Datafruit (YC Summer 2025): teardown vs XMarts

In one line: Datafruit records the customer calls of software implementation firms (the consultancies that roll out Salesforce, Oracle and similar systems), turns them into scope and deliverables, and keeps the lessons for the next project [site] [yc]. Threat: 2 because it targets the same kind of buyer and the same problem as W3 (memory from every call for services firms), but it is two people, names no customer and stops at scope and handover [yc] [site].

Big finding: this is a pivot. In about Sept 2025 the same company launched as "Devin for DevOps", an AI agent for cloud infrastructure [press]. By March 2026 it was writing about implementation consultants [blog]. The founders said on the DevOps launch that some sales were lost because buyers were content with Claude Code [press].

1. Fact file

Item	Finding	Source
What they say	Headline: "The AI workspace for software implementations." YC line: agents "capture every conversation, structure scope from day one."	[site] [yc]
Founded, team, funding	Founded 2025, YC S25, San Francisco. The YC page now shows a team of 2; an earlier YC snapshot and a Launch HN post show 3 to 4 founders (conflict; see Unknowns). Third-party directories list about $500k raised, which matches the standard YC deal (inference).	[yc] [press]
Founders	Venkat Arun (CEO): Berkeley computer science and statistics, cloud infrastructure and data pipelines at early startups. Nick Papciak (CTO): ex-Meta, Georgia Tech machine learning. Tom Jeong is listed as a cofounder in an earlier YC snapshot. What YC likely saw: technical founders with infrastructure experience (inference).	[yc] [press]
Customers	None named anywhere on the site, blog, events page or hiring page. They say "enterprise clients" only. The carousel project names look like demo data.	[site] [jobs]
Price	Not public. Only call to action is "Book a demo". The forward-deployed engineer ad ($150k to $250k, on-site with enterprise customers, on-prem installs) hints at a heavy enterprise sale (inference).	[site] [jobs]
Path	Agentic PaaS first (hard to win customers, per the founders) → AI DevOps agent, Launch HN about Sept 2025 → implementation workspace, first post 9 Mar 2026. Old domain datafruit.dev now redirects to datafruit.ai.	[press] [blog] [site]
Open roles	Forward Deployed Engineer, Founding Designer, AI Engineer. All San Francisco.	[jobs]
Compliance	"Actively pursuing" SOC 2 Type II, not yet certified. Optional zero data retention. Role-based access.	[site]
Events	Oracle AI World, Las Vegas, 25 to 28 Oct 2026; SuiteWorld 2026. 30-minute meetings with Oracle partners and implementation teams.	[site]

2. Their proposition, reverse engineered

Part	What they do
Who buys	System integrators and services teams that implement enterprise software (Salesforce, Oracle, NetSuite, SAP are all named in posts and integrations). Heads of delivery and firm founders [site] [blog].
The pain	Projects lose money in three ways: unbilled small requests (scope creep), context lost at handoff, and lessons that stay in people's heads [blog]. Discovery fails when uncertainty is mistaken for alignment [blog].
The promise	Better estimates, less risk, and every engagement becomes a traceable system of record [site].
How it works	Records calls from Zoom, Meet and Teams. Pulls context from Slack, Jira, Salesforce, HubSpot, Gong, Drive, SharePoint. Structures requirements as objects, tracks changes against a live scope baseline, drafts deliverables [site] [blog].
The output	Structured requirements, scope checks, ready-to-send deliverables, and a cross-project knowledge store [site].
Proof style	Words, not results. The only numbers are a vendor-run benchmark: they claim the best model finds 73.1% of the requirements in synthetic meeting transcripts [blog]. No customer money or time result is published.
Who sells it	Founder-led demos and conference meetings (Oracle AI World) [site]. A forward deployed engineer hire for on-site work [jobs]. Thought-leadership blog aimed at integrator founders [blog].
Price model	Not public [site]. Their own post discusses fixed bid, portfolio and gain-share pricing for integrators, not for Datafruit itself [blog].
The story	AI turns implementation from a relay race into a basketball team: fewer handoffs, less tax [blog]. Generic AI writes well but does not protect project margin [blog].

