Videos Tt2kX2sgQio

How Kepler Built Verifiable AI for Financial Services — Vinoo Ganesh

index_state ready data_status ok

AI Engineer· published 2026-07-29· 0:22:29· en-US· indexed 2026-08-10 19:42

Open on YouTube

50 shot(s).

keyframes kept every frame deduplicated

What was stored

cues: 249; whisperx 249
chunks: 40; from 249 cues
keyframes: 27; kept of 50 captured
frames with text: 27; 455 lines read
chapters: 11; from the source metadata
keyframe bytes: 5.5 MB; word timings on 249 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage	state	model	started	took
`fetch`	done	—	2026-08-09 23:21	0s
`stt`	done	—	2026-08-09 14:05	29s
`chunk`	done	—	2026-08-09 14:06	0s
`text_embed`	done	—	2026-08-10 19:42	1s
`keyframe`	done	—	2026-08-09 14:06	3m 01s
`ocr`	done	—	2026-08-09 14:09	12s
`frame_embed`	done	—	2026-08-10 19:42	5s

Frames, and what the machine read

0:02 #0 done2 line(s)

shot 0·sharpness 453.8
1. AlEngineer0.96
2. World's Fair0.97
0:03 #1 done2 line(s)

shot 1·sharpness 658.6
1. AlEngineer0.95
2. World's Fair0.99
0:10 #2 done24 line(s)

shot 2·sharpness 2731.8
1. LAB & PLATINUM SPONSORS0.99
2. Amazon AGI Lab0.98
3. ANTHROP\C1.00
4. Google DeepMind1.00
5. MINIMAX0.92
6. OpenAI0.92
7. Akamai1.00
8. arize1.00
9. aws1.00
10. Braintrust bright data0.99
11. B1.00
12. Browserbase1.00
13. docker1.00
14. :neo4j0.91
15. ORACLE1.00
16. PayPal1.00
17. qodo1.00
18. reducto1.00
19. Sonar1.00
20. Makers of0.99
21. togetherai1.00
22. Unblocked1.00
23. WorkOS1.00
24. SonarQube1.00
0:22 #3 done16 line(s)

shot 3·sharpness 1461.8
1. Kepler0.99
2. AlEngineer0.98
3. 2026- CONFIDENTIAL0.95
4. World'sFair1.00
5. JUNE 20260.97
6. tamings Call0.91
7. How Kepler built verifiable Al for1.00
8. financial services1.00
9. A framework for trusting AI-produced work1.00
10. 710€1g0.57
11. VERIFIED OUTPUT0.98
12. Vinoo Ganesh, CEO and co-founder,1.00
13. Kepler1.00
14. KEPLER.AI0.93
15. Engineering the future of Al0.98
16. World'sFair0.99
0:29 #4 done22 line(s)

shot 4·sharpness 1894.9
1. Kepler1.00
2. AlEngineer0.99
3. World'sFair1.00
4. About me1.00
5. Vinoo Ganesh CEO and co-founder, Kepler1.00
6. PRESENTED BY1.00
7. Palantir0.94
8. Verast1.00
9. Microsoft1.00
10. Led Compute / Spark. SOCOM, Maven,0.98
11. CTO of an altdata startup. Acquired.0.99
12. and Palantir Frontline (Palantir FDE1.00
13. Program)1.00
14. Citadel1.00
15. Advisor and builder1.00
16. Head of Business Engineering (FDE).1.00
17. Databand, Bluesky, Horangi, and Sync. All1.00
18. acquired.1.00
19. KEPLER.AI1.00
20. Kaplar0.72
21. Engineering the future of Al0.99
22. World'sFair1.00
1:06 #5 done18 line(s)

shot 5·sharpness 2218.8
1. Kepler0.97
2. AlEngineer0.98
3. World'sFair1.00
4. Where this talk comes from1.00
5. This talk is based on Anthropic's case study on1.00
6. Kepler's verifiable and deterministic infrastructure.0.99
7. PRESENTED BY1.00
8. Microsoft1.00
9. Anthropic customer profile – read the full version:0.99
10. claude.com/blog/how-kepler-built-verifiable-ai-for-financial-services-with-claude1.00
11. 26M SEC filings0.98
12. 14,000+ companies0.99
13. 27 markets1.00
14. built in < 3 months1.00
15. KEPLER.AI1.00
16. TRACK 3· JULY 2, 20260.94
17. Alin Finance0.96
18. World'sFair1.00
2:00 #6 done30 line(s)

