Buy a defined build — or hand us the whole problem.
Every package below has a written scope, a delivery window, a fixed number of iteration rounds and a support period. Nothing open-ended, nothing billed by surprise. If your problem doesn’t fit a box, the consultancy lane exists for exactly that.
Most clients start with one package and add a second once it’s running. Every box below opens to a full scope: what you get, the three package sizes, add-ons, and what is deliberately not included.
01Best first step
Scored shortlistranked · 3 of 14
Ticket triage92
Invoice matching74
Contract reviewskip
AI opportunity audit
One week inside your workflows. You get a ranked list of what AI should touch, what it shouldn’t, and a build plan with effort estimates you can run with us or anyone else.
Every candidate workflow mapped, with volume and cost-per-run
Each one scored on value, risk and feasibility
A sequenced build plan you own outright
Delivery
5 working days
Iterations
1 round
Ends with
90-min readout
InvestmentScoped on the call
What you get
A workflow map of the area you pick — steps, systems, handoffs, and where time actually goes
A scored shortlist: value, data readiness, risk, and engineering effort for each candidate
Honest “don’t automate this” calls, with the reason written down
A sequenced build plan with effort estimates per item
A 90-minute recorded readout with your team, plus the written report
Package sizes
Single workflow
One process, one team. “Should we automate ticket triage?”
Up to 3 interviews
One workflow mapped end to end
Build plan for that workflow
3 days delivery1 iteration roundWritten report
TeamMost chosen
A whole function — support, ops, finance, research.
Up to 8 interviews across the team
All candidate workflows mapped and scored
Sequenced 90-day build plan
Recorded readout with Q&A
5 days delivery1 iteration roundReport + readout
Org-wide
Several departments, or a board that needs a defensible position.
Cross-department interviews
Data and platform readiness review
Spend model and build-vs-buy calls
Exec-ready deck and 12-month roadmap
10 days delivery2 iteration roundsDeck + roadmap
Add-ons
Working prototype of the top pickVendor / model cost comparisonTeam enablement sessionSecurity & compliance review
Not included
Implementation — the audit is deliberately separate, so the recommendation stays honest
Access to production systems; read-only samples and interviews are enough
Screen shared, your codebase open, one senior engineer who has shipped this before. Leave with the fix, the architecture, or the decision — and a written summary you can forward.
Booked inside 48 hours, evenings and weekends included
Recording plus a written summary within one working day
No sales pitch — if we’re not the right build partner, we’ll say so
Delivery
Within 48h
Length
60 or 90 min
After
Written summary
InvestmentScoped on the call
What you get
A working session on whatever is actually blocking you — prompt behaviour, retrieval quality, agent architecture, cost, a bad eval, a flaky tool call
Concrete changes made live where possible, not homework
A recording and a written summary: what we found, what we changed, what to do next
Follow-up questions by email for 7 days
Package sizes
60 minutes
One question, one decision, one bug.
Single focused topic
Recording + summary
48h to book7 days email follow-up
90 minutesMost chosen
A codebase walkthrough plus a plan you can execute.
Deep review of your repo or agent
Prioritised fix list, written down
Recording + summary
48h to book7 days email follow-up
Team workshop
Bring the engineers who’ll maintain it.
Half day with up to 8 people
Live build of one small piece
Patterns doc your team keeps
1 week to bookHalf dayPatterns doc
Add-ons
Pre-session repo reviewNDA before we startRecurring weekly slot
Not included
Production changes on your infrastructure during the call
Ongoing implementation — that becomes a package or a retainer
Your users ask in plain English. The agent calls your real APIs, streams its reasoning, and shows exactly what it did — inside your app, under your auth, in your design system.
