The Factory That Dreams: 39 AI Agents, No Framework - Rushabh Doshi, Machinecraft

AI Engineer · 9 min · 155 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

1. 00:00Okay, I want to tell you a story about a factory that taught itself how to remember.
2. 00:07Hi, I'm Rushabh.
3. 00:08I run Machinecraft, a 100 people factory in India.
4. 00:11No data science team, no ML budget, none of that.
5. 00:15And somehow we ended up building a 36 AI agent that runs our entire go-to-market.
6. 00:21I think that's still a little ridiculous.
7. 00:24Let me show you how it happened and why you can do the same thing.
8. 00:30So, here's the thing about our company.
9. 00:32From the outside it looks like machines and metal.
10. 00:34But the actual company, the part that matters, isn't the machines, it's the knowledge.
11. 00:40Who the customer is, what we quoted them in 2019, why that one machine needed that weird custom tweak.
12. 00:47And for three generations, all of that lived in exactly three brains.
13. 00:51Initially my grandfather's, then my father's, and now mine.
14. 00:57Which is a genuinely terrifying way to run a company.
15. 01:00When you sit with it.
16. 01:02A lot of people have joined us, people have left us, the revolving door never stopped.
17. 01:08And every single time someone walked out, a chunk of our brain walked out with them.
18. 01:15We weren't scared of the competitors, we were scared of forgetting.
19. 01:18Or waking up one day and realizing the whole company only existed inside two increasingly tired heads.
20. 01:27So, I had an idea.
21. 01:29I'll be honest.
22. 01:31Sounded insane first.
23. 01:32But what if instead of writing the knowledge down in some document nobody ever reads, what if we grew a brain that just held it?
24. 01:42Not a chatbot you poke at, a twin of the company.
25. 01:46I didn't hire a sales team.
26. 01:48I tried to build one.
27. 01:50A quick detour because you need to know how messy this is.
28. 01:55We make thermoforming machines.
29. 01:57They heat up a plastic sheet and shape it.
30. 02:01Same core machine, but it ends up making hydroponic farm trays, spa bathtubs, EV car panels, medical casings, and even packaging.
31. 02:10Seven totally different worlds, seven totally different buyers.
32. 02:14So, this brain couldn't just memorize a brochure.
33. 02:18It had to know which universe a given customer lives in.
34. 02:24Step one was almost boringly simple.
35. 02:27Feed it everything, and I mean everything.
36. 02:30Years of quotes, drawings, payment schedules, timelines, email threads, hundreds of gigabytes of our own private history.
37. 02:38Not the public internet, our internet.
38. 02:42And here's the plot twist.
39. 02:44The part that surprises every engineer I tell this to.
40. 02:48We never trained a model.
41. 02:50No GPUs humming in the basement, no fine-tuning.
42. 02:53We just looked at all the history, chopped it into bite-size chunks, and let off-the-shelf models read it and pull out the facts.
43. 03:02We stored the meaning of each chunk as vectors and relationships.
44. 03:07Who's connected to what as a graph.
45. 03:10The brain isn't a smarter model.
46. 03:12It's actually a really, really well-organized memory.
47. 03:17Now, this is where it gets a little weird in a good way.
48. 03:22We stopped thinking of Era as a software and started thinking of it as something we were raising.
49. 03:27So, we gave it a body modeled on biology.
50. 03:30Senses to figure out who it's talking to, a gut to digest the documents into facts, a memory, a dream cycle, an immune system to fight off bad information.
51. 03:42Why biology?
52. 03:43Well, because evolution already spent a billion years solving how do you stay coherent over time?
53. 03:50We just copied the homework.
54. 03:52Okay, so the big question.
55. 03:54Why 36 agents instead of one genius mega prompt?
56. 03:58Because, and you already know this if you've ever tried it, one prompt that's supposed to do everything ends up doing everything badly.
57. 04:07So, Ira isn't one mind, it's a pantheon.
58. 04:10A whole cast of specialists.
59. 04:13Each one has exactly one job.
60. 04:16Athena runs the room.
61. 04:18Prometheus owns the sale.
62. 04:21Plutus does pricing.
63. 04:23Hephaestus knows every machine spec cold.
64. 04:27Vera fact checks everything, and Memnon, my favorite, guards corrections.
65. 04:33So, the second a human fixes something, it stays fixed forever.
66. 04:38One agent, one job.
67. 04:40It's a team, not a hero.
68. 04:43And here's the cool part.
69. 04:45They hold meetings.
70. 04:46Athena pulls in specialists.
71. 04:49They actually argue, and a single answer comes out the other side.
72. 04:53It's like having a board room that never sleeps, never gets tired, and somehow has no ego.
