Revenue Operations & Financial Intelligence

How do you automate a manual management reporting pack?

Unify the streams, compute one way, and gate what can be reported. This dashboard replaced a recurring spreadsheet reconciliation across pipeline, AR, AP and market data with an audited ingestion path, whitelist-only monetary metrics, and AI restricted to narration and classification — never to producing a financial figure.

Alex Scharifker · Product architect and implementation lead

Designed and led by Alex Scharifker as an internal revenue and financial-intelligence system.

Edge Functions

Data Streams Unified

Financial Figures From AI

Problem

Sales Ops Dashboard

Pipeline data lived in Quickbase, AR and AP aging lived in accounting exports, client conversations lived in email and memory, and market signals lived in browser tabs. Producing one management view meant a recurring manual spreadsheet reconciliation — slow, unauditable, and prone to definition drift where two people could produce two different numbers from the same source.

System

How the system is put together

Sources in, deterministic logic in the middle, AI restricted to interpretation.

Data sources

Quickbase REST pull and webhook push
Accounting exports (AR aging, AP aging, invoice detail, payments)
Tender boards and market news feeds
Conversations, comments and attachments logged in-app

Layers

Ingestion and sync edge functions
Normalized, org-scoped relational model
Deterministic calculation layer
Report catalog gating
AI inference and semantic retrieval
Audit, roles and tenant isolation

Deterministic logic

Whitelist-only monetary metric resolution with a non-monetary blocklist
AR/AP aging buckets, DSO, DPO and cash conversion cycle
Sunday-cutoff weekly plus YTD-cumulative series
Entity and intercompany classification
Report gating by required dataset types
Row-level tenant isolation and database-driven RBAC

Where AI is used

Classifies unfamiliar file shapes as an audited fallback, never the first pass
Extracts topics, risk flags and next-best-action from free-text conversations
Scores conversion likelihood and surfaces at-risk relationships
Narrates deterministic results — barred from producing any financial figure

Decision loop

Ingest and normalize under one audited path
Compute deterministically, gate what can be reported
Decide: who to pay, who to chase, which deals to push
Write the action back and measure the outcome

My role

Product definition and platform architecture
Ingestion pipeline and data-integrity rules
Deterministic calculation layer and report gating
Applied AI architecture and guardrails

Architecture

The architecture case file

Layer-by-layer architecture, deterministic model, AI guardrails, and the decision loop.

A monthly spreadsheet reconciliation produced a backward-looking summary nobody fully trusted. This system replaces it with one governed ingestion path, one set of encoded definitions, and a current pipeline-to-cash picture.

Ingest→Normalize→Compute→Gate→Decide

Key takeaways

—Four streams — pipeline, client relationships, market intelligence and the financial back office — consolidated into one governed multi-tenant model.
—Ingestion is deterministic first: filename, header signature, then soft tokens, with AI only as an audited fallback.
—Whitelist-only monetary resolution ended phantom revenue caused by identifier fields being summed as currency.
—Reporting standards are encoded as rules — Sunday-cutoff weeks, entity and intercompany separation, aging thresholds, DSO/DPO/CCC.

Architecture Principle

One ingestion path, one normalized model, one set of encoded definitions.

One governed model replaces a recurring spreadsheet reconciliation across pipeline, relationships, market intelligence and finance.
Tenant isolation is a database property, not application logic — no broad authenticated-role policies.
Anything authoritative lives in the database or an edge function; presentation-layer math stays pure and testable.

Sources

Quickbase REST pull and webhook push, accounting exports, tender boards, market news feeds, and human entry in the app — four streams that previously lived in four different tools.

Edge layer

27 Deno edge functions split by role: ingestion and sync, AI inference, retrieval, and platform. Auth, CORS, audit and model access are implemented once in shared modules.

Governed model

Isolation

Postgres as the system of record. Every business table carries organization_id and is scoped by row-level policy; roles live in a separate table and are read through security-definer functions, never client state.

Deterministic layer

Pure TypeScript over normalized rows: whitelist metric resolution, AR/AP aging, DSO/DPO/CCC, Sunday-cutoff weekly and YTD series, entity and intercompany classification.

Gated surface

Payoff

The report catalog unlocks only what the present datasets can actually support. Dashboards write decisions back into the same tables the analysis reads from.

Outcome

What changed

Replaced a recurring manual spreadsheet cycle with an automated, audited board pack.
Eliminated a class of phantom-revenue defects by replacing heuristic metric detection with whitelist-only resolution.
Surfaced previously invisible 61-90 and 90+ AR and AP exposure, making working capital decisions current instead of retrospective.