3. Positioning

Category they claim: "AI workspace for software implementations", "AI agents for system integrators" [site].
Category a YC partner would put them in: vertical AI for professional services firms: delivery memory and scoping for ERP and CRM integrators (inference).
What they want to be known for: the system of record for an implementation project, so scope and lessons never leak.
Where they sit on the map: layers L1/L2/L3 = 2/1/1, buyer = enterprise-facing integrators, business.

4. Three layers vs ours

Layer	Them (score + evidence)	Us (Marton's logs show)
L1 Company brain	2. Keeps engagement state and lessons across projects so later projects start sharper. Multi-tool capture. Memory is a lasting store other work reads, but it is one function (delivery), not the whole company [site] [blog].	A vault of 1,396 pages, 112 call transcripts and 8,763 customer quotes, searched by meaning. Built for one person on his own business.
L2 Agent orchestration	1. "Takes approved actions inside the systems" and has guardrails and access control. No public claim of a team of agents with handovers in the current product [blog] [site]. (The DevOps launch did describe sub-agents that hand off; that is the old product [press].)	About 5,641 jobs run by agents from three makers; each is checked by a different AI, then automatic tests, before it joins the main code.
L3 Execution and learning	1. Drafts deliverables and checks new requests against scope. No public measure of results and no claimed loop from outcomes back into the memory. Their benchmark tests extraction only [blog].	About 160 failure records, each with an automatic check. Offer feedback lines fell from 44 to 18 as 22 rules were added. Only about 15% of jobs reached the main code, and merges were unreliable in October.

5. Same and different

Topic	Same or different	Them	Us
Core belief	Same	Calls hold the truth about the work, and it is lost today [blog].	Same (brief).
Buyer type	Same	Services firms that sell expertise by the hour or project [blog].	W3 and W4 target consultancies and agencies (brief).
Buyer niche	Different	Software implementation partners, mostly Salesforce and Oracle world [site] [blog].	Not set. Brief names UK consultancies and agencies, SMEs and scale-ups.
Size	Different	Enterprise-facing, on-site engineer, on-prem installs [jobs].	Target: SMEs and scale-ups (brief).
Kind of work	Different	Captures, structures and drafts. Stops at scope and handover [site].	Marton's logs show agents also do and check work, on his own business only.
Owner of the result	Different	Firm owns a system of record [site].	We would own a number (needs proof). No customer yet.
Proof	Same (weak)	No named customer, no results [site].	No outside customer pays yet.
Security packaging	Different (they lead)	Role access, zero data retention option, SOC 2 in progress [site].	A client-data plan is open; no multi-tenant product yet.
Delivery	Different	Demo, then forward deployed engineer [jobs].	Two founders plus the system.
Price	Unknown	Not public [site].	Not set.

6. Wedge overlap

Wedge	Fit	Why
W3 Company memory from every call, for consultancies	direct	Same problem (calls become lasting memory) for the same kind of buyer (services firms). They are a vertical version for integrators [site] [yc].
W4 Fit call to tailored proposal in about 10 minutes	adjacent	They turn calls and notes into deliverables and scope for projects already won, not into a sales proposal [site] [blog].
W5 One layer for one type of small firm	adjacent	Same shape: one product for one firm type (integrators). Their firm type is bigger and more technical than a small accountancy [site].

7. Whitespace (where they stop)

Layer gap: they structure and draft; no public claim that agents run the work and measure it. To own it we need one customer workflow run for 4+ weeks with a before/after number.
Buyer gap: only software integrators are named, selling to enterprises, US-based with Las Vegas events [site] [blog]. UK consultancies, agencies and non-software services firms are not addressed (inference). To own it we need 3 to 5 such firms saying the pain is real.
Number gap: no customer result is published; the only figure is their own benchmark [blog]. To own it we need one number a firm would sign, such as hours of unbilled work caught or proposal time saved.
Price gap: price is hidden and the sale looks heavy (on-site engineer, on-prem) [jobs] (inference). To own it we need a price a 20 to 100 person firm can pay without a pilot.
Speed gap: not enough evidence either way. Skipped.

