Pitch page (one screen)

The video (60 seconds)

Slot	Strongest proven line today	Source	Gap
1. Who we are (~10s)	Marton builds the system. Xavier co-founded Girl Meets Dress (2009) as CFO/COO, then was entrepreneur in residence at Seedcamp.	brief	How they met and for how long: not on file.
2. The problem (~15s)	No buyer quote on file yet that can be shared for W1 or W8.	call notes	One named person with the pain, in their own words, is still needed (see questions below).
3. What we built (~20s)	Marton's logs show about 5,600 agent jobs from three AI makers, each checked by a different AI, with about 160 failures each turned into an automatic check. One number moved: feedback lines per sales proposal fell from 44 to 18 in two weeks.	system inventory, brief	It comes from Marton's own sales proposals, not a customer. No customer number. Count how many of the about 160 records have a real check (some may say "none, accepted risk").
4. Why us, why now (~15s)	Of 72 YC rivals we tore down, none publishes a customer number from checked results turning into rules (Bento Labs shows benchmark gains only). We run that loop every day, on our own business.	landscape, sections 5 and 7	"Why now": none found (landscape section 8). It must come from the founders.

The demo (up to 3 minutes)

Show only what runs today, in this order:

1. A job going through the line (60s): a task is picked up, built in its own copy, checked by a different AI, then put through the automatic tests.
2. The heartbeat (45s): what is stuck, what silently stopped, what it cost.
3. One failure becoming a check (45s): an incident record, then the automatic check that stops it coming back.
4. The learning number (30s): the proposal feedback count falling from 44 to 18 as rules were added.

The one number it proves: 44 to 18 feedback lines per proposal in two weeks (Marton's logs). Do not show the merge rate as a strength: only about 15% of jobs reached the main code.

Outstanding partner questions

Owners and dates are proposed; the founders confirm them.

Question a YC partner will ask	Answer that would close it	Owner	By
Who has paid for the system?	One paid pilot or a signed letter of intent.	Xavier	Mon 26 Oct
Which one buyer, which one pain?	One wedge picked, with a named person's own words on file.	Both	Wed 28 Oct
Which number moves, before and after?	A first-week number from a live pilot. A 4-week run can not finish before 1 Nov, so say what is measured and when.	Marton	Thu 29 Oct
How are you different from Conductor, Replicas or Nessie?	The DRAFT lines in landscape section 2, said in their own words.	Both	Thu 29 Oct
Why can't a rival with $22M add a check step in weeks?	A data, loop or buyer edge they lack (M3 is capped at 3 today).	Marton	Wed 28 Oct
Why now?	A dated change that helps us more than them. None found in 72 teardowns.	Both	Wed 28 Oct
Only about 15% of your jobs reached the main code. Why trust it?	The rate after the October fix, and what "checked" catches.	Marton	Thu 29 Oct
How is client data kept safe?	A short written answer: where data lives, who can see it.	Marton	Thu 29 Oct
How did you meet? What is the equity split?	Both facts written down.	Both	Thu 29 Oct

Detail behind every line: teardowns/landscape/landscape.md and teardowns/companies/.

YC landscape vs XMarts (synthesis, 19 Oct 2026)

In short: 1. No rival is hard to copy us. Of 72 deep teardowns, 0 rate our edge hard to copy (47 medium, 24 easy, 1 unknown). So rubric M3 can not score above 3 on any wedge today. 2. No "why now" found. All 72 teardowns say "None found" for a dated, sourced change that helps us more than them. M4 has no outside evidence yet. 3. W5 and W2 are crowded. W3, W4 and W9 are nearly empty. Direct rivals in the 72: W5 24, W2 19, W1 9, W8 4, W7 3, W6 2, W3 1, W9 1, W4 0. Empty can mean "no market", not "open market". 4. "Most rivals do one layer" holds, with 8 named exceptions. 44 of 72 reach level 2 on one layer or none. 8 reach level 2 on all three (list in section 4). 5. The edge every teardown lands on is the same: checked results turn into rules, so the next run is better, with a number to show it. No rival we read publishes a customer number for this (Bento Labs shows benchmark gains only). It is proven only on Marton's own business, and most rivals could add it in weeks or months.