shot 6·sharpness 1891.5
1. Kepler1.00
2. AlEngineer0.98
3. World's Fair0.96
4. Look at this track1.00
5. Most of this room is working on producing more. Almost none on0.99
6. trusting it.1.00
7. 10:451.00
8. Autonomous multi-agent research across optimization domains1.00
9. 11:101.00
10. Why off-the-shelf AI doesn't understand money0.99
11. 11:401.00
12. Let's integrate AI agents in event-sourced systems1.00
13. 12:051.00
14. How Kepler built verifiable AI for financial services ·you are here0.99
15. 1:301.00
16. Build for the memo, not the demo1.00
17. 1:551.00
18. We vetted 2,000 AI skills before they reached developers · review0.98
19. 2:251.00
20. Your finance agent's bottleneck is you · review0.99
21. 2:500.99
22. Simulation-maxxing, shipping agents faster with simulations ·review0.99
23. 3:201.00
24. Skills are new features, building a skill-centric harness · review0.98
25. 3:451.00
26. Wearing the agent, a family-and-friends personal agent0.99
27. KEPLER.AI1.00
28. TRACK 3· JULY 2, 20260.95
29. Alin Finance0.96
30. World's Fair0.99
2:14 #7 done15 line(s)

shot 7·sharpness 2491.7
1. Kepler1.00
2. AlEngineer0.96
3. World'sFair1.00
4. Al turned a writing problem into a reading problem1.00
5. For decades the hard part was producing the work. AI made producing nearly free.0.99
6. The old edge was the headcount to read everything — the ultimate form of alpha0.99
7. decay.1.00
8. When everyone's model reads everything, what stays yours is your ontology.1.00
9. Now the hard part is trusting what got produced.1.00
10. produce – the tooling is everywhere0.98
11. verify — almost nothing points here0.97
12. KEPLER.AI1.00
13. TRACK 3· JULY 2, 20260.95
14. AlinFinance1.00
15. World'sFair1.00
2:49 #8 skipped

shot 8·duplicate of #7
3:22 #9 done29 line(s)

shot 9·sharpness 2106.0
1. Kepler1.00
2. AlEngineer0.98
3. World'sFair1.00
4. Every system finance runs is read-only1.00
5. Bloomberg and FactSet look like data. They're closer to insurance - culpable data you can point at when a1.00
6. number is wrong.0.98
7. Read-only tools never produced your work, which is why every analyst still rebuilds the DCF from scratch.0.99
8. Software1.00
9. Medicine1.00
10. Aviation1.00
11. Finance1.00
12. work1.00
13. work1.00
14. work1.00
15. work1.00
16. pharmacist double-1.00
17. CI+ review0.92
18. check1.00
19. pilot + copilot0.96
20. ?0.99
21. ships1.00
22. ships1.00
23. ships1.00
24. ships anyway0.99
25. Finance never built its version of verification.1.00
26. KEPLER.AI1.00
27. TRACK 3· JULY 2,20260.97
28. AlinFinance1.00
29. World's Fair0.97
3:42 #10 skipped

shot 10·duplicate of #9
4:17 #11 skipped

shot 11·duplicate of #9
4:43 #12 done21 line(s)

shot 12·sharpness 2431.6
1. Kepler0.99
2. AlEngineer0.98
3. World'sFair1.00
4. Why verification matters in finance1.00
5. AI can produce a confident, well-written number that is simply wrong.0.99
6. Verification means the number traces to its source and gets1.00
7. checked, not trusted because it reads well.1.00
8. Regulation requires it1.00
9. Decisions depend on it0.99
10. Filings and research must be auditable. A0.99
11. Capital rides on foundational entities: the0.99
12. number you cannot trace is one you1.00
13. numbers, the companies behind them,0.99
14. cannot defend to a regulator, an1.00
15. and the directors who run them. Each has1.00
16. examiner, or an investment committee.1.00
17. to be right, and provably so.1.00
18. KEPLER.AI1.00
19. TRACK 3• JULY 2, 20260.95
20. Al in Finance0.96
21. World'sFair1.00
5:33 #13 done23 line(s)

