# Building Reliable, yet Smart Workflows for Finance, Wealth and Insurance

Karthic Rao · AI Engineer NYC 2026 · October 13

[Watch recording](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15828s)

[04:23:49](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15829s) Hello, I'm Karthic, CTO at Perpetual.

[04:23:52](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15832s) In this presentation, I'll talk about building workflows that are both reliable and intelligent for finance, wealth and insurance.

[04:23:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15839s) In regulated domains, workflows must be reliable enough to meet regulatory standards while still being intelligent enough to handle the real world.

[04:24:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15847s) In fields such as finance and wealth management, this presents both a technical and a compliance issue.

[04:24:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15852s) Reliability and intelligence are not opposites.

[04:24:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15854s) It is up to you where you draw the boundary between them.

[04:24:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15857s) And that is the point of this talk, deciding where that boundary lies.

[04:24:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15861s) For us, it's not merely a thought experiment, but a question that surfaces serious consequences.

[04:24:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15866s) Perpetual is a wealth management roll-up.

[04:24:29](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15869s) We acquire running traditional wealth management firms.

[04:24:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15872s) And we maximize their throughput and revenue with the help of AI.

[04:24:37](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15877s) Wealth management is a regulated industry.

[04:24:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15879s) And that is where we build with AI in production.

[04:24:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15882s) Everything in this talk is grounded in building for those constraints.

[04:24:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15887s) Workflows in finance and wealth management come with four facts,

[04:24:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15890s) and each of them pushes the architecture in the same direction.

[04:24:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15894s) First, they're complex and they need runtime strategies as the process unfolds.

[04:24:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15898s) You can't script every path in advance.

[04:25:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15901s) Second, the data is sensitive, so you need a path to move these workflows entirely onto privately hosted models.

[04:25:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15907s) Third, in a highly regulated industry, a policy violation isn't a bug report.

[04:25:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15912s) It has severe consequences.

[04:25:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15914s) So, fourth, execution has to be highly reliable and deterministic.

[04:25:19](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15919s) Inject intelligence only where runtime judgment is genuinely required

[04:25:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15923s) and minimize reliance on LLMs everywhere else.

[04:25:28](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15928s) AI is spreading through regulated industries,

[04:25:30](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15930s) but there's still no consensus on an architecture

[04:25:33](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15933s) that gives you high reliability with the least use of LLM intelligence.

[04:25:37](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15937s) Meanwhile, most of what developers hear from the big labs is how to burn more tokens.

[04:25:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15942s) But in a regulated industry, token maxing is the wrong goal.

[04:25:46](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15946s) These tasks are close to main.

[04:25:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15948s) You can lay them out in advance and build them into a massive state machine.

[04:25:52](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15952s) And that machine needs only a sprinkle of intelligence to navigate it.

[04:25:55](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15955s) Most steps just follow the next edge.

[04:25:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15958s) Only at a few judgment points does the model pick a branch.

[04:26:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15961s) So the intelligent way to build is to use intelligence minimally and on purpose.

[04:26:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15968s) Let's make this concrete with a real wealth management workflow.

[04:26:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15972s) A wealth advisor turns to their assistant and says one sentence.

[04:26:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15976s) Fill out the IRS W-9 for Mr. Harper to a person that's a simple instruction.

[04:26:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15982s) Watch what it actually triggers when a human being executes it.

[04:26:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15986s) The task has complexities that go beyond what meets the eye.

[04:26:29](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15989s) First, the assistant finds form W-9 from the firm's records or from the IRS website.

[04:26:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15994s) Then they look for Mr Harper in the system of record and find

[04:26:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=15998s) three Harper households. So they call the advisor. Which Harper

[04:26:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16002s) do you mean? Only once that ambiguity is resolved do they move

[04:26:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16005s) on to the form. A lookup, an ambiguity and a human in the loop

[04:26:49](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16009s) all in the first minute of a one line task.

[04:26:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16017s) Filling a form sounds like data entry. It isn't. Every form is a

[04:27:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16022s) small program. A W9 first decides who the taxpayer is, the person or the owner if it's a disregarded

[04:27:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16029s) entity. A foreign person stops right here because they need a W8 instead. Then how they're taxed?

[04:27:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16036s) Exactly one box on line 3A and an LLC adds its letter. That decides which name goes on line 1

[04:27:24](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16044s) and which on line 2. Then whose number? An SSN or an EIN. Never both. Taken from a verified document.

[04:27:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16052s) never from memory. Exemption codes apply only to some entities and never to an S-corporation.

[04:27:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16058s) And finally, who signs under penalties of perjury? Each answer changes the next question

[04:27:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16063s) and the people involved only ever see three kinds of gates, a question, a document to collect, or

[04:27:49](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16069s) a signature. Now they start filling the form and realize the system of record is missing three of

[04:27:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16079s) the values it needs. They check the documents on file and the system of record again still

[04:28:04](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16084s) missing so they start an email thread with the client asking for exactly those three

[04:28:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16089s) values. The task now depends on someone outside the firm on their timeline.

