Transcriptly

This Is How a 26-Year-Old Raised $108M in 1.5 Years | Reducto, Adit Abraham

EO13:00Added Aug 31, 2026

How do you go from manually labeling document boxes to processing over a billion pages for the world’s top AI companies? Adit Abraham is the Co-founder and CEO of Reducto, a Y Combinator (YC) alum that raised $108M from investors like Andreessen Horowitz and Ben

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00We did a ton of manual unsexy [music]
  3. 00:03work in the early days. Like we tried
  4. 00:05hiring an initial data labeling team and
  5. 00:07they weren't accurate enough. So I would
  6. 00:08spend a lot of my time just labeling
  7. 00:10boxes on documents. If you stacked all
  8. 00:12the pages that reduced processed, it
  9. 00:14would actually be something like 10
  10. 00:15times the height of Mount Everest back
  11. 00:17when we were still like fully
  12. 00:19unautomated for Stripe billing and
  13. 00:21setup. Like I would manually set up
  14. 00:23every single subscription. And those
  15. 00:24things, even though [music] they were
  16. 00:26repetitive, even though they were maybe
  17. 00:27boring in terms of the work that you're
  18. 00:29doing, were okay because the thing that
  19. 00:32I cared about is not like am I doing the
  20. 00:34most glamorous work. It was more so like
  21. 00:36[music] is the company moving forward?
  22. 00:37You're lucky to be able to do that
  23. 00:39because that means you're signing up a
  24. 00:40new customer. Like it is a privilege
  25. 00:42that you get to do that. Hi, my name is
  26. 00:44Ad Abraham. I'm the co-founder and CEO
  27. 00:46of Reductto. Reductto is a platform that
  28. 00:48helps AI teams parse, extract, and edit
  29. 00:50any sort of complex unstructured data
  30. 00:52for all sorts of language model use
  31. 00:54cases. Reduct has grown incredibly
  32. 00:56quickly. We've raised 108 million in
  33. 00:58total funding from incredible investors
  34. 01:00like Andre and Horowits, Benchmark, and
  35. 01:02First Round. Today, Reductive powers
  36. 01:04ingestion for some of the best companies
  37. 01:06in the world. Um, that includes really
  38. 01:08large Fortune 10 enterprises, but also
  39. 01:10newer leading AI companies like Harvey,
  40. 01:12Rogo, and Meror. To date, we've
  41. 01:14processed more than a billion pages for
  42. 01:16them and are continuing to grow every
  43. 01:17single week.

