Product & build hub

Where the product is, and where it's going.

One place to manage NCUE as product and build at once: what it is, the release log, the live board, the architecture, and the go-to-market. (Internal — for the team and reviewers.)

current: v0.9target: v1.0 — agents online72 tests greenlive: ncue-dev.pages.dev
1 · The product
What it is, why it exists, who it's for.

What

An agent-assisted underwriting tool for ground-up development. It turns a deal that doesn't pencil into a financeable, fully-sourced one — every number is the engine's, every assumption carries its source.

Why

Ground-up underwriting is slow, expert-scarce, and untrusted. A black-box “AI underwriter” is unusable in a high-stakes deal. NCUE shows a sourced, replayable path from no-go to go.

Who

Single-building multifamily developers (the broad market), district developers like Now City (the power user), and the LPs / lenders who consume the report.

Headline proof (live, real engine output): base case −17.1% IRR (no-go) → composed path +23.9% IRR, 2.20× equity multiple, 1.37 DSCR.

2 · Versioning & releases
Scheme: v0.MAJOR = a coherent capability era; v1.0 = the agents go online (Decision D). We ship continuously; each row is a milestone, not a date.
VerThemeStatus
v0.1Deterministic engine
Pure pro-forma engine; Python↔JS parity oracle.
✓ shipped
v0.2Zero-backend site + scenarios
Static Cloudflare build; the scenario ladder; CI/CD.
✓ shipped
v0.3Provenance + MCP
Per-assumption citation store; MCP server tools.
✓ shipped
v0.4Five agents + orchestrator
The composers, the Lead-Agent composer, /explore.
✓ shipped
v0.5The model
/model = the full tool + the agent control layer + dashboards.
✓ shipped
v0.6Mobile + transparency
Full responsive pass; /architecture + flow diagrams.
✓ shipped
v0.7Validators bind (Decision B)
/model routes through the tested, validator-aware orchestrator.
✓ shipped
v0.8One engine source (Decision C) + glossary + report
Engine generated from Python; /glossary; /report + PDF.
✓ shipped
v0.9Ten-agent fleet + contract
Steel-man typing (composers/policies/gates); two-speed architecture; agent-contract schemas + golden-deal evals (72 tests).
● current
v1.0Agents online (Decision D)
Async agentic runs behind a flag; constraint layer; first LLM composer. Needs host + key.
○ target
3 · The board
The live work, Kanban-style. Type tags: feature infra docs decision ux gtm.
Shipped 10
Deterministic engine, single source
infra Decision C · v0.8
Validators bind on the live path
feature Decision B · v0.7
/model = full tool + 10-agent layer
feature v0.5–v0.9
Full responsive / mobile pass
ux v0.6
/architecture + 8 flow diagrams
docs v0.6–v0.9
Steel-man typing: composers/policies/gates
decision v0.9
Two-speed (fast/slow loop) architecture
decision v0.9
Agent-contract schemas + golden-deal evals
infra v0.9 · 72 tests
/glossary (52 terms) · /report + PDF
docs v0.8
Risk & Returns exit provenance (real comps)
data v0.9
In progress / ready 2
Wire Risk & Returns as the 6th composer
feature ready · v0.9→1.0
0–100 deal scorecard (Q5)
feature scoped · SCORECARD.md
Next sprint 4
Decision D · step 1: the constraint layer
infra policies/gates → orchestrator
Decision D · first online composer behind a flag
feature needs host + key
Pro-forma / data-room ingestion → prefill
feature Claude long-context
Investor PDF: truth-label polish
feature proposed vs confirmed
Backlog / later 7
Buildability validator (FAR/envelope binds)
infra engine guard
Engine growth: Program & Operations levers
infra
Archetype-aware /report toggle
feature after D
Client Excel ingestion (arbitrary workbook)
feature
Multi-stakeholder views
feature
State persistence (Pro saved deals)
infra
The watch loop (scheduled re-underwriting)
feature
Needs Erik 7
Decision D: host the backend + Claude API key
decision the unblock
Branding pass: name / wordmark / tagline
gtm before marketing spend
Insurance: standalone gate vs fold into Risk
decision rec: standalone
Capital Sourcing: standalone vs Capital-Stack dim
decision rec: standalone
Entitlements: planned vs fold into Program
decision rec: planned
Retire /full → redirect /model (Decision A)
infra anytime
Responsive judgment calls (headers)
ux
4 · Architecture & the fleet
The one rule everything rests on: agents propose inputs; the deterministic engine does all the math. The full map (engine, parity, provenance, the eight flows, the steel-man pass) is on /architecture.

The engine

One source (engine.py), generated to the browser JS — proven identical by parity + fuzz. Validators (DSCR, facility cap) bind on the live path. 72 tests gate every deploy.

Two speeds

Fast loop: the instant engine you touch live (<50ms, free). Slow loop: the agent team as a background job that outputs sourced scenarios. Nobody watches tokens burn.

The contract

Typed payloads (Move / ConstraintSet / GateVerdict / QAReport), a propose-only firewall, value clamping, and golden-deal evals — the safety net that makes the agents safe to wake.

The ten-agent fleet, typed by what each is for:

ProgramCostRevenueOperationsCapital StackRisk & Returns Capital SourcingESGInsuranceEntitlements
6 composers propose engine inputs · 2 policies set constraints · 1 gate is a yes/no · 1 planned (no live levers yet). Full cards on /architecture.
5 · Go-to-market & marketing
If the product isn't compelling and eliciting good feedback as we build, we aren't doing it right. The GTM is part of the build, not after it.

Positioning

“Turn a no-go development deal into a financeable one — and show your work.”

Not an AI that grades a finished deal (acquisition screeners do that). NCUE composes the ground-up deal: what to build, and how to make it pencil — with every move sourced.

The wedge (vs. UpsideIQ & screeners)

They answer “should I buy this stabilized asset?” We answer “what should I build, and how does it pencil?” Our depth they don't have: the incentive stack (LIHTC/HUD/TIF/NMTC/OZ/C-PACE), per-assumption provenance, and an agent team that composes the path.

ICP & segments

  • Primary: single-building multifamily developers (~100+ units).
  • Power user: district / neighborhood developers (Now City).
  • Report consumer: LPs, lenders, ICs — the trust audience.

Pricing

ExploreFreeThe model in-browser; the wedge.
Report$99/dealUpload → composed path → investor PDF.
Pro$499/moUnlimited deals, saved pipeline, watch loop.
Done-for-youCustomNow City Labs runs it end-to-end.

GTM motion

  • PLG wedge: free Explore → the no-go→go “aha” → paid Report.
  • Trust-as-marketing: /architecture, /glossary, provenance — transparency sells to a skeptical institutional audience.
  • DFY + dogfood: Now City Labs runs real deals, building credibility and feedback.
  • Design partners: a handful of developers giving feedback as we build.

What “good” looks like

  • Activation: a deal explored; the gap closed.
  • The “aha”: “it found upside I'd have missed.”
  • A report generated and taken to an IC/lender.
  • Design-partner pull (“can I use this on my next deal?”).

Open GTM calls (for Erik / the PM review)

Branding pass before any spend (name / wordmark / tagline). Which segment leads — broad single-building PLG, or district/DFY high-touch first? Design-partner list — who are the first 3? Is the transparency story the lead marketing asset, or a supporting one? These are on the board under Needs Erik.