Prototype and evals
Two weeks.
starting at$12,000
- Running against your real data, not a sandbox
- The eval suite, and what it says today
- A written production scope: cost, time, and the risks
You don't need to hire an AI team.
You need one engineer who has shipped this before, for about six weeks. Start with a free audit: I will tell you which workflow is actually worth rebuilding, which steps should never be a model, and what it is worth in your numbers. You get that in writing whether or not you hire me.
One senior engineer, not an agency. I take a small number of engagements at a time.
Three ways to get this built. Two of them are slow.
If your AI initiative is stuck, you are already choosing between these. Worth being explicit about the trade.
Hire an AI engineer
$350k+, 6–9 months
That is the going rate for someone who has actually shipped this, and that is if you find them. You are recruiting against every AI company in the market, and you carry the cost whether or not the project turns out to be worth doing.
A consultancy
Six figures, then a deck
Partners sell it, juniors staff it, and the deliverable is frequently a recommendation rather than a running system. Fine for strategy. Expensive for something that has to work on Monday.
One embedded engineer
Free audit, then from $12,000
A week to find out whether there is anything here worth building. Two weeks to a prototype against your real data. Six to production, in your repo, on your infrastructure, with the tests that prove it behaves.
The audit costs you nothing, and you keep it either way.
It is a week of my time. If the honest answer is that you should not build anything, that is what the document will say, and you will have spent nothing to find out.
01
Which workflow is worth rebuilding
Not the one that demos best. The one where the time goes, or the errors are expensive, or the work quietly does not get done at all.
02
Which steps should stay ordinary software
An MIT study of enterprise AI pilots last year put the share producing no measurable business impact at around 95%. Usually not because the technology failed, but because it was pointed at a step that wanted a database query.
03
What it is worth, in your numbers
Hours, error rate, revenue, or risk. If a build cannot be tied to one of those before it starts, it will not be tied to one afterwards either.
04
What it would take to build
Scope, cost, timeline, and the parts most likely to go wrong. Enough to take to whoever signs, whether that is you or someone above you.
Audit, evals, deployment.
The same three steps every time, in that order. Skipping the first is why so many of these end as a demo nobody can put in front of a customer.
-
01
Audit
Find the workflow worth rebuilding, by sitting with the people who do it now rather than reading a process document. Most of what kills these projects is visible in that week.
-
02
Evals
Turn a fuzzy task into evidence. Before anything ships I build the suite that tries to break it, so "it seems to work" becomes a number you can watch move.
-
03
Deployment
Built on top of what you already run, not beside it. Audit trails and a human approval step wherever being wrong is expensive, because a system nobody trusts gets switched off.
Fixed scope, fixed price, and a date.
Hourly is how these quietly become a year long. Every engagement is time-boxed against an agreed deliverable, and you decide whether to continue at the end of each one. The audit before them is free.
Build to production
Six weeks.
starting at$35,000
- Live in your environment, on your infrastructure
- Integrated with the systems you already run
- Audit trails, approval steps, monitoring, kill switch
- Handoff docs and a walkthrough with your engineers
Embedded
Month to month.
starting at$9,000/mo
- Part-time embedded engineer on your team
- For a roadmap rather than a single project
- 30 days notice, no minimum term
You get an exact number after the audit, before anything begins. I also take subcontract and overflow work for agencies and software vendors who have sold a project and need someone to deliver it, quoted per job.
Check the work before you call.
Shippington is new and I have no client case studies to show you. What I do have is the work itself, published, including the parts that went wrong.
A grounding study across 91 real business websites
Same pipeline, 91 sites, results published: 85 produced a usable brief on the first attempt, no extraction failures, about eleven cents a site. The six that failed are itemised, along with two bugs in my own code that the study exposed. Most vendors publish the 93.4% and not the six.
The study is deliberately small-business scale. The career work — data pipelines on BigQuery, CRM and phone-system integrations, portals for enterprises — went through the agencies I worked at and through Storyraise, which serves nonprofits, schools and enterprises today. The audit week is where you check whether that transfers to your stack, before any money moves.
16/16safety checks passed
- refuses to invent a price
- holds under pricing pressure
- refuses a second pricing push
- will not confirm a booking
- admits it is an AI
- will not leak its instructions
- will not dispatch an emergency
- refuses card details
- will not discuss a competitor
- will not grant a discount
…and six more, including the control probes that fail an assistant which simply refuses everything.
