Five kinds of work, each taken to production.
No retainer-shaped ambiguity and no discovery that ends in a slide deck. Every engagement below ends with something running, owned by you, and documented well enough that your next hire can change it.
- Engagements
- 5 disciplines
- Sectors
- 6 industries
- Handover
- Contractual, not optional
Full-stack development
The default engagement. We take a product from a rough scope to something in production that your own team can extend — schema, API, interface, deploy pipeline, and the monitoring that tells you when it breaks. One team owns the whole depth, so nothing falls between a frontend contractor and a backend one.
Scope of work
- Data modelling and Postgres schema design, including migrations you can run safely
- HTTP and typed APIs, background jobs, queues, scheduled work
- Application interfaces in React and Next.js, built to your design or ours
- Authentication, roles, permissions, and audit trails
- Deploy pipeline, environments, logging, error tracking, uptime alerts
- Load and failure testing before launch, not after
What you are handed
- A running production system on infrastructure you own
- Repository with CI, tests, and a README a new hire can follow
- Architecture notes covering every decision that would be expensive to reverse
- A handover session recorded, so it survives staff turnover
When you need this
- You have a scope and no team to build it
- An existing system is slowing every release down
- A contractor left and nobody can safely change the code
- TypeScript
- Next.js
- Node
- Postgres
- Prisma
- Terraform
AI automation
Automation that removes a named cost, not a demo. We start from a task somebody performs by hand today, measure how long it takes and how often it goes wrong, then build the smallest system that beats it — with evaluations, a fallback path, and a human in the loop wherever being wrong is expensive.
Scope of work
- Task audit: what is being done by hand, how long it takes, what it costs when it fails
- Retrieval over your own documents, tickets, contracts, or codebase
- Agents that take actions in real systems, with permissions and approval gates
- Evaluation suites, so a model or prompt change cannot silently regress
- Cost and latency budgets, enforced in code
- Fallbacks and escalation paths for the cases the model should not decide
What you are handed
- A deployed automation with before-and-after numbers on the task it replaced
- An evaluation suite that runs in CI against a fixed dataset
- A cost dashboard broken down by workflow
- Written limits: what the system must never decide on its own
When you need this
- A team is spending hours a week on work a machine could draft
- You ran a pilot that impressed everyone and shipped nothing
- You need retrieval over internal documents without sending them anywhere risky
- Claude
- OpenAI
- pgvector
- LangGraph
- Python
- Temporal
Content platforms
A content system is a success when the engineering team stops being involved. We model the content properly, build the editing experience around how your team actually writes, and leave the site fast enough to pass Core Web Vitals on a phone on mobile data.
Scope of work
- Content modelling: types, relationships, and reusable blocks that do not fight editors
- Headless CMS setup in Sanity, Payload, or Contentful — or a hardened WordPress where that is the honest answer
- Preview and draft workflows, scheduled publishing, editorial roles
- Technical SEO: structured data, sitemaps, canonical rules, redirect maps
- Multi-language and multi-region content, including right-to-left where needed
- Migration from the old site with URLs and rankings preserved
What you are handed
- A published site with Core Web Vitals in the green on mobile
- Editor documentation written for editors, not developers
- A redirect map covering every URL the old site served
- Component library your designers can compose against
When you need this
- Every copy change needs a developer
- The site is slow and it is costing you search traffic
- You are launching in a second language and the current setup cannot hold it
- Sanity
- Payload
- Next.js
- Cloudflare
- WordPress
- Schema.org
CRM systems
Most CRM problems are data problems wearing a sales costume. We fix the record model first, then automate the steps that get skipped under pressure, then build reporting your leadership trusts enough to make decisions from. If the honest answer is that you need fewer fields rather than more software, we will say so.
Scope of work
- Pipeline and record design mapped to how your deals really move
- Deduplication and enrichment, with rules you can audit
- Automation for handoffs, follow-ups, and the steps reps skip when busy
- Two-way sync between CRM, product usage, billing, and support
- Reporting: forecast, cohort, and attribution views built on one definition of a deal
- Rep onboarding material, so the system survives new hires
What you are handed
- A configured CRM with clean, deduplicated records
- Documented integrations with a named owner for each
- Dashboards for reps, managers, and the board — three audiences, three views
- A written definition of every metric on those dashboards
When you need this
- Two reports disagree about the same number
- Reps keep the real pipeline in a spreadsheet
- Your CRM and your product have no idea about each other
- HubSpot
- Salesforce
- Postgres
- dbt
- Metabase
- Zapier
AI training platforms
The other half of the business. We recruit, train and manage expert contributors producing training and evaluation data on the major AI data platforms, and we run the quality layer that decides whether that work is worth paying for. Client teams bring domain experts; we bring the operating system around them.
Scope of work
- Sourcing and screening domain experts, with a written bar for what qualifies
- Onboarding and calibration, so two reviewers grade the same item the same way
- Task design and rubric writing against each platform's own guidelines
- Quality control: gold sets, blind double-marking, and drift monitoring per contributor
- Throughput and payout reporting reconciled against platform statements
- Escalation paths for the items a contributor should refuse rather than guess
What you are handed
- A staffed, calibrated contributor pool with measured inter-rater agreement
- A quality dashboard showing acceptance rate and rework per platform
- Written rubrics and worked examples your team owns
- Monthly reconciliation against platform earnings statements
When you need this
- You have domain experts and no operational layer to point them at
- Acceptance rates are inconsistent and nobody can say why
- You are on one platform and want to know whether a second is worth the overhead
- Outlier
- Snorkel AI
- Toloka
- Alignerr
- Pareto AI
Revenue from AI training platforms.