Runs on your existing endpoints — no data migration
Every action traced, permissioned and reversible
Ships behind a flag to a pilot cohort first
Delivery
2–4 weeks
Iterations
2 rounds
Support
30 days
InvestmentScoped on the call
What you get
A tool layer over your existing API — typed, validated, permission-aware, one tool per real capability
The agent loop itself: planning, tool calls, streaming output, graceful failure when it doesn’t know
Front-end components in your design system — streamed answers, citations, an action log the user can read
Auth and tenancy respected: the agent can never see what the signed-in user can’t
An eval set built from your real questions, so changes get measured instead of guessed at
Deployment behind a feature flag, plus a runbook for your team
Package sizes
Pilot
Prove it works on one surface before you commit the roadmap.
Up to 5 tools
One screen or panel
Internal users only
2 weeks1 iteration round14 days support
ProductionMost chosen
Real customers, real traffic, real observability.
Up to 15 tools
Streaming UI in your design system
Eval suite + tracing + cost dashboard
Flagged rollout and runbook
3–4 weeks2 iteration rounds30 days support
Scale
Multi-agent, multi-tenant, or an enterprise security review ahead.
Unlimited tools, sub-agents, background jobs
Policy engine and audit trail
Load and cost tuning
Team handover sessions
6–8 weeks3 iteration rounds60 days support
Add-ons
MCP server for your APIVoice inputMobile surfaceSSO / enterprise authPrompt-injection review
Not included
Rebuilding the API the agent sits on — if endpoints are missing, that’s scoped separately
Model provider fees; you keep your own account and keep control of spend
Built with
Claude Agent SDKVercel AI SDKLangGraphMCPSSE streamingTypeScript · Python
Yes — the difference is prorated on your next invoice.help › billing › changing plans
resolved · no human needed2.1s
Support agent that answers from your own docs
Grounded in your documentation, help centre and past tickets. Cites its sources, refuses to guess, and hands to a human the moment confidence drops — with the whole conversation attached.
Every answer carries a link to the source it came from
Escalation rules you set, not ones the model invents
Weekly report of what it couldn’t answer — your content backlog
Delivery
2–3 weeks
Iterations
2 rounds
Support
30 days
InvestmentScoped on the call
What you get
Ingestion of your help centre, docs, PDFs and resolved tickets, with a refresh job that keeps it current
Retrieval tuned on your real questions — not a default vector search that returns plausible nonsense
Citations on every answer, and an explicit “I don’t know, here’s a human” path
Handoff into the desk you already use, with full conversation context attached
An answer-quality dashboard: deflection rate, escalations, and the questions it keeps failing
Embeddable widget or integration into your existing chat surface
Package sizes
Pilot
One knowledge source, internal team first.
One doc source
Internal Slack or web widget
Baseline quality report
1–2 weeks1 iteration round14 days support
ProductionMost chosen
Customer-facing, with escalation into your helpdesk.
Up to 4 knowledge sources
Helpdesk handoff + context
Citations, refusal rules, tone tuned to your brand
Quality dashboard
2–3 weeks2 iteration rounds30 days support
Scale
Multilingual, multi-product, or it needs to take actions.
Unlimited sources + auto refresh
Account-aware actions (refunds, plan changes) with approvals
Answers the phone, books the appointment, qualifies the lead, and writes it all back into your CRM. Escalates to a person on anything it wasn’t built to handle.
Answers in under two rings, 24 hours a day
Every call transcribed, summarised and logged to your CRM
Hard limits on what it may promise a caller
Delivery
2–3 weeks
Iterations
2 rounds
Support
30 days
InvestmentScoped on the call
What you get
A phone number (or your existing one, forwarded) answered by an agent trained on your business
Call flows for the jobs you actually get called about: booking, rescheduling, qualifying, triage, out-of-hours cover
Calendar and CRM writes — the call ends and the record already exists
Transcripts, recordings and summaries, searchable after the fact
Guardrails on pricing, promises and anything legally sensitive, plus instant human transfer
Package sizes
Pilot
Out-of-hours cover, or one repetitive call type.
One call flow
Voicemail-style capture + summary email
Transfer to a human on request
1–2 weeks1 iteration round14 days support
ProductionMost chosen
Inbound calls handled end to end, all day.