8. For the YC application

Proven today: Marton's logs show his system already turns calls into lasting company memory (112 call transcripts, 1,396 pages, 8,763 customer quotes) and that a fit call becomes a personal offer page in about 10 minutes (brief). Datafruit's public results name no customer and no number [site].
Draft answer to "how are you different from Datafruit?": DRAFT. "Datafruit records calls for Salesforce and Oracle integrators. We run the same call-to-memory loop on our own business already, and we are building it for smaller services firms, where the memory also drafts the next step." Founders say it their own way.
Future claim (needs proof): "Our memory learns from checked results and owns one number for a services firm." Needs one firm's workflow run for 4+ weeks with a before/after number.
What we understand that they don't: a services firm of any kind (not only software integrators) gains most when call memory feeds the proposal, the delivery and the check, in one loop. needs proof.
Copyability (M3): medium. The call-to-memory part is their core product, so they already have it [site] (easy to copy for them). The edge we would claim, a loop that does the next step and learns from checked results, is months of work, and chasing smaller, non-software firms pulls them off the enterprise integrator buyer their hiring and events point to [jobs] [site] (inference). Not hard: they are small and have already pivoted once [press] [blog], so they could move again. M3 can not score above 3 against Datafruit alone.
Why now (M4): None found in this teardown. Their blog cites an SAP CEO remark, an Accenture charge and TCS layoffs as drivers [blog]; those are their claims and I did not verify them, so they are not used here.
Do not claim: "we have no direct rival in services-firm memory" (Datafruit is one). "Our memory beats theirs" (no side-by-side test). "Smaller firms are underserved" without a named person who says so. Any claim that implies an outside customer.

9. Ideas worth taking

Publish a benchmark of your own pipeline. Their requirements-discovery post gives a number and a method buyers can argue with [blog]. We could do the same on our call data.
Run a research sprint with the buyer: 100+ calls with integrators in about two months became their lead article [blog]. It is also a cheap way to get named people with the pain.
Show up where the buyer already gathers: they book 30-minute meetings at Oracle AI World [site]. Our equivalent is a UK services-firm event.

10. Unknowns

Real team size: YC page says 2, earlier snapshots say 4 and list Tom Jeong. Check the YC page history or ask a founder.
Whether any customer is live. Search their case studies, Oracle AI World talks and partner directories (Salesforce, Oracle).
Price and deal size. Ask a buyer who has seen a quote.
Why they left DevOps, in their own words. No post explains it; the Launch HN comments only hint at it [press].
Funding detail: the $500k figure comes from directories that disagree on date (Jan 2025 vs Sept 2025) [press]. Treat as the YC deal (inference).
"Lychee" as a former name appears only in the YC data feed; I found no public source for it (UNVERIFIED).
Whether they sell in the UK or Europe.

Sources

datafruit.ai home [site]
Careers [jobs]
Oracle AI World event page, event 25 to 28 Oct 2026 [site]
What does purpose-built AI for delivery look like?, 31 Aug 2026 [blog]
Evaluating Agents on Requirements Discovery, 24 Aug 2026 [blog]
The tool decides the org chart, 4 May 2026 [blog]
The Partner Paradox, 28 Apr 2026 [blog]
T&M is dead (not really), 20 Apr 2026 [blog]
Patterns in Discovery from 100+ Salesforce Consultants, 9 Mar 2026 [blog]
YC company page [yc]
Launch HN: Datafruit (YC S25), AI for DevOps, mirror of the founders' Sept 2025 post, shows "1 year ago" [press]
Neuronfeed profile [press]
VCBacked profile [press]
Screen data: yc/teardowns/screen/candidates.jsonl and screen.jsonl (slug datafruit) [yc]

Datafruit

What they do

How close to us

Wedges

Who buys

Full teardown