01
Pre-mortem
Rewrites the 2029 post-mortem as evidence comes in, so the team always knows its likeliest cause of death.
Every founder gets a co-founder on day one.
An AI twin that builds a team's startup in an hour, then spends the programme helping them prove it, or kill it. A pilot proposal for Citi upStart at National College of Ireland.
Scroll
01 · The bottleneck
Accelerators run on mentor hours.
Most of what a top accelerator gives a founder depends on time with someone who has done it before. In the 2025 to 2026 Citi upStart cohort at NCI, that time was shared like this:
4academic mentors
10student teams
7months to the Den
Here is what a top programme hands a founder, and how much of it an AI can now carry for every team at once.
Blunt, frequent feedback
Weekly office hours, at any hour of the day
AI carriesPressure to find customers
Weekly evidence, visible to the whole cohort
AI amplifiesPitch rehearsal
A mock Den whenever the team wants one
AI carriesThe first build
Problem, prototype spec, page and deck on day one
AI carriesTrust, introductions, money
Where every mentor hour should now go
Stays human02 · Day one, replayed
An afternoon's work in under five minutes.
09:42:3628 Oct 2026 · Europe/Dublin
A real run on a fictitious demo team called Bunkwise. Timestamps are the files' own. All 27 automated checks passed. No real founder data was used.
09:42:36
personas.json · input
60 personas weighted to CSO Census 2022
09:44:22
problem.md
One segment, three riskiest assumptions, a prototype spec
09:44:54
landing.html
A landing page whose deposit button does nothing yet
09:45:37
deck.html
Ten slides, every claim tagged
09:45:39
ledger.json
Twelve claims, none of them proven
09:46:36
simulation.json
Sixty simulated reactions. Only four fitted the segment.
09:46:52
interviews.md
Twenty non-leading questions for real people
09:47:01
postmortem-2029.md
The most likely way it dies
09:47:23
review-internal.md
It scored itself 61 out of 100 on the Citi grid
03 · The battery
Seven specialists, tuned to one team.
02
Customer discovery coach
Prepares interviews, flags leading questions, turns notes into ledger evidence.
03
Office hours
A blunt partner at 2am that asks what customers said this week before it looks at the deck.
04
Kill contract
Holds the team to the kill criteria it set in week one, and records every pivot with its reason.
05
Unit economics
Builds and stress-tests prices and costs, marking every input as measured or assumed.
06
Risk and regulation
Raises legal, data and regulatory questions early, so a human expert sees them in time.
07
Mock Den
AI judges score the pitch on the Citi grid and ask the questions that hurt. Scores go to mentors only.
Each team gets its own set, the way Victor's MSc supervision portal at NCI gives every student AI specialists matched to their project.
04 · The ledger
A prediction becomes a fact only when tested.
Every claim in a team's deck starts as a point of cold light. It turns solid when a named person records the evidence behind it, with the source, the date and what exactly it supports.
Simulated
Evidence attached
Reviewed by a person
Proven by the AI
0claims tested, of 2,600 in the field
The field is an illustration. The rule behind it is real: research on language-model "silicon samples" finds they track some aggregate patterns (Argyle et al., Political Analysis, 2023) but give reduced variation and unreliable inference (Bisbee et al., Political Analysis, 2024). So simulated customers only ever generate hypotheses.
05 · A letter from 2029
Prove it wrong
"Students did not trust Bunkwise for the same reason they did not trust the listings."
On day one the Twin writes the most likely story of how the startup dies. The team spends the programme proving it wrong. This one asked for ten paid checks from real students, a test of whether trust comes through the university channel, and a count of how many checks one person can do in a real August week.
06 · Who it serves
Everyone gets their hours back.
Founders
A working first version on day one and a co-founder available at any hour. Their work moves from polishing slides to testing claims, which is where the Citi grid gives marks.
Mentors
No more first-pass deck feedback. The dashboard tells them which team needs an introduction, a hard conversation, or encouragement.
Programme lead
One screen for the whole cohort, week by week. Quiet teams show up early.
Funder
A cohort report built from the ledger: claims tested, interviews run, pivots and kills with their reasons.
Judges
Each claim in a pitch can show whether it was tested or estimated.
07 · Many minds
The right model for each job.
Models from Anthropic, OpenAI and Google, on metered API access. Small models log and triage, mid-tier models coach daily, and the most capable models build on day one and sit on the mock Den. Every provider has to pass the same 27-check test before it gets a job.
USD 1.37for one complete day-one build at API list prices, measured once on the demo. Everything else gets metered in the pilot.
08 · The pilot
One cohort. Measured honestly.
A pilot with the next NCI Citi upStart cohort, with teams joining by consent. Citi funds it, so it costs neither NCI nor the teams anything. Start date to be confirmed.
Phase 0
Prepare
Conflict of interest disclosed, data-protection review, consent
Phase 1
Day one
Every consenting team gets its eight artefacts
Phase 2
The programme
Weekly workflows, ledger, office hours, dashboard
Phase 3
The Den
Mock Den rehearsals, then the real one
Phase 4
Evaluate
A cohort report to Citi and NCI
How we will judge it
- Teams still using it after week fiveretention
- Claims moved from simulated to reviewedlearning
- Mentor hours, and where they wenttime
- Gap between simulated and real answerscalibration
- Founders' own view, midpoint and endsurvey
What we ask Citi to fund
- Compute for every participating team[to be costed]
- Platform, dashboard and ledger[to be costed]
- Build and support during the cohort[to be costed]
- Evaluation and cohort report[to be costed]
If things slip, the specialists go first, then the dashboard, then office hours. The floor is the day-one build plus the evidence ledger for every consenting team.
09 · The path to scale
From one cohort outwards.
YBI member network · to confirm
Incubators in Ireland and abroad
New Frontiers · 18 sites
Citi upStart elsewhere · to confirm
NCI pilot
Enterprise Ireland's New Frontiers programme runs in 18 locations through institutes and technological universities. One design repeated across 18 sites each year makes it a natural national customer.
10 · Why us
Victor del Rosal's first job was country lead for Youth Business Mexico, part of Youth Business International: seed capital and mentoring for young entrepreneurs. The limit was always mentor hours.
Today Victor lectures at National College of Ireland, has mentored Citi upStart teams, and through fiveinnolabs builds AI systems in daily use, including an AI-assisted MSc supervision portal and the personal AI system that built and tested the Twin.
Disclosurefiveinnolabs would be an external vendor to a programme Victor has mentored in. We disclose that at the start, follow NCI's conflict-of-interest policy, and agree in writing how staff time, fiveinnolabs work and ownership of the platform are kept separate.
11 · What isn't proven yet
We tag our own claims too.
Evidence needed
No real founder has used it. The demo team is fictitious.
Evidence needed
Whether teams keep using it after week five.
Evidence needed
How close simulated answers land to real interviews. The census panel missed the demo segment: 4 of 60.
Evidence needed
What weekly coaching costs per team. Only the day-one build has been measured.
12 · Next
Pilot it with the next cohort.
A first conversation, a written disclosure, a data-protection review, and a costed plan from a metered trial on fictitious teams.