FRACTIONAL HUMAN EXPERTS · ON DEMAND · HAMBURG, EU
Get sum fresh eyes on it.
AI-built software ships in weeks. Whether it survives real users, real data and the first audit is a different question. A named expert answers it in days: a written verdict, the three fixes that decide go-live, a fixed price. From vibe to live. AI runs the sweep, a human makes the call. Fresh eyes help you see what needs attention before real users arrive.
SHOW US SUMTHINK · SEE A VERDICT
Named experts · NDA before access · Your material never in our marketing · Hamburg, EU
Building got fast. Knowing if it holds did not.
AI tools and vibe coding turned months of building into weeks. The demo works, the pilot runs, everyone is happy. But speed hides decisions: an access rule nobody checked, an architecture that carries ten users and not ten thousand, an AI feature nobody can prove is right, a monthly bill nobody has priced. These things rarely fail loudly. They stall the launch, the rollout or the next budget round. And the expertise that would catch them early is exactly the kind you can't justify hiring full-time.
- 45% of AI-generated code samples introduced security flaws in tests across more than 100 language models. Working code is not the same as safe code. (Veracode, July 2025)
- 40%+ of agentic AI projects will be cancelled by the end of 2027. The reasons: cost, unclear value, weak risk controls. Not the model. (Gartner, June 2025)
- 3 fix-nows, maximum, per review. A review that flags everything gets nothing fixed. We name what matters and stop. (house rule, non-negotiable)
sumthink we review.
Pick the part that worries you. Every review answers one question in writing: what blocks go-live, and what to fix first. One object, one named expert, one verdict, in days.
01 · Requirements & concept
The AI built exactly what the spec said. The spec was wrong. Closing gaps here is the cheapest fix in software.
Christoph Merkelbach
02 · System architecture
Carries ten users, folds at ten thousand? Today that decision costs a conversation. Later it costs a rewrite.
David Körntgen
03 · Infrastructure & scale
What happens when real traffic arrives, and what it costs per month. Answered before the bill.
NAMED AT SCOPING
04 · AI integration
The demo answers well. Is it right, and what does a wrong answer cost? Known before it becomes a budget line.
Refaat Al Ktifan
05 · UX, UI & design systems
Generated screens, fifty slightly different buttons, no system. One rule-level fix repairs them all.
Christian Durlej, Philipp Datz
06 · Code
AI-written or not: read end to end, with the debt named while it is still cheap to pay off.
NAMED AT SCOPING
07 · Security
Keys in the frontend, open endpoints, missing access rules. Found in a readout, not in an incident. A review, not a pentest, and we say so.
Christoph Ebeling
08 · Compliance readiness
GDPR, AI Act, NIS2: walk into counsel with a clear posture instead of a blank page. Cheaper for you, faster for them.
NAMED AT SCOPING
A verdict, before you buy one.
Demo material, fictional product, real format. This is what lands in your inbox before the readout call.
Σ READOUT · L04 AI INTEGRATION · DEMO
Harborline Claims Assist, an LLM that drafts first decisions on incoming insurance claims for a team of eleven case handlers. Pilot running 14 weeks.
DEMO · Fictional product. No client material. Pass 0 applies to every real review.
01 · The verdict
Not ready to scale. Ready to fix. The model is not the problem. The pilot proves the drafts read well. It does not prove they are right, and nobody can say what a wrong one costs.
02 · The stall diagnosis
This stalls at the first budget review after rollout. Someone will ask how many drafts were wrong, and the honest answer today is "we don't know". Handlers accept 71% of drafts unchanged, and that number is being read as accuracy. It measures trust.
03 · Fix now
- Build a 200-case truth set from closed claims. Decided by your best handlers, scored against the drafts weekly. Without it every later number is opinion. First move: Export 200 closed claims from Q2, two handlers label them independently, disagreements go to the team lead.
- Log every override with a reason code. Today an edited draft and a rejected one look the same. Five reason codes are enough to see where the model fails. First move: Add a required dropdown to the edit screen.
