Skip to content

SITE Resource Group (SRG) — Subcontractor RFP Scope-of-Work (DocGen fit assessment)

Status: candidate — assessed, not built. Verdict: strong DocGen fit.

What SRG does

SRG is an industrial / construction general contractor. The project owner hands them one big master Scope of Work (drawings, specs, requirements). SRG breaks it into per-trade subcontractor RFP packages (Fabrication, Containment, Rebar, …), each in their house template, and sends each to subcontractors to bid. Manual today: read the master SOW, pull each trade's lines, fill the house template + responsibility matrix, produce a clean RFP per trade.

Pipeline (confirmed from the project Drive + the shared skills bundle): one owner master SOW → N per-trade subcontractor RFP scope-of-work documents. KPSC example: Input = 1 owner SOW PDF → Output = 3 per-trade SOW .docx (Fabrication / Containment / Rebar).

The house template (SITE - RFP Template) is a "Subcontractor Scope of Work" with: General Requirements, a responsibility matrix (Provided by SRG / Sub-Contractor / Client), Site Info, Scope of Work (overview/detailed/instructions/exclusions), Materials (Sub/Owner/SRG), Safety & Environmental, Quality & Turnover, Project Controls.

Is DocGen a fit?

Yes — strong. Same shape as the AltaML DealDesk SOW template already built: source doc → grounded prose+table document in a fixed template, many instances.

DocGen SRG subcontractor RFP SOW
Output prose + table document prose + table document (house template) ✅
Input source documents owner master SOW (PDF) ✅
Structure fixed sections + fields fixed sections + responsibility matrix ✅
Multiplicity many template instances one package per trade per project ✅
Grounding every claim traces to source every scope line cites the owner SOW ✅
Eval score vs a gold set existing Output/ .docx are a ready-made gold set ✅

What the shared skills bundle confirms

The SRG team's Cowork skills (sow-wbs-extraction, a generator, rfp-evals) document the pipeline and the trade-split:

  1. sow-wbs-extraction — parse owner SOW → recall-first WBS (work-breakdown into trade packages) → human review gate. This is the trade-split, solved as recall-first + human approval before generation.
  2. generator — emits one rfp_<service>.json per package.
  3. rfp-evals — scores generated RFPs against ground-truth RFPs.

Output schema maps cleanly to DocGen: section state (AI Filled / From Template / Needs Input) → render modes; source (template | {section, page, quote} | estimator scope) → grounding/citation; per-package instances → template instances; ground-truth scoring → the eval harness.

Recommendation

  • Good DocGen use-case — proceed. Stand up an SRG org + a "Subcontractor Scope of Work" template, reusing the DealDesk SOW machinery.
  • The trade-split (which master-SOW lines belong to which package) is the one piece upstream of generation — model it as the WBS-extraction + human gate the SRG skills already use.
  • Use the existing Output/ docs as the eval gold set from day one.