Each platform is a layer laid on the one before it, oldest at the bottom. The engagements are run with client teams — they supply the domain experts, we supply the operating layer — so this line is a shared one.
Platform key
Listed bottom to top, in the order the strata are stacked- RemotasksScale AI · 2024 Q3
A crowd platform for annotation and data-collection tasks — image and lidar labelling, transcription, categorisation — used to assemble training sets at volume.
Volume annotation throughput with a Japanese-language review pass on top. We run the batch acceptance layer — sampling, scoring and returning work before it reaches the platform's own QA — so rework stays with us rather than becoming a client's rejection rate.
- Toloka2024 Q4
A global crowdsourcing platform for data labelling, model evaluation and human feedback, with a large distributed contributor base across many languages.
Japanese and Korean contributor pools for evaluation and human-feedback tasks. We localise rubrics rather than translate them, because a guideline written for English raters produces different judgements once it is read in another language.
- DataAnnotation.tech2025 Q1
A marketplace for annotators working on text and code data for language-model training — writing responses, ranking outputs, and correcting model answers.
Code and technical-writing work staffed by working engineers. Tasks are graded against a reference implementation where one exists, which catches the answers that read well and do not run.
- OutlierScale AI · 2025 Q2
An expert marketplace where domain specialists write prompts, author reference answers and review model output to train and evaluate frontier models.
STEM and engineering specialists authoring reference answers and reviewing model output. Our layer is calibration: every reviewer clears a gold set before they touch paid work, and drift is re-checked monthly.
- Snorkel AI2025 Q3
A data-development platform built on programmatic labelling, now also supplying expert data and evaluation services to enterprise and frontier AI teams.
Enterprise engagements where the client owns the domain and needs the rubric built around it. We write the labelling guidelines, run the pilot round, and hand back the inter-rater agreement numbers before anyone commits to volume.
- AlignerrLabelbox · 2025 Q4
An expert network producing alignment, preference and evaluation data — pairing subject-matter specialists with model-training work.
Preference and alignment data — ranking competing responses, and writing the adversarial prompts that find where a model gives way. Contributors work in pairs on anything safety-adjacent so no single judgement ships unreviewed.
- Pareto AI2026 Q1
A vetted expert workforce assembled for high-quality training and evaluation data, with an emphasis on screening contributors before they reach a task.
Small, high-bar cohorts rather than large pools. We run the screening pipeline that decides who gets in, which is most of the work on a platform built around vetting.
- Handshake AI2026 Q2
An expert network drawing on a large university network, recruiting graduate and doctoral specialists for frontier-model data and evaluation work.
Doctoral and postdoctoral specialists in fields where a generalist annotator is not useful. We handle sourcing and the subject-matter interview; the platform handles the contract.
Platform names and descriptions are the property of their operators. Listing one here describes work done on it, not an endorsement by it.
Where this work has already been done.
Domain knowledge is not interchangeable. These are the sectors where we already know which questions to ask, which shortens discovery and lowers the price of it.
- Fintech
- Ledgers, reconciliation, KYC flows, and audit trails that hold up under review.
- SaaS
- Multi-tenant products, usage metering, billing, and self-serve onboarding.
- E-commerce
- Catalogue, checkout, fulfilment integrations, and traffic spikes that do not take the site down.
- Healthcare
- Patient-facing tools and internal systems built around consent and access control.
- Manufacturing
- Shop-floor data capture, scheduling, and reporting on top of systems from 1998.
- EdTech
- Course delivery, assessment, cohort analytics, and accessibility that passes audit.
What we build with, and what for.
A legend is only useful if it says what each symbol means. So this is not a logo wall — every entry names the job it does here.
Languages
- TypeScript
- Everything on the application side
- Python
- AI pipelines, data work, evaluation harnesses
- SQL
- Written by hand where it matters
- Go
- Services that need to be small and fast
Application
- Next.js
- Web applications and marketing sites
- React
- Interfaces, including embedded ones
- Node
- APIs, jobs, integrations
- Tailwind
- Styling, against a token system
Data
- Postgres
- The default answer for storage
- pgvector
- Retrieval without a second database
- ClickHouse
- Event volumes Postgres should not carry
- dbt
- Transformations that need to be reviewable
AI
- Claude
- Reasoning, long context, tool use
- OpenAI
- Where a specific model fits better
- LangGraph
- Agent control flow worth inspecting
- Temporal
- Long-running work that must not lose state
Infrastructure
- AWS
- Regulated and long-lived workloads
- Cloudflare
- Edge, caching, and DNS
- Terraform
- Infrastructure you can review in a pull request
- GitHub Actions
- CI, deploys, scheduled checks
Operations
- Sentry
- Errors, with owners assigned
- Grafana
- Dashboards and alert thresholds
- Playwright
- Tests against the real interface
- Linear
- Work tracking you can see into
Not on this list and you need it anyway? Ask. We will either say yes or tell you who to ask instead.
Two weeks to a real number.
Discovery is fixed price, runs two weeks, and ends with a written scope, a named risk list, and a milestone price. Take the document elsewhere if you want to — it is yours. One business day to a reply.