- Price a draft at volume before anyone promises savings. Current cost per case is unknown because the pilot runs on a flat sandbox tier. First move: Run last month's volume through the metered API on a copy and read the bill.
04 · Fix next
Prompt and model version are not recorded per draft. Retention of claim data in the provider console is set to default.
05 · Later
Fine-tuning. Not before the truth set exists.
06 · What this is not
Not a certification, not legal advice, not an AI Act conformity assessment. If this system ever decides claims on its own, a high-risk classification check comes first, and your counsel decides.
Judgment by the slice.
The expertise that catches these problems costs a full salary, and you need it for a few days per decision, not every day. So you buy exactly those days: a two-day Focus Review, a five-day Deep Review, or Expert Check, a standing expert per lane with scheduled deep checks, drift scans in between and a priority slot when something urgent lands. All of the judgment, none of the headcount.
Start with one review. Keep the expert if it earns its place.
Focus Review
One object, one lane. Two-day timebox.
Fixed figure in your scope
Deep Review
One lane at full depth. Five-day timebox.
Fixed figure in your scope
Bundle
Two or three lanes on one object. Priced as the lanes, minus a bundle step.
Fixed figure in your scope
Expert Check
A standing expert per lane: scheduled deep checks, drift scans between them, a priority slot.
Fixed figure in your scope
If we build the fix with you, the review fee counts toward that engagement.
How a review runs.
Same protocol in every lane. The format is fixed, so the finding is what varies.
Show us sumthink.
Four quick steps, two minutes. What should we look at, what worries you, where the verdict goes. We reply within one business day with a fit call: if a review is not the right move, we say so before an order exists.
We fence it.
NDA signed before we see anything sensitive. Your material never appears in our showcases, decks or posts. Ever. Scope, lane and slot confirmed in writing, with a fixed figure. No hourly meter.
The review. Days, not months.
Our tooling sweeps the object and piles up evidence. Then the named expert does what tooling can't: reads it in context and forms a judgment. We ask for your team's disagreements up front. Tidy inputs hide the finding, so don't tidy up.
Verdict first. Then the readout.
The written verdict lands before the call: findings with evidence, a stall diagnosis named not hedged, max three fix-nows, and a first move your team can execute next week without us. Sixty minutes on it, decision-maker in the room. The document is yours. No follow-up sequence attached.
A verdict you can argue with.
Not a sixty-page report nobody reads. Not a maturity score. One document that tells your team on Monday what to do on Monday: what blocks go-live, why, and the first move.
Verdict first
The answer on page one. Everything after it is the evidence, so you can disagree with precision instead of vibes.
The stall diagnosis
Where this thing would realistically stop, and what that looks like when it happens. Named, not hedged.
Max three fix-nows
The two or three decisions that determine whether it ships, each with our recommendation and the reason.
A first move
Executable next week, by your team, without us. If you never call again, the review still paid for itself.
What this is not
Not a certification, an attestation, a penetration test or a legal opinion. We assess and we document; your counsel and your auditors decide. We will tell you plainly where you stand and what a reviewer would ask, and that is a different and more useful thing. And it is never a person on loan: you get a verdict, not a body.
Experts picked by what they build.
No career reviewers, and never "the team". Every review carries one named human from the sumthink collective, working the lane they have shipped in themselves. You meet your expert at scoping: by name, and by what they have built.
Christian Durlej
Senior UX UI Expert · L05, UX, UI & design systems
10+ years on data-dense enterprise products in MedTech, FinTech, public sector, ERP and AI interfaces. Fixes arrive as rules and tokens, not redraws.
Christoph Ebeling
Senior IT Security & DevOps Expert · L07, Security
20+ years from development and DevOps into application security. Judges findings by real exploitability, not scanner volume, and names the three that actually block the ship.
Christoph Merkelbach
Senior Requirements Engineer · L01, Requirements & concept
Requirements engineering and frontend architecture across public health, finance and consumer software. Turns fuzzy scope into decisions a team can build against, and closes the gap before it becomes a change request.