Up to 4 call flows
Calendar booking + CRM writes
Transcripts, summaries, escalation rules
Tone and script tuned over a week of real calls
2–3 weeks2 iteration rounds30 days support
Scale
Outbound campaigns, multiple locations or languages.
The repetitive half of the job runs itself; the judgement stays with your team. Built as real software with retries, audit logs and alerts — not a no-code flow that dies silently on a Tuesday.
Approvals land in Slack or email, one click to accept or correct
Every run logged: input, decision, output, who approved it
Failures alert a human instead of vanishing
Delivery
2–3 weeks
Iterations
2 rounds
Support
30 days
InvestmentScoped on the call
What you get
One workflow taken end to end: trigger, extraction, decision, action, confirmation
Human-in-the-loop approvals wherever a mistake would be expensive — with the reasoning shown
Integrations with the tools you already run: email, Slack, CRM, accounting, spreadsheets, internal APIs
Idempotency, retries and dead-letter handling, so a flaky API doesn’t corrupt your data
An operations dashboard: throughput, approval rate, time saved, failures
Runbook and handover so your team can change the rules without calling us
Package sizes
One workflow
The single process that eats the most hours.
One trigger, up to 3 systems
Slack or email approvals
Run log
1–2 weeks1 iteration round14 days support
DepartmentMost chosen
Three connected workflows and the dashboard over them.
Up to 3 workflows, up to 6 systems
Approval queue with audit trail
Ops dashboard + alerting
Team runbook
2–3 weeks2 iteration rounds30 days support
Operations platform
Automation as a system your team keeps extending.
Unlimited workflows on shared infrastructure
Role-based approvals
Self-serve rule editing for your ops team
Monthly reliability review
5–8 weeks3 iteration rounds60 days support
Add-ons
Document extraction (OCR)WhatsApp approvalsn8n / Make migrationOn-call alerting
Not included
Replacing systems that don’t have an API — we’ll flag those before you buy
Changing your team’s process for them; we automate the process you decide on
Built with
Python · CeleryClaude Agent SDKGmail · Slack · WhatsApp APIsn8n where it fitsPostgres
90 days. Either party, in writing.MSA-2024 · §12.3 · p.14
Document intelligence that cites its sources
Contracts, reports, policies, research, tickets — ask a question, get an answer with the paragraph it came from. Built so a reviewer can check it in ten seconds instead of trusting it blindly.
Page- and paragraph-level citations on every claim
Tables, scans and messy PDFs handled properly
Access rules enforced per document, not per index
Delivery
3–4 weeks
Iterations
2 rounds
Support
30 days
InvestmentScoped on the call
What you get
An ingestion pipeline that survives real documents: scans, tables, multi-column layouts, appendices, revisions
Chunking and retrieval tuned against a question set from your own domain experts
Answers with page-level citations, and a viewer that jumps to the source passage
Permission-aware retrieval — people only ever see what they’re cleared for
An eval set that measures answer quality and catches regressions before users do
Re-index jobs so new documents are searchable the day they land
Package sizes
Pilot corpus
One document set, one team, proof of quality.
Up to 1,000 documents
Search + cited answers
Quality baseline report
2 weeks1 iteration round14 days support
ProductionMost chosen
The whole library, in front of the people who need it.
Up to 100k documents
Permission-aware retrieval
Source viewer + eval suite
Scheduled re-indexing
3–4 weeks2 iteration rounds30 days support
Regulated
Audit trails, retention rules, reviewers who sign off.
Millions of documents
Full audit log of every query and answer
Reviewer workflow and sign-off
Knowledge graph over entities where it helps
6–10 weeks3 iteration rounds60 days support
Add-ons
OCR for scanned archivesEntity graph (Neo4j)Clause comparisonBulk summarisation
Not included
Legal or clinical interpretation of what the documents mean — the system surfaces sources, your experts decide
Digitising paper archives; we integrate with a scanning vendor if you need one
revenue by region, last quarter…same thing, split by planSQL ✓
EMEANAAPACLATAMUKDACHANZ
Talk-to-your-data analytics
Plain-English questions over your warehouse, answered with a chart and the SQL underneath it. Built on a semantic layer your analysts define, so the numbers match the numbers they publish.
