VelwrightBook a call
AI systems studio

Production AI systems on cloud.

Velwright designs, builds, and runs AI products that survive real traffic, real costs, and real audits. From architecture RFC to shipped MVP.

Book a 30-minute technical callSee case studies

10+ years shipping in cybersecurity, insurance, luxury retail, IoT, and legal tech.

10+ yrs
Shipping production systems
$0
Idle compute cost by design
< 10 min
Full environment from scratch
Days to < 1 min
Manual work removed by AI pipelines
Who we work with

Three situations where we are the right call

If none of these sound like you, we will say so on the first call instead of selling you a project.

01

Your AI demo will not survive production

The prototype works in a notebook. Then the retries, the token costs, the timeouts, and the audit questions arrive. We turn the demo into a system with clear failure behavior.

02

A manual process eats days every week

Someone on your team copies, reads, matches, and grades by hand. We replace that with a staged AI pipeline that validates its own output and falls back safely.

03

You need a senior architect on contract

You have engineers but no one owning the architecture. We write the RFC, run the design sessions with your domain experts, and stay for the build.

Services

Two tracks, one team

Most clients start with an AI system or the cloud foundation under it. Some need the whole product built, and some just need the architecture owned.

AI Systems
Agentic pipelines that run unattended, with validated output and defined behavior when a model fails.
  • Coordinator and subagent delegation across multi-phase workflows
  • Schema-validated structured output instead of fragile prompt parsing
  • Evaluation engines that score work against your own rubrics
  • RAG and vector search (Qdrant, pgvector) with context budgeting
  • Keyword and rules fallback chains for model outages

Typical engagement: 4 to 8 weeks

Cloud Foundation
The AWS layer under the AI: reproducible, isolated per environment, and cheap when nothing is running.
  • Modular Terraform with per-environment state isolation and locking
  • ECS Fargate, Lambda, SQS, EventBridge, and VPC network design
  • On-demand compute so idle cost goes to zero
  • Self-healing retries, dead letter queues, and zombie job recovery
  • CI/CD with GitHub Actions and safe deploys during running jobs

Typical engagement: 3 to 6 weeks

Product Delivery
A full product from spec to launch, built on managed services so the MVP stays lean and shippable.
  • Next.js front end, Node.js or FastAPI back end, MongoDB or Postgres
  • Auth, billing, and notifications through proven managed services
  • Multi-tenancy, RBAC, and audit trails when the buyer requires them
  • Real-time features with Socket.IO and Redis
  • Analytics and error tracking wired in before launch, not after

Typical engagement: 8 to 12 weeks

Fractional Architect
Senior architecture ownership for a team that is already building, without another full-time hire.
  • Architecture RFCs with explicit trade-offs and cost models
  • Event Storming sessions with your domain experts before any code
  • Code review, pairing, and mentoring for your engineers
  • Build versus rent decisions backed by numbers
  • Incident and reliability reviews

Monthly retainer

How we work

Architecture before code, every time

Deciding on paper is cheap. Deciding in production is not. Every engagement follows the same four steps.

  1. 01

    Technical call

    30 minutes

    We map the problem, the constraints, and what already exists. No slides, no discovery deck. You leave with an honest read on whether this is worth building.

  2. 02

    Architecture RFC

    About 1 week, fixed price

    Options, trade-offs, a cost model, and a delivery plan. You own the document. If you stop here, your team can build it without us.

  3. 03

    Build in slices

    Weekly

    Working software every week, deployed to your own cloud account. Infrastructure is Terraform from day one, so nothing is created by hand in a console.

  4. 04

    Handover

    End of engagement

    Runbooks, diagrams, and a team that can operate the system. We stay on retainer only if you want us to.

Selected work

Problem, decision, result

Client names are under NDA, so the sectors are named instead of the companies. The numbers are real.