shot 13·sharpness 2563.1
1. Kepler0.99
2. AlEngineer0.98
3. World'sFair1.00
4. Why verification matters in finance1.00
5. AI can produce a confident, well-written number that is simply wrong.0.99
6. Verification means the number traces to its source and gets1.00
7. PRESENTED BY1.00
8. checked, not trusted because it reads well.1.00
9. Microsoft1.00
10. Regulation requires it1.00
11. Decisions depend on it0.99
12. Filings and research must be auditable. A0.99
13. Capital rides on foundational entities: the0.99
14. number you cannot trace is one you0.99
15. numbers, the companies behind them,1.00
16. cannot defend to a regulator, an1.00
17. and the directors who run them. Each has1.00
18. examiner, or an investment committee.1.00
19. to be right, and provably so.1.00
20. KEPLER.AI1.00
21. TRACK 3 • JULY 2, 20260.93
22. Al in Finance0.98
23. World'sFair0.98
5:49 #14 done22 line(s)

shot 14·sharpness 1759.5
1. Kepler0.99
2. AlEngineer0.99
3. World's Fair0.99
4. A citation is not a verification0.99
5. It does not confirm the number matches the source.0.99
6. PRESENTED BY1.00
7. A citation0.99
8. A verification1.00
9. Microsoft1.00
10. names where a number came from0.99
11. confirms the value equals the source1.00
12. claim·$416,161M1.00
13. $416,161M1.00
14. the 10-K1.00
15. one document, thousands of numbers0.99
16. source·$416,161M1.00
17. value never checked1.00
18. values match·value checked1.00
19. KEPLER.AI1.00
20. TRACK 3· JULY 2, 20260.95
21. Al in Finance0.97
22. World's Fair0.96
6:15 #15 skipped

shot 15·duplicate of #14
6:51 #16 skipped

shot 16·duplicate of #14
7:16 #17 done18 line(s)

shot 17·sharpness 1585.7
1. Kepler0.99
2. AlEngineer0.97
3. World'sFair0.96
4. Verification lives in the path, not the answer1.00
5. Same number. Different trust.1.00
6. Pulled from the 10-K, periods0.99
7. matched1.00
8. Analyst A0.99
9. 26.9% net margin1.00
10. Analyst B1.00
11. From a summary slide, eyeballed,1.00
12. rounded1.00
13. solid green = sourced and checked1.00
14. dashed red = eyeballed1.00
15. KEPLER.AI1.00
16. TRACK 3· JULY 2, 20260.96
17. Al in Finance0.96
18. World'sFair1.00
7:45 #18 skipped

shot 18·duplicate of #17
8:09 #19 skipped

shot 19·duplicate of #17
8:55 #20 skipped

shot 20·duplicate of #17
9:35 #21 done22 line(s)

shot 21·sharpness 1672.2
1. Kepler1.00
2. AlEngineer0.99
3. World's Fair0.98
4. The framework, three principles1.00
5. model reasons · deterministic code retrieves, computes, cites · ontology maps words to line items0.98
6. 011.00
7. 021.00
8. 031.00
9. Atomic1.00
10. Scoped1.00
11. Derivation1.00
12. provenance1.00
13. determinism1.00
14. chains1.00
15. binds to one datapoint1.00
16. model never computes0.99
17. from verified inputs1.00
18. Each verifies to a single datapoint1.00
19. KEPLER.AI1.00
20. TRACK 3· JULY 2,20260.97
21. Al in Finance0.99
22. World's Fair0.98
9:56 #22 skipped

shot 22·duplicate of #21
10:28 #23 done25 line(s)

shot 23·sharpness 1968.2
1. Kepler0.99
2. AlEngineer0.99
3. World's Fair0.94
4. What atomic provenance buys you1.00
5. The model writes a value at an address0.99
6. Provenance ledger1.00
7. PRESENTED BY0.97
8. f-78 = $416,161M0.99
9. Ships1.00
10. Microsoft1.00
11. on match1.00
12. Model writes a1.00
13. Deterministic check1.00
14. number1.00
15. does the number match?0.99
16. $416,161M· cites f-780.98
17. Stripped1.00
18. 94% first-pass line-item match0.98
19. on mismatch1.00
20. frontier models alone: 38-46%1.00
21. A wrong number is stripped before it reaches the analyst.0.99
22. KEPLER.AI1.00
23. TRACK 3· JULY 2, 20260.94
24. Al in Finance0.99
25. World's Fair0.98

