MEMICHAEL EDLINLet’s talk
Choose your view
MICHAEL EDLIN / CONSTRUCTION TECHNOLOGY & APPLIED AI

I scale the work.
I build the software.

13+ years in construction, from installing floors to leading estimating. I grew a bid department, then built and deployed AI software that helps each estimator deliver more.

Head of Estimating & Preconstruction, Summit Flooring Group
Founder, Kentucky AI · Creator, OpenTakeoff

Read ConTech résuméContact Michael

Latest ConTech résumé · September 24, 2026 · PDF, 3 pages

DEPARTMENT GROWTH · BEFORE AI$2.5M to $7.5M

Annual awarded work

Tripled annual awarded volume with no added headcount.

PER-PERSON OUTPUT · WITH DEPLOYED AI$2.5M–$3.5M

In bids per person, per week

Up from approximately $550K–$1M per week, using AI software I built and deployed.

Business results reported by Michael, October 2026. Weekly figures describe bid value produced; annual figures describe awarded work.

WHAT I BRING

Operations experience.
Hands-on AI engineering.

01 / LEAD

Grow a department

Lead a three-person estimating team across a 13-state commercial flooring business. Own the workflow from bid selection and pricing through proposals and award handoff.

02 / BUILD & DEPLOY

Put software to work

Built OpenTakeoff and Spline, deployed software in live estimating, trained users, and improved the workflow from their feedback.

03 / ADAPT MODELS

Work inside the model

Build computer vision models from pretrained weights, modifying initialization and task-specific heads. Work with fine-tuning and evaluation; know LoRA, QLoRA, and knowledge-distillation methods.

SEE THE WORK

From building plans
to working products.

OpenTakeoff connects people and AI assistants to the same measurement tools. Spline brings that work into the estimating workflow.

Watch the product demo
OpenTakeoff displays measured floors and walls from a construction plan
OpenTakeoff + SplineConstruction measurement, estimating, and AI tools developed from firsthand experience.
ROLE FIT

Construction knowledge.
Technical delivery.

I’m interested in roles connecting customer problems, product development, and successful adoption.

Applied AISolutions engineeringAI product developmentTechnical implementationConstruction technology

Shelbyville, Kentucky · Open to relocation and travel

Technical skills at a glance

Computer vision & model adaptation
PyTorch, pretrained weights, initialization, custom prediction heads, training, and evaluation.
Language-model methods
LoRA, QLoRA, fine-tuning, and knowledge distillation.
Software & deployment
Python, TypeScript, MCP tools, MLX, and cloud GPU workflows.
Data & experiment tracking
Label Studio, MLflow, DVC, and Hugging Face.
LET’S TALK

Bring me a real problem.

michael@kentucky-ai.com
Download ConTech résuméView my open-source work
APPLIED AI BUILDER · KENTUCKY, USACONSTRUCTION × APPLIED AI

Michael Edlin.

Domain expertise.
Working software.

I’m the founder of Kentucky AI and creator of OpenTakeoff. I bring 13+ years of construction experience to applied machine learning, agent tooling, and the evaluation of AI in real workflows.

See selected workOpenTakeoff on GitHub
01 / OPENTAKEOFF · NATIVE 3D VIEW
REAL WORK. OPEN CODE. MEASURABLE RESULTS.Founder / Kentucky AI
2,300 DOWNLOADSOpenTakeoff MCP · npm~477 VISITSCloudflare · filtered production traffic161 STARS68 forks · built in the open
NVIDIA InceptionKentucky AI · program member

Built, released, and reaching people. Cloudflare recorded approximately 3,160 page views across 477 filtered visits to opentakeoff.kentucky-ai.com, September 10–26, 2026.

Sources, dates, and counting method

Cloudflare Web Analytics snapshot saved September 26: production hostname only, bot-flagged traffic excluded, preview deployments and downstream self-hosts excluded, and 270 identified automation visits removed. Earlier days were sampled, so counts are approximate. Repeat visits and some owner traffic remain possible; visits are not unique people or completed takeoffs.

The 2,300 npm downloads are Michael’s reported OpenTakeoff MCP server figure as of October 7, excluding mirrors and bots; the reporting window and filtering have not been independently verified. GitHub: 161 stars and 68 forks, October 7. These sources cover different periods.

AT A GLANCE

The facts.

