@devanandb thanks! Devanand, actually considerate questions. Let me solution in two portions.
1) “Style-agnostic routing”: how does Leni make a decision which LLM handles what, and does the person have a say?
At a prime degree, it’s extra dynamic than “one fashion for math, one fashion for writing,” however we do use that spirit (specialization) underneath the hood.
We use a planner/executor structure:
Planner: breaks your request right into a typed step graph (as an example, “retrieve the appropriate supply information,” “compute and reconcile numbers,” “write the memo,” “validate outputs”).
Executors: each and every step is dispatched to the most efficient “employee” for that task, which could be a other fashion (or instrument) relying at the step’s necessities (reasoning intensity, velocity, value, strictness, context window, and so forth.).
Effects waft again to the planner, which is able to adapt the remainder plan in keeping with what got here again, together with working verification passes.
So sure, it’s task-aware, but in addition context-aware and adaptive step by step, no longer a static mapping.
On person keep an eye on, we do each:
For the general public, it’s at the back of the scenes with a powerful Auto default (so that you don’t must develop into an LLM ops engineer).
However we additionally imagine enterprises must have the ability to standardize on authorized fashions/suppliers, and in some instances pressure routing constraints for compliance, safety, procurement. The workflow must no longer exchange when your company’s fashion coverage adjustments.
2) How did we get the primary institutional consumers to have confidence AI with actual cash choices?
You’re precisely proper: have confidence is the chilly get started drawback.
The truthful solution: we didn’t get started via asking someone to “have confidence the AI.” We began via incomes have confidence operationally, in a couple of planned steps:
Get started the place the ache is prime however the blast radius is managed
Early use instances have been time-sink analyst paintings (information pulls, consistency exams, first-pass drafts, reconciling numbers throughout resources) the place the workforce may nonetheless overview outputs temporarily.
Win on verifiability, no longer vibes
Establishments don’t care if the solution sounds good. They care if it’s proper, and if you’ll be able to display why. So we concerned with:
deterministic information retrieval from their programs,
particular calculations,
consistency exams,
trust-building behaviors like surfacing assumptions and tying outputs again to supply artifacts.
Meet them the place their information already lives, and stay it protected
Numerous early have confidence got here from with the ability to perform throughout the truth of institutional stacks (belongings control, reporting programs, deal doctors, fashions) and being transparent on safety barriers (no cross-client leakage, no coaching basis fashions on Jstomer information, and so forth.).
Increase scope best after repeated “no-surprise” results
As soon as groups noticed the similar degree of high quality throughout more than one cycles (per month reporting, portfolio tracking, underwriting improve), they naturally moved from “low-stakes” to “actual choices,” for the reason that gadget had already confirmed it would behave like a competent analyst.
So no unmarried buyer “went all in” on day one. It was once extra like: turn out accuracy, turn out safety, turn out repeatability, then scale.
If you need, I will percentage a concrete instance of what a routed step graph seems like for one thing like “construct an IC memo, tie-out numbers to the fashion, generate a lender-ready package deal.” That has a tendency to make the routing idea click on speedy.