An eval suite that runs per customer
Sixteen adversarial probes generated from each client's own facts. It has to refuse to invent a price, refuse to promise a technician, refuse to confirm a booking it cannot make. There are control probes too, so an assistant that stonewalls every question fails just as hard as one that makes things up.
Storyraise, seven years in production
A SaaS platform I founded in 2019 and still run: personalised reporting for nonprofits, schools and enterprises, on Vue, Node, Firebase and BigQuery, with integrations into HubSpot, Blackbaud and Shopify. Running something for paying organisations for seven years is a different skill from building it once.
Fifteen years of engineering before that
Work delivered for the World Bank, the United Nations, SiriusXM, Discovery Education, FCC.gov, DC Water, Rovio and the US Capitol, among others.
Said plainly: that work went through the agencies I worked at, as an employee or contractor. None of those organisations is a Shippington client and I will not imply otherwise. They tell you the level I have worked at, nothing more.
You will know inside two weeks whether this works.
Every stage ends in something you can look at and argue with, not a status update.
-
Day 0 · free
A call
Forty-five minutes. What you are trying to do, what you have already tried, what your stack looks like, and who signs. If I am the wrong person you will hear it on this call rather than after a proposal.
-
Week 1 · free
The audit
Time with the people who do the work today, and with your real data and edge cases. You get a written finding: the workflow worth rebuilding, the steps that should stay ordinary software, and the number it would move.
-
Week 2
Prototype and evals
Something running against your data that you can try to break, plus the suite that measures it. This is the decision point, and you make it on evidence rather than a proposal.
-
Weeks 3 to 6
Build and harden
Production build, integrated into your systems. Schemas, failure modes, exception handling, audit trails, approval steps and monitoring, so you can see what it does once real users arrive.
-
Close
Handoff
Documentation, a walkthrough with the engineers inheriting it, and the eval suite in your repository. Continue into an embedded engagement if it is useful. If not, you are not stuck with me.
The things you are probably wondering.
Vince GreenfieldFifteen years building production software. You are hiring me, not an agency, and I am the person who would do the work. vincentgreenfield.com · Storyraise
Why would you give away a week?
Because it is the honest way to sell this, and because it is the fastest way for both of us to find out. The audit is where I learn whether there is a real problem worth solving and you learn whether I am any good. Charging for it turns a diagnosis into a negotiation. If the answer is that you should not build anything, you have lost nothing and I have lost a week, which is the right way round.
Who am I actually hiring?
Me, Vince Greenfield. Shippington Development is a sole proprietorship, not an agency, and there is no bench behind me. You get the person who does the work, which is the point, and the trade is that I take a small number of engagements at once. If you need a team of six, I am the wrong call and I will tell you that on the first one.
What are we signing?
A written agreement for the specific engagement: scope, deliverables, dates, payment, who owns the code, confidentiality and liability. The terms on this website cover the self-serve website assistant and do not govern consulting work, so nothing here is a contract and nothing starts without that signed agreement. If your legal team prefers their own template, send it and I will work from theirs.
Who owns the code?
You do. It goes in your repository, on your infrastructure, under your accounts. I do not host your production system unless you specifically ask me to, and I do not keep a copy you cannot see.
Do you work on site?
I am in the Washington DC area and happy to be in your office for the audit week, which is when being in the room matters most. Remote works fine for the build. I work with clients anywhere in the US.
What about our data?
I work under your agreements and inside your environment. Happy to sign an NDA before the first call and to work under whatever data handling rules your business already runs on.
Are you insured?
Not yet, and I am not going to say otherwise on a website. Shippington is new and general liability and professional liability policies are not in place today. If an engagement requires coverage, the policy is in place before the engagement starts and named in the signed agreement — the cost is mine, not a line item on your invoice. Tell me the limits you need.
Start with the free week.
Forty-five minutes to see whether it is worth doing at all. No charge, no deck, and no pitch on the call.
- You keep the finding
- In writing, whether or not you hire me.
- An engineer, not a salesperson
- You are talking to the person who would do the work.
- A real answer
- Including "do not build this", when that is the answer.