Every answer shows the query it ran — auditable, not magic
Metric definitions owned by your data team, not the model
Read-only, row-level-security aware
Delivery
3–4 weeks
Iterations
2 rounds
Support
30 days
InvestmentScoped on the call
What you get
A semantic layer: your metrics, dimensions and joins defined once, in version control
Question → SQL → chart, with the generated query visible and copyable
The whole thing: auth, billing, dashboard, admin, infrastructure — and the AI core that makes it worth paying for. Shipped as a codebase your own engineers can pick up on day one.
Weekly demo on a real URL from week one
Conventional stack, documented, no bespoke framework
Your repo, your cloud, your accounts, from the first commit
Delivery
4–8 weeks
Iterations
Weekly cycles
Support
30 days
InvestmentScoped on the call
What you get
Product scoping: the smallest version that proves the thing, with everything else written down for later
Front end, API, database, background jobs, and the AI layer — one coherent codebase
Auth, roles, billing, admin tooling and the boring parts founders forget to budget for
CI/CD, environments, error tracking and logging set up from the start
Weekly demo on a live staging URL, and a written changelog you can forward to investors
Handover: architecture docs, runbook, and a walkthrough with whoever takes it over
Package sizes
Prototype
Something real to show users, investors or your board.
One core flow, clickable and working
Demo auth, seeded data
Deployed staging URL
2–3 weeksWeekly cycles14 days support
Launch MVPMost chosen
Real users, real payments, real support load.
Full product: auth, billing, admin, the AI core
CI/CD, monitoring, error tracking
Docs + handover sessions
30 days of bug fixes after launch
5–8 weeksWeekly cycles30 days support
Launch + runway
Launch, then keep shipping while you hire.
Everything in Launch MVP
Three months of continued delivery
Hiring support for your first engineers
Gradual handover to your team
8 weeks + retainerWeekly cyclesOngoing support
Add-ons
Mobile appDesign systemSOC 2 groundworkData migrationAnalytics instrumentation
Not included
Brand and marketing design — we’ll work to your designer’s files or a clean default system
Ongoing feature work after handover, unless you move to a retainer
Take the AI feature you already have and make it safe to put in front of customers — measured, traced, permissioned, and cheaper per run than it is today.
A regression gate in CI, so quality stops depending on luck
Traces that show why a bad answer happened
Typical cost reductions come from caching and routing, not cheaper models
Delivery
2–3 weeks
Iterations
2 rounds
Support
30 days
InvestmentScoped on the call
What you get
An eval suite built from your real traffic and failure cases, running in CI on every change
Tracing across the whole agent loop: prompts, tool calls, retries, latency, cost per request
It hallucinates, it times out, or last month’s bill was four figures more than it should be. Diagnosis in 72 hours with evidence, then a fix plan you can hand to anyone — including us.
First response same day, diagnosis inside three
Findings backed by traces and reproductions, not opinion
Works on code we didn’t write, including abandoned projects
Delivery
72h diagnosis
Then
1–2 week fix
Support
30 days
InvestmentScoped on the call
What you get
A reproduction of the failure, or proof that it can’t be reproduced the way it was reported
Root cause written plainly: retrieval, prompt, tool contract, context limits, concurrency, or spend
A fix plan ordered by impact, with effort estimates — yours to keep either way
The fix itself, if you want us to do it, with tests that stop it coming back
A short post-mortem your team can learn from
Package sizes
Diagnosis
You have engineers; you need the answer.
72-hour turnaround
Root cause + fix plan
Handover call
72 hours1 iteration roundWritten report
Diagnose & fixMost chosen
Stop the bleeding, then make it stay fixed.