Current work
Founder, Kentucky AI · Creator and maintainer, OpenTakeoff
Domain experience
13+ years in construction and flooring · field work through preconstruction leadership
Published ML work
3 Hugging Face model repositories · 5 downloadable September model components
MCP adoption
2,300 OpenTakeoff MCP server downloads · founder-reported, excluding mirrors and bots
Open source
Apache-2.0 browser application + MCP server · 161 stars / 68 forks, October 7, 2026
Technical practice
PyTorch · MLflow · Label Studio · Lambda · MLX · DVC · TypeScript / React
Production reach
~477 filtered visits · ~3,160 page views · Cloudflare, September 10–26, 2026
01 / SELECTED WORK

Built because
I needed it.

Deep domain knowledge, turned into working systems. Start with the code. Then watch it run.

FLAGSHIP · OPEN SOURCE

OpenTakeoffOne engine.
Humans + agents.

A construction measurement engine with a browser canvas for estimators and an MCP interface for AI agents. Both use the same geometry and quantity calculations.

Measurements keep their scale, method, and review history. The person doing the work stays in control.

Apache-2.0TypeScriptMCPGeometry
Try OpenTakeoffRead the source
Measured rooms on a construction finish planDRAWING → MEASUREMENT → REVIEW
02 / APPLIED RESEARCH

Three model projects.
Documented results.

A model-by-model record of the architecture, training, evaluation, and release status published on Hugging Face.

01 / DIVISION 9 ESTIMATOR

Commercial flooring
bid-estimating model

div9-flooring-estimator-gemma4-31bPUBLIC MODEL CARD · PRIVATE WEIGHTS
Base / method
Gemma 4 31B · rank-8 LoRA over a 4-bit quantized base
Training data
748 supervised pairs from verified historical bids
Holdout
51 projects dated after the training window
Reported result
12.3% median absolute percentage error on bid totals
Structured output
48 / 51 holdout outputs passed strict JSON parsing
Hardware / stack
Single 128 GB MacBook Pro · MLX / mlx-lm · 32.7M trainable parameters

Author-reported evaluation for one company’s cost structure. Weights and pricing data remain private; no public inference endpoint. The model card documents preliminary results and failure cases.

02 / VERTICAL SPECIALIZATION

Generic Gears Gym
estimating experiment

gears-div9-estimator-gemma4-31bPUBLIC MODEL CARD · PRIVATE WEIGHTS
Base / method
Gemma 4 31B · QLoRA · 4-bit NF4 · PEFT
Data split
39 training projects · 4 validation · 4 held-out projects
Experiment
Three training seeds · six epochs · rank-8 adapter
Reported result
13.0% median APE for the best seed; 13.0 / 13.2 / 21.9% across seeds
Output format
4 / 4 strict JSON parses for every seed
Finding
Vertical-only training did not outperform the parent model’s fitness-project slice.

Pilot-scale holdout (n=4). This is a documented specialization experiment, not a general accuracy claim. Adapter weights are not distributed.

03 / CONSTRUCTION VISION

September
floor-model components

september-floor-modelFIVE CHECKPOINTS + FEATURE CODE
Visual roles
roles-e070b.pt · DeepLabV3 / ResNet50 · wall, fixture, annotation, hatch, background
Wall context
walls-e068.pt · SegFormer B2 wall segmentation
Vector walls
wall_v0_noband.pkl · 36-feature wall classifier
Doorway seals
seal_v0.pkl · paired doorway-seal classifier
Nonwall veto
expanded_head.pkl · visual-context rejection of false wall candidates
Validation
All five checkpoints loaded in a Lambda A10 smoke test; both visual networks produced finite outputs.

Component compatibility is validated; full-sheet accuracy is not certified. This release does not package the complete pipeline. Component licenses differ; SegFormer is research/evaluation-only.

DOCUMENTED EVALUATION

From an untuned model
to a domain-specific adapter.

Median bid-total error on the model card’s 51-project temporal holdout. Lower is better.

31B untuned base62.8%
7B tuned · v036.1%
31B tuned · v213.6%
31B tuned · v2.112.3%

Source: published model card. Base results use lenient parsing; tuned results use strict parsing. Single-run comparisons; the v2 → v2.1 improvement was not statistically resolved.

AGENT SYSTEMS / BUILT AND SHIPPED

Tools agents
can actually operate.

I build the server, the measurement contract, and the evaluation that checks whether the resulting work is useful.