Transcript

249 cues· 3,798 words· 21,208 chars

chunk 0 · 0:12–0:56 · 137 words · 761 chars
0:12 Hey, everyone.
0:13 Thank you for being here.
0:14 My name is Vinu Ganesh, and I'm the CEO and co-founder of Kepler.
0:18 Today, I'm going to talk to you about how we built verifiable AI for financial services.
0:24 First, a little about me.
0:25 My career has been working in fairly difficult places to work in terms of numerical accuracy and verifiability.
0:33 Began my career at Palantir, where I led the compute platform, as well as a lot of our USG engagements.
0:38 Built and sold a data startup, then was head of business engineering at Citadel.
chunk 1 · 0:43–1:27 · 143 words · 748 chars
0:43 I have some Citadel colleagues here in the audience as well.
0:46 I've advised a bunch of startups and been lucky that all of them reached pretty positive outcomes.
0:52 So this whole talk is a distilled version of a case study Anthropic did on Kepler.
0:56 If you scan that QR code, we're the only company they've ever done a case study of, which is kind of cool.
1:03 And so this is a distilled version of that that has a lot more information about how we were able to do what we do, why it's been pretty impactful in financial services, and really where we go from here.
chunk 2 · 1:15–1:58 · 128 words · 696 chars
1:15 So my only goal with this whole talk is to convince you that AI is going to start doing some very powerful things in terms of producing work product.
1:24 So everything that I tell you is inevitable.
1:27 Whether it's Kepler or whether it's anyone else, we are already on this journey and this trajectory.
1:33 So we all better be ready.
1:36 So I first want to observe every talk that we've seen in the financial services space so far has been about producing more.
1:43 Like how do I token max?
1:45 How do I get AI to do more?
chunk 3 · 1:47–2:27 · 98 words · 600 chars
1:47 Very little of it is about how to actually make AI trustworthy or produce trustworthy products.
1:54 And what's interesting is even terms like trust and verifiability have actually been largely abused.
1:59 Evals are not verifiable.
2:01 You cannot take a non-deterministic LLM and eval your way to something deterministic.
2:06 These are probability machines.
2:09 And the kind of underlying reason for this is that AI has made a writing problem a reading problem.
2:15 We can produce insane amounts of content, whether that's code, whether it's marketing, whether it's like a DCF in record time, but we can't easily verify this.
chunk 4 · 2:27–3:11 · 132 words · 730 chars
2:27 And that's because for years, the hardest part about this whole process was actually producing the work.
2:33 Edge and alpha came from people like Sutterdell being able to hire hundreds of analysts who could scour the internet and understand where any source of alpha could exist.
2:43 So it really came from this idea of being able to consume content.
2:48 The problem is when a model reads everything, you have the most real version of alpha decay that you possibly can.
2:54 There is no edge if everyone can look at Tegas and get all the same information.
chunk 5 · 2:59–3:41 · 120 words · 670 chars
2:59 So the hard part now is trusting what actually got produced by the model.
3:04 And this is kind of funny.
3:05 This is not necessarily a finance problem.
3:08 Every system that exists has some form of this.
3:12 Software, we run CICD.
3:15 We do unit tests.
3:16 We do integration tests.
3:17 We have code reviews.
3:19 When a doctor writes a prescription, a pharmacist fills that prescription.
3:23 So if it says 10,000 milligrams of a medication, someone catches that.
3:28 We have a pilot and a co-pilot.
chunk 6 · 3:30–4:14 · 128 words · 738 chars
3:30 We have an EMT that's a primary EMT and a secondary.
3:33 In finance, we have maybe an overworked VP as a verification layer, but that concept doesn't really exist.
3:40 Now I'm going to say something even more aggressive.
3:43 The reason that people buy products like Bloomberg and FactSet is to displace culpability.
3:49 When you buy a tool like that, you know that information's free.
3:52 It exists in SEC filings.
3:54 But you believe that because a bunch of contractors or folks overseas vetted this data and stuck it in a central instance, at least if it's wrong, everyone on Wall Street has the same incorrect information.

Chapters

0:00 Introduction: a data background in finance
1:42 Why trust and verifiability matter now
2:57 Models are probability machines
4:27 Why analysts still put in the hours
8:22 Modeling AI like an overworked VP
9:26 Atomic provenance
12:01 Scope determinism
13:52 Reconciliation and pulling real numbers
15:06 Extracting entities without misses
16:34 Toward zero invented securities
20:31 Where a number really comes from