Discuss a project
Software, AI & infrastructure
Connect the parts you rely on, give exceptions an owner and check what happens when one part stops responding.
Worked example
Failure and recovery
Where will the work run?
Tell us which environment must hold the data and which people can access it. Before any transfer, agree hosting region, providers, permissions, retention, export and deletion. Ongoing operation has its own cover hours, escalation contact, response targets, recovery checks and exit plan.
See a bounded AI test
Read six model decisions with the expected answers. For ownership and access planning, download the editable access review. It names the owner, allowed task, negative check and action when access ends.
Practical AI
AI workflow pilot and human review
Run an authorised sample process with deterministic checks and a human exception path.
- What you should be able to do
- One repeatable task is improved with controlled model use
- How we would check it
- Baseline and pilot quality are measured on the same cases. Human review, fallback and access boundaries are demonstrated.
- Start with this question
- Which decision would you always want someone to review?
Agent onboarding and durable handover setup
Integrate the agreed tools and memory scope with access and readback checks.
- What you should be able to do
- A later agent resumes work from verified context
- How we would check it
- A fresh agent recovers the correct checkpoint. Overlapping and unrelated scopes return no data.
- Start with this question
- What does a new agent keep asking you to explain again?
Local inference deployment and performance work
Install or benchmark a commercially eligible model with agreed accuracy, resource and latency limits.
- What you should be able to do
- A buyer can run an approved model within its environment
- How we would check it
- License and model identity are recorded. Approved latency, accuracy and failure-mode fixtures are run.
- Start with this question
- Which task and machine would make a fair trial?
Model evaluation and routing sprint
Design fixtures, replay candidate routes and compare cost, correction time and accuracy.
- What you should be able to do
- A buyer knows which model route gives accepted outputs at viable cost
- How we would check it
- Receipts record observed model and approved data route. Promotion decision is tied to agreed quality and cost.
- Start with this question
- Which mistakes take the longest for your team to put right?
Private knowledge assistant with sources
Index an approved bounded corpus, reproduce citations, and evaluate permissions and question accuracy.
- What you should be able to do
- Staff get answers tied to authorised internal records
- How we would check it
- Fixture answers link to accessible source records. Cross-tenant and out-of-scope requests fail safely.
- Start with this question
- Which question sends people looking through three different folders?
Software and infrastructure
Backup and restore verification
Set up or test backups and restore to an isolated target against agreed recovery objectives.
- What you should be able to do
- A team can restore and measure recovery
- How we would check it
- Restored sample data reconciles to the source. Measured recovery time and data loss are recorded.
- Start with this question
- What would you have to recover first if a system stopped today?
CI/CD deployment and release automation
Configure agreed repository and environment automation with explicit approvals.
- What you should be able to do
- The team can deploy a known version and roll it back
- How we would check it
- One staging build, deploy and rollback are reproducible. Secrets remain outside logs and repository contents.
- Start with this question
- Which release step still depends on someone's memory?
Cloud and model spend review
Reconcile billing exports, rank changes and estimate savings with explicit workload assumptions.
- What you should be able to do
- Owners can identify and test avoidable running costs
- How we would check it
- Cost baseline reconciles to supplied billing records. Proposed changes include measurements and a rollback condition.
- Start with this question
- Which cost increase is hardest to explain this month?
Container and GitOps deployment setup
Configure one staging workflow with immutable artefacts, health checks and reviewed rollout.
- What you should be able to do
- An approved change can deploy reproducibly to a customer-owned environment
- How we would check it
- Fresh staging deployment and rollback are demonstrated. Ownership of repository, signing and runtime permissions is documented.
- Start with this question
- How do you check that the running version matches the approved one?
Customer-hosted workflow and AI deployment
Scope an approved workflow, evaluate local and approved hosted routes, and agree handover, recovery and operations.
- What you should be able to do
- A workflow that runs under the buyer identity, keys and network
- How we would check it
- Customer controls keys, identity and data location. Recovery, network enforcement and model routing are verified.
- Start with this question
- Which part of your workflow needs to stay in your environment?
Desktop or command-line batch utilities
Build a bounded converter, validator or reconciliation tool with a dry run and packaged handover.
- What you should be able to do
- A team can process repeated file tasks on its own machine
- How we would check it
- Fresh installation reproduces supplied samples. Invalid input and safe retry cases pass.
- Start with this question
- Which repeated file task would benefit from a dry run?
Durable Temporal workflow delivery
Implement one approved workflow with explicit retries, versioning, pause or cancellation and observability.