David Körntgen
Senior System Architect & Backend Engineer · L02, System architecture
20+ years building clinical-grade software in HealthTech and MedTech: privacy-first, regulatory-ready, still shipped at product velocity. Names the architecture decisions that are still a conversation before they turn into a rewrite.
Philipp Datz
Senior Product Designer · L05, UX, UI & design systems
15+ years of product design, from Apple Maps and the Siemens Healthineers design system to unu Motors, automotive UI and public-health software. Verdicts land as system decisions, not screen mockups.
Refaat Al Ktifan
Senior AI Solution Architect · L04, AI integration
10+ years building cloud platforms, commerce systems and production AI: multi-agent runtimes, RAG, MCP tool platforms with governance built in. Judges whether the AI holds up under real load, real data and a real bill.
Build things yourself? The bench grows by invitation from the collective.
Every reviewer's first deliveries are checked by a second expert before they carry a review alone.
The expert your AI can hire.
Your agents write the code, the spec, the architecture. Sumbody still has to own the verdict. This page is machine-readable today. The booking endpoint for agents is in build. Either way, a booking is confirmed by people on both sides, and NDAs are signed by humans.
- Findable by machines. This page, its llms.txt and its structured data describe the offer for assistants.
- Bookable by agents. The booking endpoint is in build. Until it is live, the four steps below do the same job in two minutes.
- Closed by humans. Your agent requests, you confirm, a human on our side accepts the slot.
sumthink you'd ask.
- Is it a problem that our product was built with AI tools?
- No. Speed is the point, and we build with the same tools. The only question is whether the decisions inside it were checked: access rules, data handling, architecture, cost, what the AI is allowed to decide. That is exactly what a review reads, and it reads it in days.
- How fast is "on demand"?
- We publish our slot count and hold it: a fixed number of review starts per week, first come first served. You get a start date in the scope confirmation, typically within days. We would rather tell you "next week" than start late.
- What does it cost?
- A fixed figure, named in the scope confirmation before you commit, never an hourly meter. Reviews run in a two-day or five-day timebox depending on lane and depth. And if we end up building the fix with you, the review fee counts toward that engagement.
- How is this different from a pentest or an audit?
- A pentest actively attacks a system to prove vulnerabilities. An audit or certification checks against a standard and is issued by accredited bodies. Our review judges decisions: spec, architecture, code, AI use or security posture, and says what to fix first. It replaces neither a pentest nor a certification, but it often prepares both for less.
- Why only three fix-nows?
- Because a review that flags everything gets nothing fixed. You get every finding with its evidence. Three of them are marked as the ones that decide whether it ships. The rest are labelled fix next or later.
- What do you need from us?
- Read access to whatever exists: repo, spec, tickets, the deck, the prototype. No preparation, no tidying up. Tidy inputs hide the finding. Plus 30 minutes with the people who disagree about it.
- Who sees our material?
- The named expert on your review, under NDA, and nobody else. Your material never appears in showcases, posts, decks or sales conversations. This rule has no exceptions and predates this page.
- What if we disagree with the verdict?
- Good. Every finding ships with its evidence, so you can argue with the document instead of with an impression. Some of our most useful readouts ended in a well-founded "no, and here is why". The point is the decision, not the applause.
- Do you work in German?
- Yes. Reviews, readouts and NDAs in German or English, your call. There is a German version of this page at sumthink.io/de/expertsreview.
- What is a fractional human expert?
- Expert judgment bought by the slice instead of by the hire. You get a caliber of reviewer you could not justify employing, for exactly the days the decision needs. Fixed scope, fixed figure named in the scope confirmation, one named human.
- Can our AI agents really book you?
- They can already find us: this page ships machine-readable. The direct booking endpoint for agents is in build. Either way the contract and the NDA happen between people; your agent just saves you the search.
Show us sumthink.
Through the request form below, in about two minutes. A human gets back to you within one working day for a short fit call, with scope and start date if it fits.