[04:28:19](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16099s) now the workflow waits. The request is with the client and all three values are still

[04:28:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16103s) missing. If the client doesn't reply within two days, the assistant follows up and brings

[04:28:28](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16108s) it to the advisor at the daily stand up. If the client replies with only two of the three values,

[04:28:33](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16113s) they follow up again for the missing one. When all three arrive, they complete the form and write

[04:28:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16118s) the new information back to the system of record, including the trail of how they obtained those

[04:28:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16123s) values. In this industry, how you know a value matters as much as the value itself.

[04:28:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16136s) Once the form is filled, it goes out for e-signature.

[04:29:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16140s) If the client doesn't sign within a day or two, the assistant nudges them again.

[04:29:04](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16144s) Once it's signed, they file it for the client, send the acknowledgement and update the advisor

[04:29:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16149s) at the next day's stand-up.

[04:29:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16151s) One sentence from the advisor became a process that runs for days, crosses several systems,

[04:29:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16156s) weights on two people and loops back on itself more than once.

[04:29:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16174s) Look at what's inside that one sentence. Ambiguity to resolve

[04:29:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16179s) data together. Approvals, weights measured in days, follow-up loops and an

[04:29:46](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16186s) evidence trail at the end.

[04:29:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16193s) Workflow automation tools exist, but they rarely handle this real-world complexity and

[04:29:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16198s) ambiguity.

[04:29:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16199s) A fixed flowchart breaks at the first witch harbour and the lack of a native system of

[04:30:03](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16203s) record with an integration and an agent that improvises its way through is exactly what

[04:30:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16209s) a compliance team can't sign off on.

[04:30:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16211s) One breaks on ambiguity, the other can't be signed off.

[04:30:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16214s) neither handles what's inside that sentence. So what do we do about it? The fashionable framing

[04:30:20](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16220s) is more agency and better models. Given agent the whole task and wait for the next model to

[04:30:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16226s) make it reliable. The reflexive answer, especially inside regulated firms, is the opposite. Less

[04:30:33](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16233s) agency, more rules, lock it down and hard code every path. Both are the wrong axis.

[04:30:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16239s) In regulated verticals, reliability comes from where you place the boundary between the two,

[04:30:46](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16246s) between the parts that must be deterministic and the parts that genuinely need judgment.

[04:30:51](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16251s) Here's the principle.

[04:30:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16253s) Determinism governs control flow, evidence, and side effects.

[04:30:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16256s) What runs next, what we know, and how we know it, and anything that changes the world,

[04:31:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16262s) like a record written, an email sent, or a form filed.

[04:31:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16271s) Intelligence provides interpretation and judgment. Reading an ambiguous instruction, deciding which Harper, understanding a client's reply. The model answers a typed question and the deterministic system decides what to do with the answer and the interplay is recursive. A judgment can open a new smaller plan that's compiled, gated and deterministic again. You'll see that in a moment with follow ups. The boundary between the two is the design.

[04:31:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16298s) Here's how that works in the workflow engine we've built.

[04:31:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16310s) The request goes to a planner. That is a model call and its

[04:31:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16314s) only job is translation. It emits a typed semantic graph of

[04:31:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16319s) intents, entities, facts, artifacts, events, and

[04:32:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16322s) constraints, each grounded in the words of the request with any

[04:32:06](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16326s) any ambiguities it found. Every node points back to the words

[04:32:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16330s) it came from. That graph goes to a compiler that never calls a

[04:32:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16334s) model. In 11 deterministic stages, it turns the graph into

[04:32:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16337s) an execution tag or it fails closed. There are only three

[04:32:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16341s) verdicts. Valid needs clarification or refused.

[04:32:28](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16348s) Which Harper comes back as a question, not a guess. The plan is

[04:32:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16352s) is compiled, immutable and pinned by its hash.

[04:32:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16365s) The model translates. The plan is compiled. The compiler checks

[04:32:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16370s) the plan against the grammar. The vocabulary of what the firm

[04:32:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16373s) is allowed to do. 42 typed action contracts across forms,

[04:32:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16378s) documents, email, e-signature, calendar and human steps.

[04:33:03](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16383s) Each action the planner asks for has to match a contract. Here

[04:33:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16387s) form fill finds its forms contract and locks in. If the

[04:33:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16391s) planner asks for an action that isn't in the grammar or wires

[04:33:15](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16395s) one up wrong, the compiler refuses.

[04:33:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16398s) There are about 40 ways to fail closed from a cycle in the graph

[04:33:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16402s) to an unsupported action

[04:33:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16403s) And the same contract digest is pinned in the engine, the backend and the UI. So all three agree on what an action means. That is what a closed domain buys you. The model can only ask for things the firm already knows how to do safely. Here is a compiled plan for a W9 chase. Send the client the form, wait for it to come back, check it and file it.

