Pi
All case studies
02 How · case study
Live

Portfolio Intelligence Platform

14 Days. 290 Was the Estimate.

290 days. That was the estimate.

Four specialists: a frontend developer, a business analyst, a data modeller, a backend engineer.

One operator. No code. 14 active working days.

66,380
Lines of production code — zero written by hand
44.7B
Rows across 3 Snowflake databases
3,328
Source tables
7
Cortex AI agents with shared semantic layer
346
Formally documented architectural decisions
119
Consecutive sessions — zero hallucination-forced resets
251
Markdown documentation files produced
14
Active working days for PIP v1
Timeline
Mar 18, 2026

Session 1 — Blank Environment

Blank Snowflake environment. No prior context. Scope: build a portfolio intelligence platform for an energy retailer managing 14,768 sites.

Mar 18 – Mar 26

FDB Phase — Data Foundation

9 active days. Sessions 1–15. Consumption stack, wholesale cost mapping, DUoS timeband structure (5 bands — the first significant domain correction), synthetic forecasts seeded.

Mar 27 – Mar 31

Phase 1 — Core UI

5 active days. Sessions 16–27. 9 Streamlit application pages built: Home, Portfolio Health, Consumption, Gross Margin, Site Intelligence, Signals, Alerts, Forecasting, Reconciliation.

Apr 1 – Apr 3

Phase 2 — Intelligence Layer

6 active days. Sessions 28–40. 7-agent Cortex AI swarm deployed. Full 14,768-site portfolio rebuild. Signal tables, alert tables, inter-agent handoff table.

Apr 3

PIP v1 Complete

14 active working days from blank environment. Core platform delivered.

Apr 4 – Apr 17

Volta Suite + New Account Deployment Build

Volta Suite — the commercial and new account deployment layer. 7 additional active days. PIP + Volta: 21 active working days total.

Apr 18 – May 7

Customer Demo and V2 Design

All 5 streams complete. 51 calendar days total.

In every fork across 119 sessions: domain expertise, not AI capability, was the decisive input. The AI is the execution layer. The operator is the intelligence layer.
Key learnings

1. Domain expertise was the decisive input — not AI capability.

The AI executed. The operator decided what the AI needed to know. DUoS timeband structure was five bands, not the three the model assumed — caught mid-build, corrected, logged as DEC-164. Settlement crystallisation windows: what constitutes confirmed vs provisional data changes signal reliability. A Jul/Aug 2024 MPAN anomaly contextualised against known market events the AI had no way to know. Without energy market knowledge, the intelligence layer would not have been intelligent.

2. Zero warm-up across 119 sessions.

Every session closed with a verbatim handover document — a complete start prompt for the next session. No re-explaining context. No drift. 119 consecutive sessions, full context loaded from the first message. This was the single most valuable process innovation across the entire build.

3. Every decision logged and immutable.

346 formally documented architectural decisions across 5 build streams. Each marked APPROVED, REVERSED, SUPERSEDED, or DEPRECATED. Nothing deleted. The append-only log is why the system is maintainable, not just delivered.