Every fact below comes from a file in teardowns/companies/. Their numbers are their claims. Our numbers are "Marton's logs show", from research/system-inventory-2026-10-19.md or the brief.

---

1. The map: layers by buyer

Count of companies scoring 2 or more on each layer, by buyer size and type.

The 72 deep teardowns:

Layer	Enterprise, business	Mid-market, business	SME, business	SME, developer	Unknown size
L1 Company brain	4	6	11	4	1
L2 Agent orchestration	0	6	23	6	4 (+1 enterprise developer)
L3 Execution and learning	0	3	16	3	7

The 664 on the deep list (screen scores, from YC text only):

Layer	Enterprise, business	Mid-market, business	SME, business	SME, developer	Unknown size
L1 Company brain	36	36	61	6	55
L2 Agent orchestration	40	38	79	14	66
L3 Execution and learning	23	36	51	6	49

• Crowded: agents that run work for small businesses (L2, SME, business). This is the vertical "AI layer for one trade" pattern: accounting, insurance, law, clinics, distributors.
• Thin: L3 for mid-market and enterprise buyers. Very few rivals both do the work and measure it for a company of 100+ staff.
• Thin: anything for mid-market developer buyers (0 in the deep 72).
• Caution: the deep 72 were picked by relevance to us, so they over-count our own space. The 664 row is the wider view.

2. Per wedge

"Direct" counts the 72 deep teardowns. In brackets: all 2,870 companies, using the teardown score where one exists and the screen score elsewhere (yc_matrix.py).

Wedge	Direct	3 closest rivals	Whitespace in one sentence
W1 Trusted 24/7 coding	9 (14)	Conductor ($22M, they claim), Replicas ($4M seed, 13 Oct 2026, they claim), GitStart	Nobody shows that accepted and rejected work teaches the agents, with a fall in review time or rework.
W2 Are your agents producing?	19 (24)	Weave ($13.5M Series A, Jul 2026, they claim), Lemma, Raindrop	Rivals watch traces, chats and cost; "did the work actually land" for a team's own agent fleet is less covered, but easy to add.
W3 Company memory from calls, services firms	1 (1)	Datafruit (direct); Foaster and Nixo (adjacent)	Call memory that drives the next step (proposal, delivery, check) for smaller, non-software services firms.
W4 Fit call to proposal in 10 min	0 (0)	Covera, Praxis, CopyCat (adjacent: intake to quote inside one trade)	No YC company owns it for consultancies. That may mean no buyer: test demand before claiming space.
W5 One layer for one type of small firm	24 (24)	Combinely (UK accountancy, £180 a seat, they claim), Double (all three layers, 4,000+ firms, they claim), Bluebook	Only "a UK trade nobody serves, plus a published before/after number" is left. UK accountancy is taken.
W6 Cursor for product managers	2 (8)	Kestral, Backdrop, Linzumi (adjacent)	No rival checks the work or shows a number; Kestral already has the buyer and the call inputs.
W7 Find and fix margin leaks	3 (3)	Glimpse (a16z-backed, US consumer goods), Handled, Autostep	One leak in one US industry is covered; cause finding (price or demand) for UK mid-sized firms is open.
W8 The other 90% of a skill	4 (5)	Nessie, Glen, Wato	They store, share and track skills; none shows skills improving from checked results with a number. Easy to copy.
W9 Ask my agent before you book my time	1 (1)	Memory Store (inference by the teardown)	Weak: the brief already names Viven and Twin1 (outside YC) as funded.