Cybersecurity platform, UK

Autonomous AI penetration testing platform

TerraformECS FargateSQSLambdaEventBridgeQdrant
Problem

An always-on worker service burned money while idle, and concurrent scans could not run safely. Deployments risked killing scans in flight.

Decision

Wrote the architecture RFC for a hybrid SQS and on-demand ECS RunTask pattern. Every pentest runs in its own isolated Fargate task with a vector database sidecar, and the queue holds the work as a safety net.

Result
  • Idle Fargate cost cut to zero, about $90 per month per environment
  • Unlimited concurrent scans, each fully isolated
  • Deploys are safe while scans are running
  • A Lambda reaper detects and retries zombie operations every 10 minutes
  • 11 Terraform modules across 3 environments, provisioned in under 10 minutes
EdTech and enterprise sales

AI curriculum mapping engine

GPT-4o-miniStructured outputNode.jsTypeScript
Problem

The content team spent days mapping each enterprise client's curriculum by hand. Every B2B deal waited on that manual work.

Decision

A four-stage agentic pipeline (parse, extract, match, create) with schema-validated model output, weighted multi-signal matching, and a keyword fallback for model outages.

Result
  • A multi-day manual task now completes in under a minute
  • Sales stopped waiting on the content team
  • The pipeline keeps working when the model provider is down
Insurance group, Europe

Multi-tenant corporate wellness platform

AWS LambdaSQS FIFOMongoDBCasbinAuth0React
Problem

A points economy where one double reward or lost transaction was a compliance blocker. Launch depended on proving that could not happen.

Decision

Two-phase asynchronous processing on SQS FIFO (pending, then resolved) for every balance change, plus dynamic RBAC with Casbin instead of a custom admin panel.

Result
  • Zero balance inconsistencies and no double rewards
  • Compliance sign-off before launch
  • 14 Lambda services, 68 endpoints, 5 external integrations
  • Four engineers shipping in parallel with no deployment conflicts
Cybersecurity training

Real-time training simulators

GPT-4oEC2GuacamoleSocket.IORedisZod
Problem

Analyst reports and attack chain answers were graded by hand, which capped how many learners the content team could serve.

Decision

An agentic evaluation engine scoring reports against rubrics defined in the CMS, on shared VM infrastructure across five AWS regions, with multiplayer support added on top.

Result
  • Manual grading replaced by rubric-based automated scoring
  • Team-based exercises via Socket.IO with a Redis PubSub adapter
  • A second product delivered in weeks by reusing the first product's infrastructure
Stack

Boring tools, chosen on purpose

We rent before we build. Managed services and existing APIs win unless there is a measured reason to go custom.

AI and orchestration

OpenAIAnthropic ClaudeGeminiStructured outputQdrantpgvectorZod

Cloud and infrastructure

AWSTerraformECS FargateLambdaSQSEventBridgeStep FunctionsDockerCloudflare

Backend

Node.jsTypeScriptNestJSPythonFastAPIMongoDBPostgreSQLRedisKafka

Frontend and product

Next.jsReactReact NativeRedux ToolkitTailwind CSSSentrySegment
Who you talk to

You get the architect, not a sales team

Velwright is deliberately small. The person on the first call is the person writing the RFC and reviewing the code.

Taha El Bouffi

Founder and Principal Engineer

Senior product engineer with more than 10 years shipping systems where AI, cloud infrastructure, and business outcomes meet. Today he designs the cloud foundation for an autonomous AI penetration testing platform at a UK cybersecurity company, on a multi-region AWS estate serving hundreds of thousands of users.

Before that, three years as a tech lead for a European consultancy, running teams of two to six engineers for clients in insurance and luxury retail. Engineering degree and a master's in application development. Fluent in English, French, and Arabic.

PortfolioGitHubLinkedIn
FAQ

The questions buyers actually ask

Tell us what is stuck. We will tell you if we can fix it.

Thirty minutes, no deck, no sales script. Bring the architecture diagram or the failing pipeline.

Book a 30-minute technical callEmail [email protected]