  44. Build What Customers Pull for Now, Not the Future

  45. 01:28Ever since I was young, I always used to
  46. 01:29have side hobbies. In high school, I saw
  47. 01:31an article that said something like the
  48. 01:33creator Flappy Birds making $50,000 a
  49. 01:35day on ad revenue. So, me and my best
  50. 01:37friend in high school just immediately
  51. 01:39had this gut reaction of, you know,
  52. 01:40forget school, forget all of that. we're
  53. 01:42just going to make apps and that's going
  54. 01:43to be our future. At some point, we even
  55. 01:44discussed not going to college. Things
  56. 01:46didn't work out that way. Uh we tried a
  57. 01:48few things but did end up going to
  58. 01:49college and you know pursuing a longer
  59. 01:51career from there. But I think it was a
  60. 01:53really nice inspiration that kind of
  61. 01:55showed how going off the beaten path
  62. 01:57[music] can lead to outlier outcomes for
  63. 01:59folks. Even though we don't work on game
  64. 02:00development today, I do think there's
  65. 02:02something very valuable about seeing
  66. 02:04individual effort that's you know
  67. 02:05sometimes just start as side projects
  68. 02:07spiral into something much much bigger.
  69. 02:09that eventually led to me going to MIT
  70. 02:11did my undergrad in computer science.
  71. 02:14I remember I was taking my first grad
  72. 02:16level ML course. Um so it was a course
  73. 02:18on metalarning like teaching models to
  74. 02:20learn and on the first day of the course
  75. 02:23the professor introduces Ronic who at
  76. 02:25this point is a freshman like it's his
  77. 02:27first week on campus probably and he
  78. 02:29frames it as hey everyone meet Ronic
  79. 02:31he's going to walk you through how to do
  80. 02:32the first pets. Um, so Ronic was a
  81. 02:34learning assistant um for this course
  82. 02:36that was primarily PhDs. And that was
  83. 02:38crazy to me. Like it was this person
  84. 02:39that even though he had just come on to
  85. 02:41campus um a campus with really smart and
  86. 02:43exceptional people, he was already at
  87. 02:45sort of the top um and so we became
  88. 02:47really close from there. The first time
  89. 02:49Ronic suggested that we could work on
  90. 02:50something together, that was an
  91. 02:52immediate yes for me. Like I didn't
  92. 02:53think twice about leaving my job or
  93. 02:55anything like that. He was just somebody
  94. 02:56that I admired enough for it [music] to
  95. 02:57just be a no-brainer. So in the course
  96. 03:00of the company before the YC batch we
  97. 03:02actually gave up on revenue multiple
  98. 03:04times. We tested different ideas got to
  99. 03:07a point where people were willing to pay
  100. 03:08for it but decided that the urgency with
  101. 03:11which they were willing to pay for it or
  102. 03:13like the need to which they wanted the
  103. 03:14product wasn't high enough for us to
  104. 03:16want it. And so just to give you a sense
  105. 03:18of what this looked like tangibly when
  106. 03:20we were selling remember all um we would
  107. 03:22constantly find you know at best people
  108. 03:24were willing to pay $50 a month or maybe
  109. 03:27$100 a month. So remember all as a
  110. 03:29product was at that time the first
  111. 03:31long-term memory API [music]
  112. 03:32for language models to remember things
  113. 03:34that you'd mentioned in the past. We
  114. 03:35would store context that [music] was
  115. 03:37important and retrieve it when it was
  116. 03:38relevant. As you would talk about things
  117. 03:40like implementation times, it was never
  118. 03:42the number one thing that they needed to
  119. 03:43focus on. And this was kind of one of
  120. 03:45those things that was nice [music] to
  121. 03:46have. In comparison, one of the things
  122. 03:48that we built for remember all is people
  123. 03:49would say, "Hey, you're managing the
  124. 03:51user's chat history. [music] Can you
  125. 03:53also manage the files that they upload?"
  126. 03:54Um almost like a managed drag service.
  127. 03:56And we saw that as you know a simple
  128. 03:59feature that we would add with
  129. 04:00off-the-shelf tools. [music] When we
  130. 04:01would demo remember all we would find
  131. 04:03that people would get really excited
  132. 04:05about the fact that we were managing the
  133. 04:06files that they uploaded. We had put so
  134. 04:08much time into making that [music] file
  135. 04:09management better. Um we started
  136. 04:11training our own models. Um we did a
  137. 04:14technical blog in YC's forum talking
  138. 04:15through how we segment [music] documents
  139. 04:17that wasn't packaged as you know a clean
  140. 04:20demo or anything like that. It was a
  141. 04:21really simple streamlit app. It was you
  142. 04:23would upload a documents and we would
  143. 04:24draw boxes on that document. And
  144. 04:26surprisingly here it was almost like
  145. 04:28they [music] were pulling us. They
  146. 04:29immediately started replying with, "Hey,
  147. 04:32these are better results than what I'm
  148. 04:33seeing from my existing [music] vendor.
  149. 04:35Is this a hosted API? Do you have a
  150. 04:36Stripe link? Like can I purchase this?
  151. 04:38Can I start using this?" It's almost
  152. 04:39like a slap in the face in terms of how
  153. 04:41much the market wants the product. And
  154. 04:43so when we were considering whether or
  155. 04:45not we should, you know, have high
  156. 04:47conviction in the space or not, that was
  157. 04:49the biggest thing that concerned [music]
  158. 04:50us. We knew that in a year, two years,
  159. 04:53three years, in some span of time, um
  160. 04:54long-term memory would need to exist.
  161. 04:56But what we wanted was to solve the
  162. 04:58problems that people needed [music]
  163. 04:59solved immediately to solve the things
  164. 05:01that they were actively looking for
  165. 05:02solution for. And we decided that
  166. 05:04remember all was not that.

