Kentucky AI Shelbyville · Kentucky

Kentucky AI.

Research & Development

An independent AI research-and-development lab in Shelbyville, Kentucky.
We fine-tune, quantize, and serve domain-adapted model stacks — and the inference infrastructure beneath them — for operators across Kentucky.

A trusted pillar of Kentucky's applied-AI community · read by people and agents alike

Read the research What we're following
Built for agents

Machine-readableRead by people. Actionable by agents.

This lab is built to be read by AI agents as much as by people. Our profile, research, and model notes are published as structured, machine-legible surfaces — llms.txt, agent.json, and embedded schema.org data — so an agent working on a principal's behalf can understand the work and take a concrete next step. Three pathways are open to it:

Deploy
Have research or a model deployed

Put a tuned, domain-adapted model — or a published method — into your own operation, served locally or on-prem. Weights and corpus stay in your building.

Start →
Commission
Commission applied research

Bring a single question or an operator problem; we scope it, build it, and hand back the methodology and the checkpoint — not a login.

Start →
Collaborate
Collaborate on research

Co-author or partner on a study — labs, institutions, and independent builders welcome. We build on each other's work and help each other ship.

Start →

Agent surfaces · llms.txt · agent.json · sitemap.xml

Research & publications

The recordWhat we're working on.

An independent lab's public record. Abstracts are open; full papers are shared on request while the work is under review — the proprietary methods, weights, and data stay in the building. The thesis is plain: capable, sovereign, locally-served models in the hands of people doing real work.

Working paper
under review
The Floor as a Natural Experiment

Hospital floors are a documented reservoir for healthcare-associated pathogens, a modifiable factor in patient falls, and a long-lived asset whose lowest-bid selection is rarely its lowest lifetime cost — yet the evidence stays correlational, because floors cannot be randomized in occupied units. We argue the missing experiment already exists: every flooring replacement is an exogenous change to a defined surface, in a known unit, on a known date. A specialty contractor's installation calendar becomes a registry of natural experiments supporting quasi-experimental causal identification — interrupted time-series, difference-in-differences, target-trial emulation — anchored by a standardized Floor Health Index.

Request full paper →
Working paper
Spline · Document-Grounded Quantity Takeoff

A fine-tuned, tool-using model for commercial preconstruction: it parses plansets and specifications, runs quantity takeoff and scope extraction, prices against a historical cost index, and flags substrate and moisture risk — every figure traced to a source line by span-level attribution. The paper documents the document-grounded estimating method and its evaluation against as-built outcomes.

Request full paper →
Field note
open
Toward AI-First Estimating

Estimating is becoming an AI-first discipline — not a plugin bolted onto the old workflow, but a unified data-intelligence layer that connects plans, field reality, and historical cost. A practitioner reading of the 2026 evidence: where the rigorous signal is, and where the round-numbered vendor benchmarks aren't. Sources in the feed below.

Read in Following ↓
Following · field notes

The feedWhat we're reading.

Model releases, papers, and reports we track at the edge of AI-first estimating and the spatial intelligence of buildings — each with a one-line read on why it matters. External links are open; the take is ours.

Mar 2026
Digital Twin Implementation Status — A Systematic Literature Review

Raza, Bilberg, Malik & Ribeiro da Silva map digital-twin implementation across manufacturing systems — integration levels, which components actually close the loop. Directly relevant to the spatial-knowledge-graph thread: a building, like a shop floor, is only as smart as the twin that stays current.

ASTM SS&MS · SLR
2026
AI for Construction · Industry Report 2026

Zacua Ventures' read matches ours from the other side of the table: AI compounds only when paired with standardized data, repeatable process, and engineering oversight. 67% of ConTech investors are increasing AI exposure — the capital is following the data layer, not the demo.