OPENTAKEOFF MCP

A working tool surface
over a shared geometry engine.

2,300MCP server downloads

Founder-reported October 7, 2026; mirrors and bots excluded.

Agents load plans, calibrate scale, measure geometry, inspect overlays, and export takeoffs through a stdio MCP server. The browser and server import the same measurement logic.

  • Publishable npm package, Docker workflow, and a bundled desktop extension.
  • Tool-call tracing on stderr keeps diagnostic output off the protocol channel.
  • Scale changes recompute geometry-derived quantities; reviewed measurements have explicit edit protections.
  • The published benchmark exercised the real server across 48 retained model trials.
Inspect the MCP servernpm package
COMMONWEALTH AGENT UNION

A proving ground
for construction agents.

READ.
CALIBRATE.
MEASURE.
VERIFY.

Built the Commonwealth Agent Union and OpenTakeoff Academy: a place for agent evaluation, public evidence records, membership, and human-proctored qualification.

  • Separate self-reported reference runs, independent agents, and certified result records.
  • Document tool-operation provenance and signed run bundles.
  • Define apprentice, journeyman, and master qualification paths.
  • Connect measurement evidence to review and acceptance criteria.

Early-stage platform. The site identifies its current benchmark reference as a single-plan, self-reported run; the hiring hall is preparing for funded work.

Visit the Union
PUBLISHED BENCHMARK / OCTOBER 2026

Evaluate what
the model actually draws.

I built a harness that makes models use OpenTakeoff’s MCP tools, then scores their committed geometry against reviewed takeoffs.

8models compared
40 + 8floor + base/wall trials
66approved floor regions

91.0% floor overlap.
14.9% boundary F1.

Kimi K3 led mean floor IoU in this run, but its wall-face placement was much weaker. The benchmark separates geometric correctness, quantity error, completion, runtime, and cost.

Published results,
explicit limits.

Eight models, the same five private sheets, one retained trial per model/sheet. Anonymous scalar results and summary-reproduction code are public. Source drawings and answer geometry remain private.

Read the benchmark
03 / HANDS-ON TECHNICAL EXPERIENCE

The tools.
The work behind them.

Specific systems I have built and experiments I have run, across data preparation, training, evaluation, and delivery.

01

MLflow

Experiment tracking I can interrogate.

Built labhub, a Python CLI that imports experiment registries, compares runs on the same evaluation protocol, and records metrics, parameters, checkpoints, and review artifacts.

Implementation detail

Remote GPU runs log locally; full metric histories are pulled back before the instance is removed. Process completion and model acceptance are recorded separately.

Python · MLflow · experiment lineage
02

Label Studio

A bridge between predictions and human judgment.

Implemented task import, local image storage, and PNG-mask conversion to RLE brush predictions. Configured polygon and brush labeling for rooms, walls, doors, and hatch.

Implementation detail

Review categories capture escaped rooms, wrong shapes, missing door gaps, and faulty reference labels. Predictions remain distinct from human annotations.

Label Studio SDK · RLE masks · annotation export
03

PyTorch

Training loops, losses, and model internals.

Built segmentation trainers with custom datasets, ignored-label masking, class-weighted cross-entropy, AdamW, OneCycleLR, mixed precision, and checkpoint selection by validation IoU.

Implementation detail

I build from off-the-shelf pretrained weights and modify initialization and task-specific prediction heads. Architecture work spans DeepLabV3, SegFormer, SAM 2 decoder adaptation, and a ConvNeXt/FPN pilot with a native drawing-segment transformer.

PyTorch · CUDA · MPS · bf16
04

Lambda

From a checked dataset to a multi-GPU run.

Ran an eight-A100 pilot: six GPUs for the main training job and one GPU for each of two smaller comparison arms. The main run reached 39,918 steps.

Implementation detail

Used checksummed data transfer, project-grouped splits, training snapshots, and deadline watchdogs with verified instance termination. Also validated released checkpoints on a Lambda A10.

Lambda GPU Cloud · distributed training · lifecycle automation
05

MLX + PEFT

Language models trained on domain records.

Fine-tuned a 31B model on a 128 GB MacBook Pro with MLX. Explored rank-8 LoRA and NF4 QLoRA/PEFT for structured commercial estimating. Familiar with knowledge-distillation methods for transferring a larger model’s behavior to a smaller model.