  167. Build a Win-Win Product with Your Customer

  168. 05:10There are quite a few different ways
  169. 05:12that somebody can demonstrate how much
  170. 05:13your product means to them. It's not
  171. 05:15just the money, it's the time that
  172. 05:16they're willing to put into making the
  173. 05:18product great together. We've put a ton
  174. 05:20of time into, you know, aggregating
  175. 05:22data. Um, it's a big part of what we do
  176. 05:24and it's a big part of why we've been
  177. 05:25able to train state-of-the-art models,
  178. 05:27but production data is different. We
  179. 05:30work with really intensive financial,
  180. 05:32healthcare, insurance use cases that
  181. 05:34you're never going to find on the
  182. 05:35internet. And so, really quickly, we
  183. 05:37started having customers that, you know,
  184. 05:39had tried public documents and saw
  185. 05:41exceptional performance, but they would
  186. 05:43come to us with the most esoteric
  187. 05:45examples [music] imaginable. Like we've
  188. 05:47seen really hard cases where a doctor
  189. 05:49annotated things and you know they just
  190. 05:51put things at the bottom of the page and
  191. 05:53you were supposed to understand that it
  192. 05:54related to the thing at the top. We see
  193. 05:56really intensive financial tables with
  194. 05:58thousands of rows of data everything
  195. 06:01along those lines. The nice thing is our
  196. 06:03customers want us to solve those and so
  197. 06:05we've always had this almost design
  198. 06:07partner like relationship where they
  199. 06:09will come to us with that sort of
  200. 06:10feedback and we will iterate day after
  201. 06:12day after day to make the models better.
  202. 06:14And when you fix that feedback, they end
  203. 06:16up telling you whether or not that
  204. 06:17worked or it didn't. And you iterate by
  205. 06:20the end of that first week, you've
  206. 06:21already made a ton of progress with
  207. 06:22them. And that is really meaningful in
  208. 06:25that they care to make sure that your
  209. 06:27product is great. Like you're on the
  210. 06:28same team, you want to make the product
  211. 06:30better together because the work that we
  212. 06:32do directly helps them too. And so from
  213. 06:35the early days even to now, we would set
  214. 06:37up individual Slack channels with all of
  215. 06:39our customers. I have their phone
  216. 06:41numbers like we would call directly and
  217. 06:43if they ran into issues they would just
  218. 06:44call us um like they would tell us hey
  219. 06:46like this isn't working we need this for
  220. 06:48a big customer and we would work late
  221. 06:50into the nights to make sure that it was
  222. 06:51working for them because we don't take
  223. 06:53it lightly that people decided to trust
  224. 06:55us from an early stage. They have many
  225. 06:57reasons to not um they have all the
  226. 06:59reasons in the world to choose an
  227. 07:00established company that you know has
  228. 07:02been around for a decade and part of the
  229. 07:05way to pay back the trust that they've
  230. 07:07given us is to be there for them on an
  231. 07:10individual level. So even today you know
  232. 07:12if a company has an issue they can just
  233. 07:13pay Ronic or me directly. Part of what
  234. 07:16they're getting with Reduct is us as
  235. 07:17their ingestion team.