The line we would say against the closest rival (all DRAFT, built only from what Marton's logs show; the founders say it their own way):

• W1 vs Conductor: "Conductor is a window where a person runs and reviews coding agents. Our system runs a queue of agents day and night, has a different AI check every job, and turns each failure into an automatic check, on our own business."
• W2 vs Weave: "Weave measures coding agents from the outside. Our system runs a fleet of them, and our own logs show what got stuck, what failed silently and what it cost."
• W3 vs Datafruit: "Datafruit records calls for big software integrators. We already run call-to-memory on our own business, and we are building it for smaller services firms."
• W4: no direct rival to answer. The proof is demand, not difference.
• W5 vs Combinely: "Combinely sells accountants a seat with a library of agents. We are not an accountancy tool; we want to run one workflow for a firm and own one number." Weak: we have no customer and no trade picked.
• W6 vs Kestral: "Kestral decides what the team should build. Our system runs agents that do the work and check it, every day, on our own business."
• W7 vs Glimpse: "Glimpse proves agents can win back cash on one leak for consumer brands. Our system already runs checked agent work every day, and we are picking the first leak with real buyers." Weak: no buyer calls on W7 yet (brief).
• W8 vs Nessie: "Nessie remembers what your agents did and tracks how your skills are used. Our system runs the agents, has a second AI check the work, and turns each failure into a rule."
• W9: not worth a line yet.

3. What YC funds now (M2)

Themes across every screened company in the last 4 batches that have started (Fall, Summer, Spring and Winter 2026), from yc_matrix.py, top 10:

Theme	Count
other	130
developer tools	115
vertical ops layer	92
agent team	64
back office automation	58
finance ops	37
company brain	34
agent monitoring	32
ai services firm	32
coding agents	27

Set against YC's requests for startups (research/yc-landscape-2026-10-19.md):

YC request	Season	What YC funded next to it
Company Brain; AI Operating System for Companies	Summer 2026	34 "company brain" companies in 4 batches. Funded and crowded: a brain alone is not a pitch.
AI-Native Service Companies; AI-Native Agencies	Summer and Spring 2026	92 vertical ops layers and 32 AI services firms. The most funded pattern near us. It backs the W5 shape, which is also why W5 is crowded.
Cursor for Product Managers	Spring 2026	Few: W6 has 8 direct in the whole screen. Open, but YC has asked and few have answered.
Multiplayer AI (current list)	Fall 2026	Agent teams (64) and coding agents (27). W1 sits here.
Infrastructure for Multi-Agent Systems	Fall 2025	Agent monitoring (32). W2 sits here, and it is crowded.

Read: YC is funding "agents that do one job for one buyer" far more than "a brain for every company". This backs the brief's wording rule: lead with the result on one workflow; the brain is how, never the headline.

4. Full-loop check

Companies at level 2 or more on all three layers:

• In the 72 deep teardowns: 8. Combinely (UK accountancy), Double (US accounting), burnt (food distributors), Handled (consumer brand finance), careCycle (Medicare agencies), Replicas (coding), Trellis (short-term rentals), Rulebase (scored 2, 2, 3, on its own claims).
• In all 2,870 screened: 23.
• What it means: the brief's claim "most funded rivals do one layer" holds by count. But "nobody does the full loop" is false. Seven of the eight sit in one vertical with one buyer. Replicas is the only full-loop rival on coding, and it scores L1 2 on thin evidence.

5. Draft form answer: "Who are your competitors? What do you understand that they don't?"

The top 3 below follow the brief's angle tests (1. W1 + W8, 2. W1, 3. W8). The sealed rubric scores may pick others. All DRAFT; the founders say it in their own words.

1. W1 + W8 (agent team whose skills improve from accepted and rejected work):

DRAFT. Our closest competitors run coding agents (Conductor, Replicas, GitStart) or store and track agent skills (Nessie, Glen, Wato). We understand that the limit on agent work is trust in what comes back, not how many agents you can run. Marton's logs show our system has run about 5,600 agent jobs, each checked by a different AI, and every failure became a written record with an automatic check (about 160 so far).

