Build Phases

Three tracks. One calendar. The winter trough is the load-bearing element; missing it costs a year.

Calendar reality
Window What happens Why then
Sept to Oct Portal, NDA, discovery signed Attention is thin during harvest. Ask for something small.
Oct to Nov Track 0 discovery runs Harvest ending, benches clearing
Dec to Feb Phase 1 and most of Phase 2, Program tier The winter trough. Technicians have time to talk. Reference machines are parked and accessible. This window does not come back until next December.
Mar Step to Build tier Their season starts, our heavy build window closes
Mar to May Phase 2 completes Bench rungs 1 and 2, cross-site triage, and intake transcription in place before planting
May onward First real season of the tool Measured against the Track 0 baseline captured in October
The winter trough is the load-bearing element. Corpus capture, technician interviews, and reference machine recording are only possible when the benches are quiet, and all three are prerequisites for everything else. Missing the December window costs a year. This should be said out loud at the next call.
Track 0: Discovery and Corpus Assessment

Track 0 Discovery and Corpus Assessment

This is what the next call is asking for. Everything diagnostic, nothing constructive. Four to six weeks. Fixed fee. Credited in full against the first retainer month if you proceed within thirty days.

Track 0 exit

A written document that says either "the corpus supports this and here is what Phase 1 costs" or "it does not, and here is what would have to be built first." Both are successful outcomes of a discovery engagement.

Phase 1: Substrate and Corpus

Phase 1 214 agent hrs

Requires almost nothing from their staff beyond the interviews already run in Track 0.

Phase 1 exit gate

"An agent, asked about a real historical failure it has not seen, returns the resolution that was actually performed, citing the ticket. If it cannot, do not proceed. Fix the corpus."

Phase 2: Bench, Triage, Intake, Trust Layer

Phase 2 554 agent hrs

Three workstreams in parallel at Program tier, two at Build tier.

Workstream 1: The bench

Hardware interface on one bench in Des Moines. Device identification on the top device family. Guided diagnostic at Rung 1. Verdict with evidence and calibrated confidence. Trace capture and replay at Rung 2 on one device family. Every session archived as a labeled example. Verdict controls and technician sign-off ship with Rung 1, not after it.

Workstream 2: Cross-site triage, the safe win

Repair queue orchestration across four sites by capacity, capability, and turnaround risk. Turnaround forecasting that flags units about to miss the published 3-5 day window. Near-zero technical risk. Hits the seasonality caveat on their own website. If the bench runs long, this is what carries the engagement.

Workstream 3: Corpus, trust, and continuity

Technician authorship and review workflow. Reference machine capture program, must happen in the winter trough or not at all. Continuity package with a rehearsed handover drill. Voice transcription and intent classification, listening only; no caller speaks to an agent yet. Operator console so your staff can see node activity without us.

Phase 2 exit gate: the real one

"One technician, on one bench, uses this daily by preference rather than instruction, for four consecutive weeks. If they stop when nobody is watching, the tool is wrong and no amount of Phase 3 fixes it."

Second gate

The handover drill was actually run and what it revealed was fixed.

Phase 3: Expansion

Phase 3 528 agent hrs

Voice intake live, one intent class at a time, each measured against human handling before the next is enabled. Consent and disclosure ship before the first live call. Bench rolled to all four sites. Rung 3 behavioral modeling funded per device family behind its own go/no-go, decided on measured repair volume rather than enthusiasm. Field tablet for internal Ag Express technicians. Parts demand forecasting. Web node: consolidation, catalog repair, generated catalog content.

Phase 4: Extension and Data Products

Phase 4 604 agent hrs

Dealer tenancy and licensing. Dealer technical backup portal. Domain-adapted model fine-tuned on the accumulated corpus. ICW production support. Onboarding and training agent. Inbox nodes. Electronics residual risk scoring and external distribution.

Phase 4 does not start on a schedule.

It starts when the dealer network asks for access without being prompted, or when a lender or auction house asks what an electronics risk score would cost. Those requests are the market signal. Their absence is also a signal.