FFrameAlpha
Technical approachDiscussion pointsWho am I

00 — Overview

Treat the transformation as one platform problem, then ship thin verticals on top of it.

Seven Finance-facing transformation projects, built separately, each re-implement identity, data definitions, retrieval, evaluation and audit. This document is an alignment brief on how we would build a shared spine instead — not a pitch.
0.1

Three rules that decide whether this is trusted after month three

01

Numbers are computed, language is generated

A model never emits a figure it did not fetch from a tool.

02

Permissions are inherited, never re-implemented

Queries run under the asking user's identity, RLS unchanged.

03

An answer without provenance is not an answer

Measure, filters, as-of time and source version travel with every response.

Why — why platform-first, not project-first

Six of the seven archetypes below need the same four things: a metric definition, a permission model, a knowledge base, and a way to prove an answer was right. Built once, each vertical is weeks of work; built seven times, each becomes its own small platform team with its own drift.
0.2

Transformation map: seven projects, seven archetypes

Executive

Effort and impact are a qualitative assessment for sequencing discussion, not measured data.

HigherTransformation impactLower
LowerBuild effortHigher

Hover or tap a marker to read it.

A1

Deterministic calculation engine

Approved source data + frozen spec

Assists extraction/docs, never computes

A2

Governed conversational analytics

Certified semantic model + identity propagation

Intent routing + narration only

A3

Document intelligence

Stable target data model + human validation

Proposes fields; humans confirm obligations

A4

Controlled knowledge system

Named owners + review cadence

Retrieves & summarises; authorship stays human

A5

Planning & simulation UX

Planning-platform API access + validation harness

Scenario narration, anomaly flags — not submission

A6

Workflow automation at scale

Inventory of existing flows, ranked by criticality

Classification & drafting inside deterministic flows

A7

Semantic & data foundation

Agreement on the authoritative definition

Makes AI answers defensible, not plausible

0.3

Dependency order — and the reason for it

FOUNDATIONS — unblock everything

F1 Semantic layerF2 Identity & RLSF3 Knowledge architectureF4 Platform services (audit, eval)

VERTICALS — thin once foundations exist

A1 Calculation engineA3 Document intelligenceA6 Workflow automationA5 Planning UXA2 Conversational analytics (needs ALL four)

Hypothesis — sequencing

We expect conversational analytics to be the most requested and least ready, since its dependency on definition governance is hidden. See the assumptions page for the discovery questions that confirm or refute this.

Design decision — parallelism

One foundation stream, at most two verticals concurrently in phase one — more removes the feedback that keeps standards real rather than theoretical.
0.4What changes in engineering practice when the system is AI-nativeImplementation
0.5

How to read the rest of this document

01

Reference architecture

Seven layers from ingestion to experience, plus two worked walkthroughs.

02

Solution patterns

Each archetype in build detail: design, determinism boundary, evaluation, when not to use it.

03

Delivery model

Engagement scenarios, phase gates, decision records.

04

Quality & evaluation

Testing pyramid for AI systems, offline eval sets, CI gates, drift monitoring.

05

Risks & mitigations

Failure modes with mitigations and early-warning signals.

06

Metrics & baselines

What to measure per archetype, and baselines to capture before build.

07

Assumptions & questions

Everything unknown, stated as a question — the first-workshop agenda.