Estym8: built from scratch, AI-first, AI everywhere

Canonical product framing for all /docs. Every guide, runbook, and stakeholder doc in this folder describes the same product: Estym8 was designed and implemented from scratch as an AI-first construction preconstruction platform—not a legacy estimating or takeoff product with AI features added later.

---

What “AI-first from scratch” means

• Greenfield architecture: Folder ingest, classification, takeoff jobs, plan metadata, estimates, reports, Estee, and estimate-to-submittal were built as one coherent system where models are core dependencies, not optional plugins.
• Not bolt-on AI: We do not treat AI as a marketing layer on top of manual counting workflows. Upload → classify → takeoff → intelligence → export is model-driven end to end, with human review and confirm-before-save where trust requires it.
• Multi-model by design: Different stages use the right model modality (text, vision, structured JSON) via our job pipeline—Grok, Claude, OpenAI vision/CV, and supporting extractors—rather than a single generic chat wrapper.
• Workflow product, not a chatbot: AI runs inside cohesive job UIs (folder → takeoff → review → export; ClearTo* intake → completeness → comment log → back-check). Scoped assistants (e.g. Estee) help on a case — they are not the product.
• Traceable intelligence: Verbatim “As Printed on the Sheets” harvest, bid-bucket reconciliation, and sheet references keep AI outputs grounded in the documents, not free-form guesses.

---

Where AI is implemented (across the board)

• Bid-package ingest — Classify files (drawing vs spreadsheet vs supporting), path hints, dedupe drawing sets, email/snippet extraction, supporting-context LLM passes
• Run planning — AI project overview (game plan, estimate strategy, file roles, gaps) before execution
• Takeoff — Multi-discipline quantity extraction from plans and schedules; symbol/CV paths; image-tail vision for dense power-device pages; schedule-first governance with model-assisted reconciliation
• Plan intelligence — Conflicts, code posture, optimization recommendations, draft RFIs; cross-file intelligence across folder bundles
• Harvest & QA — Per-PDF verbatim harvest; code-edition mismatch detection; printed-count mining and bid-bucket reconciliation
• Estee — Optional run-grounded helper on projects and estimates (confirm-before-save); not the primary UX
• Estimate-to-submittal — Draft spec AI cross-check against catalog PDFs and linked line items
• Ongoing product — Prompting, coercion, and evaluation loops tied to ground-truth regression—not one-off demos

Operational docs (setup, testing, migrations) describe how we run and verify this AI-first stack; vision and investor docs describe why it matters commercially.

---

Portfolio (same spine, Estym8-first)

The folder-native plan-reading spine also powers adjacent products under the same company (raise remains Estym8-anchored). Each is a workflow app with AI assist — not a chatbot SKU. ClearToPermit makes city reviewers more efficient; it does not replace them.

Detail: Investor overview · plan-intelligence SKUs · municipal UX principle.

Competition (suite): There is no AI-first competitor building this suite (takeoff + city permit review + A/E QA + ICT BOD on one greenfield spine). Lanes have point tools and legacy platforms with bolted-on AI — that is different. See Suite competitive position.

---

Model strategy & independence (roadmap)

Today — orchestrate, don't train. Estym8 runs on best-in-class third-party models (Claude and others) behind a model-agnostic job pipeline. This is deliberate and capital-efficient: we stay on the frontier, pay only for inference, and carry zero foundation-model training cost. The model layer is swappable by design—stages are wired to the right modality (text, vision, structured JSON), not locked to one vendor.

The moat is the data, not the model. Our durable advantage is the proprietary ground-truth + calibration corpus (real won projects with human takeoffs/BOMs) and the regression harness that compounds with usage. That corpus is what makes a proprietary model possible—and hard for others to replicate.

Post-funding — model independence on the narrow domain. As the corpus compounds, the roadmap is to train proprietary fine-tuned / distilled domain models for the specific, repeatable tasks (symbol/device classification, schedule extraction, count reconciliation)—not a general-purpose foundation model built from scratch. Intended payoffs: lower inference cost, reduced vendor dependence/lock-in, and accuracy gains on our exact problem. We keep the orchestration layer model-agnostic so proprietary models slot in alongside (or replace) third-party calls task-by-task as they prove out on regression.

How to say it (one line): We orchestrate frontier models today and own the ground-truth data that lets us bring proprietary, domain-tuned models in-house over time—lowering cost and dependence without betting the company on training a foundation model.

---

How to talk about it (consistent language)

Short (one sentence):
Estym8 is built from scratch as an AI-first full-project preconstruction platform—AI drives ingestion, multi-trade takeoff, plan intelligence, and submittal draft review in one cohesive workflow, not a chatbot and not add-ons to legacy estimating software.

Medium (elevator):
We built Estym8 from the ground up around AI and plan intelligence: upload a plan PDF or whole bid folder, and models classify, quantify across disciplines, flag conflicts and code gaps, draft RFIs, and produce bid-ready outputs—with estimate-to-submittal and an optional scoped assistant (Estee) on the same spine.

Contrast (vs competitors):
Legacy and point tools often preserve the old rhythm (export, retype, stitch in Excel). Estym8 collapses preconstruction into one AI-native run because the product was authored for models from day one.

---

Doc map (start here)

---

When you add or revise any doc under /docs, keep this framing visible—either link here or restate the “built from scratch, AI-first, AI across the board” message in the doc’s introduction.