Evidence-backed draftingfor constructionproject documents.
- Type
- Web app, AI SaaS
- Category
- AI & LLM
- Stack
- Next.js / React, TypeScript & Tailwind, Python & FastAPI
- Live site
- app.marshaltech.us
The challenge
Project knowledge is locked in past reports, RFIs, change orders and schedules. Generic chatbots cannot reliably complete a deliverable from it, reviewer corrections are normally lost, and writing a monthly report by hand means re-reading prior months to pull cost, schedule, risk and change items together.
Marshal turns a construction company's past project documents into an evidence-backed context graph, then completes new deliverables, such as a monthly report, section by section from that precedent. Every claim is traceable to its source document and page.
What it does
- 01
Cited chat drafting
Answers cite their sources as [S#]. Marshal fills a template or writes from an outline section by section, with versions, AI revision of a single section and DOCX or PDF download.
- 02
Vault and intelligence map
Upload PDF, DOCX, XLSX, TXT or MD files with live processing status, then explore a company-wide or per-project map with real evidence behind every dot.
- 03
Review that learns
Reviewers accept, edit or reject each section with a reason. Project managers approve immediately, members need PM approval, and corrections become reusable company context.
- 04
Evidence check
Each claim is marked supported, unsupported or numeric_mismatch, so weak statements are flagged before anyone signs off.
- 05
Context graph pipeline
Documents are parsed into sentence-level source spans, and facts that do not cite a source are dropped. Retrieval scores each section against seven signals, and one model call writes each section.
Built with
- Next.js / React
- TypeScript & Tailwind
- Python & FastAPI
- PostgreSQL
- Qdrant
- S3-compatible storage
- OpenAI structured outputs
- AWS, nginx & systemd
At a glance
- 10
- Entity types in the context graph, each tied to a source or a review
- 7
- Retrieval signals used to select context for every section
- [S#]
- Source citation on every claim in a draft
- DOCX / PDF
- Export that keeps the template structure