Implementation detail

748 verified supervised bid pairs; a 51-project temporal holdout. Private training data and adapters stay private; methodology and limitations are published.

MLX · mlx-lm · LoRA · QLoRA
06

DVC + Hugging Face

Connect the data, run, and released artifact.

Versioned a wall-training dataset with DVC and linked its content hash to MLflow runs. Added checkpoint publishing that records the Hugging Face repository and commit on the originating run.

Implementation detail

Published five vision/geometry checkpoints with feature code, checksums, load instructions, and component-specific license notes.

DVC · Hugging Face Hub · reproducibility
SELECTED ENGINEERING RECORDS

What the experiments actually showed.

Summaries from my 2026 lab records. These are research measurements with their original scope, not claims of production accuracy.

SYSTEMS / APPLE SILICON

A CUDA-dependent model,
running through plain PyTorch.

Ported VecFormer operations using hash-gather sparse convolution, scaled dot-product attention, and native scatter reductions. Used the same replacement operations for CUDA training and MPS inference.

0.594CUDA validation F1
0.595Mac validation F1

Same pilot checkpoint, 120 held-out tiles. A 66-tile full-sheet runtime benchmark took 51.7 seconds. This establishes portability; the pilot remained precision-limited.

VISION / SAM 2 ADAPTATION

Measure the task.
Then separate the failure modes.

Fine-tuned 4.22M decoder parameters in SAM 2.1 Hiera Small on 19,593 samples across 242 jobs. Evaluated rooms separately from narrow strips.

0.382Room mIoU · baseline
0.647Room mIoU · adapted

36-room evaluation bucket. The separate 10-strip bucket barely changed (0.225 → 0.232). The aggregate headline was withdrawn in favor of the task-specific results.

EVALUATION / PRODUCT JUDGMENT

Better component scores.
Still a failed release gate.

A paired-face labeling change improved segment wall F1 on 100 held-out validation sheets. Whole-plan review still found floor spill, wall intrusion, and incomplete usable room boundaries.

0.734Segment F1 · baseline
0.844Segment F1 · revised

The model was not promoted on these scores. Full-plan utility remained the acceptance criterion.

04 / OPEN COLLABORATION

Built in Kentucky.
Developed in the open.

I collaborate internationally through GitHub. The work is visible in issues, reviews, and merged contributions.

On-device OCRKevin N. Murphy · scanned finish schedules read in the browserMerged PR #487PDF engine compatibilityOdalo Eguabor · path decoding and color handlingMerged PR #433Product refinementSabih Sarowar · early-access request flowMerged PR #473A shared codebaseExplore the contributors, reviews, and development history.GitHub contributors
05 / VISUAL WORK

From a drawing
to something tangible.

Native 3D, material layouts, and visual storytelling. Open any image for a closer look.

Watch the work happen.

Autonomous agent takeoff videoPLAY ON YOUTUBE · 2:47

An agent runs the takeoff

A continuous demonstration of the MCP workflow.
OpenTakeoff 3D view demonstrationPLAY ON YOUTUBE

The takeoff, in three dimensions

Floors, walls, openings, and finishes in the native view.
Terminal-driven takeoff demonstrationPLAY ON YOUTUBE

From the terminal to a measured plan

An agent works through a studio flooring takeoff.

More videos on Kentucky AI
06 / ABOUT MICHAEL

I know what the output
has to survive.

I’ve installed floors, run crews, owned a flooring business, managed projects, and led commercial estimating. That experience shapes the software I build and the questions I ask of a model.

An estimate has to survive a real project. A model has to survive a real drawing. I care about the gap between an impressive demo and a system someone can depend on.

Through Kentucky AI, I build the tools, curate the data, and publish the work. I’m interested in AI teams solving difficult problems where domain expertise matters.

Recognition: Kentucky AI is an NVIDIA Inception program member. OpenTakeoff ranked #1 Tool / #2 overall on the BUIDL_QUESTS leaderboard on October 7, 2026 (Attention Award reference standings).

Applied AI & evaluation

Domain-specific model development, failure analysis, data quality, and human review.

AI products & agent systems

Tools that connect model capabilities to the workflows and decisions of real users.

LET’S BUILD SOMETHING USEFUL.

Hard problems.
Real-world impact.

Michael Edlin · Kentucky, USA
GitHubHugging FaceYouTubeKentucky AI