  236. Don't Explain, Show: Let Them See the

  237. 07:24There's a world where we just relied on
  238. 07:26marketing. Hey, it's the best product.
  239. 07:28Hey, it's state-of-the-art. All those
  240. 07:29things. But there are many companies
  241. 07:31that can say that. And on the flip side,
  242. 07:33the other thing that we could do is
  243. 07:34actually put the product in front of
  244. 07:36people even if it wasn't a perfect
  245. 07:37platform to let them see on their
  246. 07:39hardest documents that it works to prove
  247. 07:41what you're saying is true. And that
  248. 07:43translated to the company growing really
  249. 07:45quickly. At least in our case, being
  250. 07:47[music] public in that way um just meant
  251. 07:49that companies that otherwise probably
  252. 07:52would have ignored Reducto became really
  253. 07:54interested. Um like when we were
  254. 07:56twoerson company, [music]
  255. 07:57trillion dollar enterprise decided to
  256. 07:59book a demo and the reason why they
  257. 08:01booked a demo is because we had that
  258. 08:02public playgrounds where they uploaded
  259. 08:04hard documents that they'd seen fail on
  260. 08:06every other vendor. And once they saw
  261. 08:08that work, that justified reaching out.
  262. 08:10And if we hadn't done that, if we were
  263. 08:11this twoperson company of, you know,
  264. 08:1320some year olds, I find it hard to
  265. 08:15imagine that they would even be
  266. 08:17interested in engaging with us. Um, if
  267. 08:18we'd been shy about what we were
  268. 08:20building, we probably would have never
  269. 08:22gotten on the phone with them. When we
  270. 08:23say that we are the most accurate
  271. 08:24product in the market, we really mean
  272. 08:26it. Here are some examples, but if you
  273. 08:28want to see further, like you can test
  274. 08:29that for yourself. We've had companies
  275. 08:30[music] that I've tried to sell to two,
  276. 08:33three times and for one reason or
  277. 08:35another, they weren't sure if they
  278. 08:36could, you know, trust this early stage,
  279. 08:38seedstage company with what they were
  280. 08:40doing, even though they like the
  281. 08:41product. And what's interesting is
  282. 08:43pretty much all of those companies have
  283. 08:45since come back to us. Like they have
  284. 08:47come inbound saying, "Hey, we've been
  285. 08:49really impressed by the work that you've
  286. 08:50been doing. We see the progress that
  287. 08:52Reduct keeps making month over month."
  288. 08:54And they're ready to buy. Um and so as
  289. 08:56the company's grown, the companies that
  290. 08:57we struggle to sell to in year one, um
  291. 09:00we're fortunate to call customers today
  292. 09:02in year two.