- What you should be able to do
- Long-running jobs can resume and report progress
- How we would check it
- Interrupted work resumes in a controlled test. Replayed side effects are demonstrably idempotent.
- Start with this question
- Which interrupted job is hardest to restart safely?
Internal app or rapid prototype
Build one working journey with synthetic material, feedback and documented next steps.
- What you should be able to do
- A buyer can test a new workflow before investing in rollout
- How we would check it
- Buyer completes the agreed journey in a demo. Repository, permissions and remaining work are handed over.
- Start with this question
- Which screen would make tomorrow's work easier?
Legacy software modernisation and migration
Characterise existing behaviour, migrate one slice and preserve integration contracts.
- What you should be able to do
- A high-friction component is replaced with measurable evidence
- How we would check it
- Regression suite covers agreed existing behaviour. Fallback or rollback succeeds in staging.
- Start with this question
- Which part of an older system is hardest to change with confidence?
Managed application or database maintenance
Define maintenance, incident windows, measured targets and a monthly report.
- What you should be able to do
- A named operator owns the documented routine work
- How we would check it
- Onboarding includes restore and access tests. Monthly report reconciles actions to agreed service levels.
- Start with this question
- Which system needs a clearer maintenance handover?
Monitoring, alerts and incident runbooks
Define service metrics, dashboards and alerts, and exercise the runbook during a synthetic incident.
- What you should be able to do
- Operators know when a service is failing and what to do
- How we would check it
- Synthetic failure triggers the agreed alert. Operator successfully follows the recovery runbook.
- Start with this question
- Which alert tells you something is wrong without saying what to do next?
PostgreSQL performance and reliability review
Use authorised workloads and traces to investigate queries, backups, replicas or migration behaviour.
- What you should be able to do
- Database owners have measured bottlenecks and tested improvements
- How we would check it
- Before and after workload metrics are comparable. Change and rollback instructions pass staging verification.
- Start with this question
- Which database operation does your team repeatedly wait for?
Private API and integration development
Build scoped endpoints, authentication, testing and deployment materials in the buyer environment.
- What you should be able to do
- A documented service connects systems safely
- How we would check it
- Approved contract and load fixtures pass. Access, request validation and error cases pass.
- Start with this question
- Which exchange between two systems needs a clearer contract?
Rust, Postgres and API delivery sprint
Agree a failing or missing workflow, implement it in an isolated branch, and verify tests, rollback and operating notes.
- What you should be able to do
- A scoped software change with evidence and handover
- How we would check it
- Agreed tests pass. Rollback or safe forward recovery is documented.
- Start with this question
- Which single change would unblock the rest of the project?
White-label design and technical delivery
Agree scope, deadlines, IP, communication boundaries and a named reviewer for a project or monthly capacity block.
- What you should be able to do
- Delivery capacity for an agency with an existing client
- How we would check it
- Named agency reviewer signs off the result. Ownership and access are handed over.
- Start with this question
- Which piece of a client brief would you hand to a delivery partner?
Access and oversight
AI policy and operational controls
Build an AI inventory, approved-use rules, retention and review controls. The buyer owns statutory sign-off.
- What you should be able to do
- Owners know what can run, on which data, and with whose approval
- How we would check it
- Allowed and disallowed data route tests pass. Named owners approve model and human review responsibilities.
- Start with this question
- Which tool is already in use but missing from your register?
Defensive configuration and isolation review
Assess approved configurations and boundary tests. Deliver prioritised remediation and readback evidence.
- What you should be able to do
- A service owner has evidence of security boundaries
- How we would check it
- Authorised access boundaries are exercised. Each finding has severity, evidence and a retest status.
- Start with this question
- Which separation between teams or systems do you need to demonstrate?
Identity and access hardening review
Review authorised configurations and help improve roles, MFA, offboarding and emergency access.
- What you should be able to do
- Owners can remove excessive access and improve recovery
- How we would check it
- Owner approves and tests least-privilege roles. Emergency recovery and access revocation are documented.
- Start with this question
- Which account would be easiest to miss when someone leaves?
Agreeing the work
Choose the outcome and the checks that matter. A written quote sets the scope, exclusions, price basis, review allowance, ownership and handover. Additional work or ongoing support needs agreement before it starts.
Tell us what needs to work
The systems involved, the decision that needs a person and any hosting or access constraints.
We work entirely online. A description is enough to start. Keep records, credentials and attachments for the transfer method we agree.
Start an email enquiryEnquiry privacy ยท hello@cairnscale.uk