2. W1 (trusted 24/7 coding):

DRAFT. Conductor and Replicas give engineers agents and a place to review them; GitStart sells reviewed pull requests with people behind the AI. We understand that teams don't need more agent output; they need output they can trust without reading every line. Our system checks every job with a second AI and turns each failure into a permanent check.

3. W8 (the other 90% of a skill):

DRAFT. Nessie, Glen and Wato store, share and track the skills that agents use. We understand that a skill gets better from checked work, accepted or rejected, faster than from reading traces. Marton's logs show this on his own sales proposals: feedback lines per proposal fell from 44 to 18 in two weeks as 22 rules were added.

Note for the founders: that number comes from Marton's own sales proposals, not from a customer.

6. What we can not claim yet

Line we want to say	Why not yet	Proof it still needs
"They map the work; we do it and own the number" (vs Foaster, Within)	No outside customer has run the system (brief).	One customer workflow run 4+ weeks with a before/after number.
"Our brain learns from checked results"	Shown on Marton's own work only.	A customer's rule-change log tied to checked runs, with rework falling.
"Trusted 24/7 coding"	Marton's logs show only about 15% of jobs reached the main code, and landing was unreliable in October.	One outside repo run 4+ weeks with merged work and a bad-merge rate.
"Value in days, priced on the result, for UK mid-sized firms"	No pilot, no price, no UK lead named.	3 to 5 paid pilots or letters of intent.
"Nobody does the full loop"	False: 8 of 72 do (section 4).	Drop it. Say "most stop at one layer" and name the exceptions.
"Nobody watches agent fleets"	False: 19 direct W2 rivals.	Drop it.
"UK accountancy is open"	False: Combinely sells there, and Bluebook counts as European.	Drop it, or name a different trade with a named buyer.
"Company brain" as the headline	34 funded in 4 batches; Hyper, Memory Store, Savant, Nessie and Glen use the words.	Drop it as a headline (the brief says so too).
"Every failure becomes an automatic check"	The inventory says each record must name a check, and a check can be recorded as "none, accepted risk". The brief uses the stronger wording.	Count how many of the about 160 records have a real check, and say that number.
"We are cheaper or faster" (vs Covera, Combinely, Cranston)	They publish prices and speeds; we have neither.	A price and a timed first result.

7. M3 check (can they copy our edge?)

Top-3 wedge	Closest rivals and their copyability	M3 cap
W1 + W8	Conductor easy, Replicas easy, Nessie easy, Glen easy, Wato easy	3
W1	Conductor easy, Replicas easy, GitStart medium	3
W8	Nessie easy, Glen easy, Wato easy, Method medium	3

No rival in the 72 is marked hard, so M3 can not go above 3 on any wedge. To earn a hard, the edge must need data, a loop or a buyer the rival lacks, and chasing it must hurt their core business (lens). The best candidate (inference): months of checked results for a non-technical UK buyer that developer-tool rivals do not sell to. That is unproven today.

8. M4 check (why now)

None found. All 72 teardowns looked for a dated, sourced change from the last 2 years that helps XMarts more than the rival. None found one.

Two dated changes in the brief cut the other way (inference): Claude Enterprise reviews every shared skill version by default since 2 October 2026, which turns skill sharing into a feature, and the August 2026 study of 105 software releases shows the skill-upkeep problem is real for everyone, not only for us.

So M4 has to come from the founders' own evidence, not from the market. Section 9 of the pitch page lists it as an open question.

---

Sources and method

• Screen: teardowns/screen/screen.jsonl (2,870 rows, lens_hash 179096c479), deep list teardowns/screen/deep-list.csv, fixes teardowns/screen/screen-fixes.md.
• Deep teardowns: teardowns/companies/*.md (72). The list is teardowns/screen/pass2-list.txt.
• Matrix: teardowns/landscape/matrix.csv.
• YC requests: research/yc-landscape-2026-10-19.md.
• Us: research/system-inventory-2026-10-19.md and brief/xmarts-brief.md.
• 7 of 72 teardowns were checked against their sources (1 in 10); see the commit message for the result.