  293. A Good Investor Stays When Things Get Tough

  294. 09:08I had known quite a few investors from
  295. 09:10just the course of building the company.
  296. 09:12And I think a lot of early stage
  297. 09:14founders think in terms of firm brand.
  298. 09:16um like they only think of tier one VCs
  299. 09:19as the actual firm which you know many
  300. 09:21of these firms have been around for
  301. 09:23decades and you know have their own
  302. 09:25reputation from [music] them but at the
  303. 09:27end of the day the thing that matters
  304. 09:29most is ideally whoever you're raising
  305. 09:31money from like that individual partner
  306. 09:33is somebody that you're going to be
  307. 09:34partnering [music] with for the next 10
  308. 09:36years. They're going to be there in all
  309. 09:38of your great successes like your future
  310. 09:40fundraising rounds when you close the
  311. 09:41great contracts but they'll also be
  312. 09:43there for the bad moments of the
  313. 09:44company. They'll be there when you have
  314. 09:45to, you know, let an employee go. Um,
  315. 09:47they'll be there when you lose a
  316. 09:49contract. They'll be there when you have
  317. 09:51a big, I [music] don't know, media
  318. 09:52incident, whatever could happen in the
  319. 09:54lifetime of the company. But it's really
  320. 09:57important to see how their interactions
  321. 10:00changed when things weren't going well.
  322. 10:02I remember there was a moment where Liz,
  323. 10:05our seed investor, actually basically
  324. 10:07never takes time off. If I text her at
  325. 10:0910:00 p.m., she's replying at 10:05.
  326. 10:11[music]
  327. 10:12She's getting on the phone like doing
  328. 10:13whatever. And one of the only moments
  329. 10:16where she was taking time for herself, I
  330. 10:18think she was at a Broadway show with
  331. 10:20her husband tragically. Like I I wish we
  332. 10:23hadn't done this. Um [music] but we had
  333. 10:25a opening eye outage at the same time.
  334. 10:28And so Veronica was like frantically,
  335. 10:30you know, messaging her like, "Hey, like
  336. 10:32what do we do? Our keys aren't working.
  337. 10:34Customers are upset." And even though it
  338. 10:36was one of the only times that she had
  339. 10:38to herself, she just immediately stepped
  340. 10:40out. She started calling people in her
  341. 10:42network and very quickly actually had
  342. 10:44the chief product officer at the company
  343. 10:46on the phone trying to help us with our
  344. 10:48issue and we were not an important
  345. 10:49enough customer for them to be doing
  346. 10:50that. These partners are committed to
  347. 10:53helping the company succeed and [music]
  348. 10:55that is really important. So if you're
  349. 10:56an early stage founder thinking about
  350. 10:57who to raise from, take the time to
  351. 11:00actually understand [music] what that is
  352. 11:01going to look like um because it's one
  353. 11:03of the most important decisions you'll
  354. 11:04have to make.
  355. 11:07I think with every moment that is really
  356. 11:10exciting in a company, the thing that
  357. 11:12isn't discussed in interviews is [music]
  358. 11:14what it took to get to that moment. Um,
  359. 11:16like when we were landing our first
  360. 11:18really big enterprise contract, it was
  361. 11:20an on-prem deployments and we had never
  362. 11:21done an on-prem deployment before. You
  363. 11:23know, we [music] didn't have
  364. 11:24infrastructure engineers on team, we
  365. 11:26weren't this large org that could divvy
  366. 11:29up responsibilities. We would wake up,
  367. 11:31we would immediately go to the office
  368. 11:33and we would be in the office until
  369. 11:34[music] we were too exhausted to
  370. 11:35continue working. We would sleep for at
  371. 11:37most a few hours and then we would go
  372. 11:39back and we would try again and again
  373. 11:40and again. People diving in and doing
  374. 11:43anything [music] at the company. There's
  375. 11:45no sort of notion of hey if you're an
  376. 11:48engineer you don't need to do customer
  377. 11:49support. There's no notion of like hey
  378. 11:52you know if you're an ops person you
  379. 11:53don't need to label data for the ML team
  380. 11:56because everybody just wants to see the
  381. 11:58company succeed. Um, and the company
  382. 11:59succeeds when all of these things work,
  383. 12:01when the product works, [music] when
  384. 12:02customers are happy. And people don't
  385. 12:03think of their job in terms of whatever
  386. 12:05their role title is. They think of it in
  387. 12:07the capacity that they [music] can help
  388. 12:08the company move forward. And so what I
  389. 12:10see reductive as it, it's not really
  390. 12:12just parsing. It's what does it mean to
  391. 12:15have this layer that connects human data
  392. 12:17to this new level of intelligence that
  393. 12:20applies across all of that data. Um,
  394. 12:22we're seeing products built with
  395. 12:24productto today that don't just read the
  396. 12:26documents, they actually create net new
  397. 12:27documents for their end customers. Like
  398. 12:29they do end-to-end work with agentic um,
  399. 12:31[music] workflows. In the future, most
  400. 12:34AI products will be some component of
  401. 12:36intelligence. That is what the
  402. 12:37foundation model companies provide, but
  403. 12:39it will be some components of context as
  404. 12:41well. [music] And we want reductus to be
  405. 12:42the best way that you interact with that
  406. 12:44context, like a building block that you
  407. 12:46aggregate together and apply [music] it
  408. 12:47to a specific use case.
This Is How a 26-Year-Old Raised $108M in 1.5 Years | Reducto, Adit Abraham — Transcriptly