73. 05:00So, what does all this actually run?
74. 05:03Honestly, the whole front business.
75. 05:06Everything between a stranger exists somewhere, and now they're a customer.
76. 05:13Nine concrete jobs every single day.
77. 05:15Outbound emails that actually reference my real world.
78. 05:19Account briefs built from cross-checked truths before a call.
79. 05:24Quotations.
80. 05:25A swipe left, swipe right mode for outreach.
81. 05:29Reviving dead leads, which I call blast from the past.
82. 05:34Inbound replies, and figuring out before we waste an hour whether a company is even a fit.
83. 05:40Nine jobs, one operator who never sleeps.
84. 05:45Where does all this live?
85. 05:47One cursor tab.
86. 05:48That's genuinely it.
87. 05:50You type and Eira reaches out with a dozen hands.
88. 05:53Searches the knowledge base, reads the inbox, drafts the email, builds the code, and then shows you before anything actually goes out.
89. 06:03Under the hood, it's genuinely a real stack.
90. 06:06Not a demo held together with duct tape.
91. 06:08Databases for vectors, for relationship graph, for the CRM.
92. 06:13Three different model providers, each picked for the job it's actually best for.
93. 06:17Tools for Google, for swallowing documents, for every communication channel, plus monitoring.
94. 06:23So, we can see what it's thinking.
95. 06:26All of it, every capability exposed as 213 tools over one protocol.
96. 06:32And the golden rule, the one we never break, Eira drafts, human sends.
97. 06:38Now, memory.
98. 06:40And this is the part where most AI quietly lies to you, because a raw language model is basically a goldfish.
99. 06:46Brilliant for about 30 seconds, and then you close the tab and forgets you ever existed.
100. 06:52So, we engineered memory on purpose, in layers.
101. 06:55Working memory for the last few minutes.
102. 06:58Pinned facts about someone who who is.
103. 07:02Episodes, whole conversations as little stories.
104. 07:05Relationships with warmth that grows from stranger to trusted.
105. 07:10And a bouncer at the door.
106. 07:12A salience gate that decides what's even worth remembering, so the brain doesn't fill up with junk.
107. 07:19When two facts disagree, corrections win.
108. 07:22Continuity without making things up.
109. 07:26And then, I genuinely love this part.
110. 07:28At night, it dreams.
111. 07:32Every night, Eira runs a sleep cycle.
112. 07:35It replays the day, locks in useful stuff, hunts for contradictions, gently forgets the stale junk, and turns the day's work into reusable skills.
113. 07:45In the morning, there's a little dream report waiting for me to read.
114. 07:49Here's what I consolidated.
115. 07:51Here's what I Here's what I let go of.
116. 07:54Here's what I figured out while you were asleep.
117. 07:56The thing literally gets smarter overnight.
118. 08:01And here's the part I care about the most.
119. 08:03Every agent has a conscience.
120. 08:05And it is emphatically not to be helpful, be harmless.
121. 08:12It's a soul file written from the principles of a Jain family business that's been doing this for the last three generations.
122. 08:20Five old ideas turn into engineering rules.
123. 08:23No single source has the whole truth.
124. 08:25So, cross-check before you speak.
125. 08:28Never say things absolutely.
126. 08:30Cite the document and the date.
127. 08:33Do your own job, not someone else's.
128. 08:35Report the truth even when the truth is ugly.
129. 08:38And nobody works alone.
130. 08:40Ancient philosophy running as guardrails in production.
131. 08:44Now, let's talk money.
132. 08:46Because this is the part that should make the whole industry a little uncomfortable.
133. 08:50There was no training bill.
134. 08:51Zero.
135. 08:53The expensive part was never compute.
136. 08:54It was teaching a company to remember itself.
137. 08:57An agency quoted us 230 grand to build this.
138. 09:01We built it for around 30.
139. 09:03That's cheaper than a nice watch.
140. 09:05And it runs on a couple of thousand dollars a month.
141. 09:09So, here's the move.
142. 09:11We pulled the whole architecture out and made it forkable.
143. 09:14We call it Brain OS.
144. 09:16It ships as an empty nervous system.
145. 09:18The agents, the memory, the dream cycle, the soul file.
146. 09:22All there, completely blank.
147. 09:24You pour your own company's truth into it and from inside out.
148. 09:28Because here's the thing nobody can outsource for you.
149. 09:32Only you can build your company's brain.
150. 09:37We are a 100 people factory with no data scientists.
151. 09:40If we can grow a brain, you can too.
152. 09:42We are not selling ours to you.
153. 09:44We are helping you build your own.
154. 09:46forkmybrain.org Go build something that remembers.
155. 09:51Thank you.