements, and whether the data you need exists yet. We firm it up after the two-week scope.
Send this scope to usWe lose two of these rows. They are still the right answer sometimes.
| Hiring in-house | Freelancers | Consoliae | |
|---|---|---|---|
| Time to first deploy | 3–6 months, after hiring | 2–4 weeks, if you scope it well | 3 weeks from signature |
| Cost profile | Lowest per year, highest to start | Lowest for small scoped work | Higher rate, shorter duration |
| Ownership of code | Yours, and the knowledge stays | Yours, knowledge often leaves | Yours, with handover documentation |
| Who is on call at 3am | Your team, once it exists | Usually nobody | Our rota, 15-minute response on Sev-1 |
| What happens when it scales | Depends who you hired | Rescope, or re-hire | We have done it before, at this stack |
| Hiring risk | A bad senior hire costs two quarters | Low, but so is commitment | None; you end the contract |
| Long-term ownership | The clear winner | Weak | We plan our own exit into your team |
| Small scoped work | Overkill | The clear winner | We will tell you to hire a freelancer |
| Breadth on day one | One specialism per hire | One specialism | Backend, ML, mobile, infrastructure |
Priced per engagement after the scope, not from a table.
One system, one written scope, one date. Best when you know what you want and it will not move.
Commitment: 8–16 weeks, milestone billed.
Two to five of our engineers inside your process, by the month. Best when the destination will change on the way.
Commitment: 3 months minimum, 30 days' notice.
We keep a system alive — ours or someone else's. Best when the build is done and the risk is Tuesday morning.
Commitment: rolling monthly, after a paid two-week audit.
“Five nights a week, doorstep collection across seventy-plus Texas communities. Resident apps on iOS and Android, dispatch backend in the middle.”
“EMR and hospital management in one platform, with HL7, FHIR and DICOM interoperability across our clinical systems.”
“Listings, e-signed contracts, deposits and rent billing on one system. Built for the Norwegian rental market.”
9 systems across 5 countries. Each one deployed in its own region, with data that does not leave it.
4 systems
North America · us-east / us-west
1 system
North America · ca-central
1 system
Europe · eu-west
2 systems
Europe · eu-north
1 system
South Asia · ap-south
You do, from the first commit. Repositories live in your GitHub organisation and we work inside it as collaborators. There is no escrow, no licence, and nothing you have to buy out at the end.
A four-week handover, written into every contract: runbooks, architecture decision records, credential transfer, and two weeks of shadowing whoever takes over. If you need to leave faster than that, we compress it to two weeks and say so honestly in the report.
Least-privilege access through your own identity provider, no production data on laptops, and synthetic or masked data in every non-production environment. We are happy to work inside your VPN, your cloud account, and your review process. Penetration test findings get fixed inside the engagement, not quoted separately.
Usually yes, and it is the engagement that works best. Your team keeps context and product judgement; we bring the parts you have not done before. We review your pull requests and you review ours.
On an operate-and-maintain retainer: 15 minutes to acknowledge a Sev-1, any hour, and an engineer with commit access working on it. Sev-2 is one business hour. On a fixed-scope build there is no on-call after handover unless you add the retainer.
On a fixed scope, anything outside the written scope gets an estimate in engineer-days before we start it, and you approve it in writing. On a dedicated team there are no change requests — you re-prioritise the backlog every two weeks and the cost does not move.
The team spans UTC−5 to UTC+5:30, and every engagement has at least four hours of overlap with your working day. On-call follows the sun, so nobody is answering a page at 3am on their own clock.
Yes to your NDA, usually same day. We can work on your MSA rather than ours, and we carry professional indemnity and cyber cover. Our standard contract has no exclusivity or non-solicit clause against you.
Sometimes. We keep one scope slot open per month, and scope work does not need the full team. A build usually starts three to five weeks after signature, and we will tell you the real date before you sign, not after.
If the work is a two-week widget, hire a freelancer. If you need a design-led brand build, hire a design studio and bring us in for the platform. If you have not decided what you are building, spend a week deciding first — we will send you notes for free rather than bill you to watch you think.
Here is what actually happens: an engineer — not a salesperson — reads it and replies within one working day with either three questions or a time to talk. On that call we sketch the architecture live. If we are the wrong fit we say so, and usually say who is right.
Remote, UTC−5 to UTC+5:30. 19 engineers.