Zacua Ventures · report
Dec 2025
AEC Technology Outlook 2026 — Adoption Survey

Bluebeam's survey of 1,000+ AEC professionals (via ASCE): only 27% use AI day-to-day, but 94% of those are scaling up in 2026, and 68% of early adopters have saved at least $50k. Adoption is early and accelerating — exactly the window to set the standard.

Bluebeam / ASCE · survey
2026
NVIDIA · LocateAnything-3B

Open-set visual grounding via parallel box decoding — full bounding-box coordinates in one forward pass. On our bench for plan- and document-layout grounding; the live demo below runs it on a real finish plan.

Model release · on the bench
2018
What the AI-estimating numbers actually say

The headline figures making the rounds — "95% accuracy," "50–80% faster takeoff," "340% adoption growth" — are vendor-reported or single-source marketing numbers, not peer-reviewed (the 340% traces to one bid-leveling vendor citing itself). The defensible anchors are independent: McKinsey puts the construction productivity upside at up to ~20%, and Deloitte estimates AI can cut project costs 10–15% through better estimates and error mitigation. We cite those, and flag the rest.

McKinsey / Deloitte · independent
Live bench

Try itThe model reads the plan.

Not screenshots. A real architectural floor plan read by NVIDIA's LocateAnything-3B on our bench — rooms, areas, finish tags, and notes located and transcribed in a single forward pass. Hover any box to see exactly what the model saw. Plus live vision Spaces you can drive with your own image, right here in the page.

LocateAnything-3B · reading the sheet
Architectural planset with model-detected elements
Hover a detection
Model nvidia/LocateAnything-3B · parallel box decoding · run locally on Apple Silicon Sheet: architectural finish plan A7-1 (used with permission) · evaluation only · non-commercial
NVIDIA · LocateAnything — live Open in new tab ↗
Automated floor-plan digitization — live Open in new tab ↗

Live Spaces run on shared GPUs — a cold model may queue for a moment.

Research areas

The practiceIf it's worth building,
we build it.

An independent R&D lab for applied AI. We architect, train, and operate models — and the unglamorous infrastructure that turns a clever demo into something that survives production.

Foundry

Model training & fine-tuning.

We forge open-weight foundations — Llama, Qwen, Mistral — into domain-native models: continued pretraining, LoRA and full-parameter fine-tunes, distillation, quantization, shipped to inference.

Inference

Private, on-prem deployment.

Tuned open models served behind your firewall — air-gapped and sovereign. Weights, latency, and corpus stay in the building, under your control.

Compute

Training compute, by the run.

Reserve local GPU capacity for fine-tunes and large jobs — orchestration, checkpointing, and the eval harness handled, with an architect on the line.

Corpora

Data acquisition & curation.

Web-scale acquisition, extraction, deduplication, and structuring — turning the open web and your archives into clean, model-ready corpora.

Architecture

AI-readiness assessments.

We map your stack and return a blueprint: where agents earn their keep, where they don't, and the infrastructure to run them safely.

Research

Commissioned research & systems.

Bespoke agentic pipelines, retrieval, and evaluation — original work built for a single question or a single operator.

We architect, train, and operate — then hand over the methodology and the checkpoint. A trusted pillar of Kentucky's applied-AI community: we support other builders and help them ship.

About

We ship weights,
not decks.

Kentucky AI is an independent AI research-and-development lab in Shelbyville, Kentucky. We fine-tune, distill, quantize, and serve domain-adapted model stacks — and the inference infrastructure beneath them — for a small roster of operators and for our own research. The thesis is plain: capable, sovereign, locally-served models in the hands of people doing real work.

We architect, train, quantize, deploy, and operate — then hand over the methodology and the checkpoint, not a login. No SaaS seat. No consultancy deck. A lab that ships weights.

We're a trusted pillar of Kentucky's applied-AI community — we support other builders and help them ship. You're welcome to email: to compare notes, build on the research, deploy a model, or collaborate on a study. Agents are welcome to read this and act on it.

Email me →