[04:33:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16425s) Before anything runs, a person approves the plan itself.

[04:33:49](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16429s) Find the form, prepare it for the client to fill, render the PDF and draft the email.

[04:33:55](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16435s) Drafting is a model call, so it is violet.

[04:34:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16442s) Sending is an action gate. Nothing leaves the firm until a person approves that exact email.

[04:34:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16448s) Then the plan waits for a reply and a judge steps in to read it.

[04:34:13](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16453s) The returned form is verified. If it matches, it's archived behind another gate. If it doesn't,

[04:34:20](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16460s) the plan follows up. Now look at the dashed steps. Rendering the PDF, attaching it, judging the reply.

[04:34:29](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16469s) The planner never asked for them. The grammar added them because the contracts require them.

[04:34:35](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16475s) In the story, waiting, approvals, and follow-ups were most of the work.

[04:34:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16479s) Here there are nodes in the plan, not prompts. A reply wait defaults to 72 hours and an explicit

[04:34:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16487s) deadline can be anything from a minute to 30 days. When it passes, a rule decides what happens,

[04:34:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16493s) not a model. One approval covers the plan and every external action like sending an email or

[04:35:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16500s) a signature request has its own gate. The approval is tied to a hash of the exact payload and it's

[04:35:06](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16506s) checked again right before sending. When a client replies with only part of what we asked,

[04:35:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16511s) the reply judge marks it partial. That judgment is provisional. The step stays blocked until a

[04:35:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16518s) person confirms it. A person confirms it and the step unblocks. Then the follow-up runs as a new

[04:35:24](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16524s) child plan. The approved parent is never changed and there are at most five rounds.

[04:35:35](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16535s) Once a plan is approved, a state machine runs it. It picks every step and records every result.

[04:35:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16542s) When a step needs judgment, it makes one fixed schema model call. Compose an email, judge or

[04:35:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16548s) reply, spot missing values, extract fields, verify a document. The answer comes back typed,

[04:35:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16554s) and the state machine decides what to do with it. Side effects get a life cycle of their own.

[04:36:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16560s) Prepare, dispatching, succeeded, or when the world doesn't answer cleanly, outcome unknown,

[04:36:06](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16566s) and manual review. There are no automatic retries. A chat message can approve a step

[04:36:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16572s) or answer a question. It can't change an already approved plan. The model steers. It never drives.

[04:36:24](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16584s) Now look at what is left for the model to do. Outside the planner, the judgments are

[04:36:29](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16589s) small and typed. Is this reply complete? Does this document match? Which field is which?

[04:36:35](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16595s) They don't need frontier intelligence. A fast small model, the flash tier, can handle

[04:36:40](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16600s) them. In this illustrative run of 20 decisions, 19 go to the flash tier and one goes to the

[04:36:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16607s) frontier. That is why we designed for about 95% of model decisions in a workflow never

[04:36:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16613s) never needing a frontier model and the flash tier can run on

[04:36:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16617s) hardware you control. Only the genuinely hard cases go to a

[04:37:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16621s) frontier model.

[04:37:04](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16624s) the 5% is not a permanent tax because every judgment is

[04:37:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16628s) typed and logged and the provisional ones are confirmed

[04:37:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16631s) or rejected by a person. Each frontier decision produces

[04:37:15](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16635s) eval grade distribution real training data by construction.

[04:37:19](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16639s) That data can post train the flash tier. Logging and evals

[04:37:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16645s) exist today. The training loop is next. So the frontier share

[04:37:28](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16648s) shrinks release over release until even the hard cases run

[04:37:33](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16653s) locally.

[04:37:40](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16660s) Remember the requirement we started with because data is

[04:37:44](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16664s) sensitive. You need a path to move workflows entirely to

[04:37:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16668s) privately hosted LLMs. So can this design really get there? Take

[04:37:52](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16672s) one type call from the Harper Burbinine plan. Every model call

[04:37:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16676s) is small, typed and checked against a schema. So pointing it

[04:38:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16681s) at a privately hosted model is a configuration change, not a

[04:38:04](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16684s) rewrite. One line changes and the same answer comes back. Every

[04:38:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16689s) other call has that same small type shape. So the same change

[04:38:13](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16693s) applies to almost all of them.

[04:38:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16694s) That is how about 95% local inference becomes viable as an architectural consequence, not a policy promise.

[04:38:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16703s) That is the path that sensitive data demands and compliance risk goes down as an engineering roadmap, release by release.

[04:38:31](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16711s) So if you remember one thing, don't ask whether your system should have more agency or less.

[04:38:37](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16717s) Ask where the boundary is.

[04:38:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16719s) Let determinism govern control flow, evidence and side effects.

[04:38:44](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16724s) Let intelligence handle interpretation and judgment.

[04:38:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16727s) At typed points, you can inspect and use every judgment you log to need less of it next time.

[04:38:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16734s) Ask where the boundary is. Thank you.

[04:38:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16736s) I'm Karthic, CTO at Perpetual.