Diagnosis plus the top fixes implemented
Regression tests and monitoring
Post-mortem + runbook
1–2 weeks2 iteration rounds30 days support
Takeover
The people who built it are gone.
Full codebase audit and documentation
Stabilisation sprint
Infrastructure and access recovered
Ongoing maintenance option
3–4 weeks3 iteration rounds60 days support
Add-ons
Emergency same-day startCost-reduction passOn-call cover for a launch
Not included
Guaranteed rescue of a system built on a broken premise — the diagnosis will say so honestly and early
Don’t see your problem? That’s what the consultancy lane is for — or run the matcher below.
Scope matcher
Describe it in your words. We’ll name the package.
Four questions, about twenty seconds. You get the package that fits, the size we’d recommend, and what the first two weeks would look like — before you talk to anyone.
How buying works
A fixed quote before anyone writes code.
We don’t quote from a form. We spend thirty minutes on the actual workflow, then send a written scope with a fixed price. If it isn’t worth building, we say that instead — it costs you nothing.
Step01
Scope call — 30 minutes
You walk us through the workflow or the product. We ask about volume, systems, who approves what, and what “working” would mean in numbers. No deck.
Free · booked within 48 hours
Step02
Written scope and fixed quote — 48 hours
Deliverables, package size, delivery window, iteration rounds, support period, assumptions, and what’s explicitly out of scope. One page, no ambiguity.
Free · yours to keep
Step03
Build in the open
Your repo, your cloud. A demo every week on a real URL, a shared channel for questions, and a changelog you can forward. No black box, no month of silence.
Weekly demos
Step04
Handover and support
Docs, runbook and a walkthrough with the people who’ll own it. Then the support window: we fix what we built, no argument about whose fault it is.
14–60 days depending on package
You own everythingCode, prompts, evals and docs, in your repo from the first commit. No licence, no lock-in.
Your accountsModel and cloud usage bills to you directly, so you can see and cap the spend.
Fixed scope, fixed priceChange the scope and we requote before doing the work — never after.
NDA on requestSigned before the scope call if you need it. Client names stay private unless you say otherwise.
Questions
The ones that come up every time.
Why aren’t there prices on the cards?
Because the same package can be a two-week build or a two-month one depending on how many systems it touches and how clean your data is. Quoting a number before we’ve seen that would be guessing, and you’d pay for the guess. The scope call is free, the written quote is fixed, and both are yours to keep even if you go elsewhere.
What counts as an “iteration round”?
One consolidated pass of changes within the agreed scope — you try it, you send a list, we work through the list. Adding a new integration, a new surface or a new use case isn’t an iteration; that’s a scope change, and we requote it before starting rather than surprising you on the invoice.
What does the support period actually cover?
Anything we built behaving differently from the written scope — bugs, regressions, breakage from a provider change. It runs 14 to 60 days depending on the package. It doesn’t cover new features, and it doesn’t expire into a support contract you have to cancel.
Can you work inside our existing codebase?
Yes — that’s the normal case. We work in your repo, on branches, through your review process, in your stack. If the codebase needs work before the AI feature can land safely, we’ll tell you that during scoping rather than discovering it in week three.
Who owns the code and the prompts?
You do — all of it, including prompts, eval sets and infrastructure config. Nothing runs on our servers, nothing requires a licence from us, and nothing stops working if you never speak to us again.
What if AI isn’t the right answer?
Then we say so on the call. A lot of “AI problems” are a missing integration, a rules engine, or a report nobody built. We’ll tell you what would actually fix it, including when that’s cheaper than anything we’d sell you.
Which time zones and languages do you work in?
Remote and worldwide, with overlap arranged around your team — we currently work with clients across the Gulf, Europe and North America. All work and documentation in English.
Next step
Tell us the workflow. We’ll tell you the fit.
Thirty minutes, no deck, no obligation. You leave knowing whether this is worth building, roughly what it takes, and which package it maps to.