Automating the Roller Shade Takeoff Process — Findings from initial technical validation | 7/25/26
Findings Presentation
What I tested, what I found, and what it means for the build.
The Challenge
Takeoff bottleneck & estimator goals
What Was Tested
Real sample data from DPS_Sandoval
Four Findings
Tag extraction, dimensions, AI accuracy, competitive landscape
Build Implications
What's ready vs. what needs validation
The Lu-Tek Estimation Challenge
The Bottleneck
Takeoff averages 4–6 hours per bid which is a time-consuming step in estimating. That time is better spent on site visits and higher-value work.
The Goal
Objective: remove Bluebeam from the workflow
The Target
80% of projects are "simple," 20% are complex. The simple bucket is the near-term automation target. The highest-leverage starting point.
What Was Tested
The Sample File
Real data from the DPS_Sandoval plan set — AC102 shade plan sheet, finish schedule, and your completed CSV takeoff. Not a hypothetical.
Two Core Questions
- Can we count and classify shade tags automatically from the plan file? 1.
- Can we derive width and height dimensions automatically without manual measurement? 2.
Goal: determine what's reliably automatable vs. what still requires human input.
Quantity & Tag Extraction — Validated ✓
Method
Extracted shade tags directly from the file's vector text layer, not OCR, not an image guess. Real embedded data.
Result
5× RWS-1A + 3× RWS-1B = 8 total, matching your CSV exactly (ML1 qty 5, ML2 qty 2, ML3 qty 1).
Why It's Reliable
Fully deterministic with zero hallucination risk. The system reads real embedded text rather than interpreting pixels.
The Dimension Gap Is Real and It's the Pain Point
I searched the full vector text layer for the actual width/height numbers (48, 64, 78, 104). They don't exist as text anywhere in this file. The finish schedule explicitly lists shade sizing as "VARIES; REQUIRES VERIFICATION IN FIELD."
Confirmed
Your estimators are deriving dimensions themselves, not reading them off a table, which was as described in our meeting.
Implication
Dimension automation is not a "read the PDF better" problem. It's a "measure geometry against a calibrated scale" problem which is a different engineering challenge entirely.
Why Generic AI Tools Landed Arent As Accurate
Raw LLM-reads-an-image approaches eyeball pixel distances against a flattened picture, not measuring against real scale. This is why general-purpose AI underperforms on this specific task.
Reliable Architecture
Separate two jobs: deterministic code handles geometry/measurement math; AI handles language tasks — reading labels, matching shade types, flagging ambiguous cases.
Kreo Test Results
At ~77% confidence, Kreo returned close to the known 8 shades. Width accuracy landed within 1–3" of the known 48" dimension strong for majority-size shades.
Key Limitation
One shade missed the wider RWS-1B (~78" class). Lowering threshold from 50% → 23% to recover it nearly doubled detections (33 → 65). No single threshold cleanly returns all 8 with zero noise.
The Competitive Landscape
Togal.ai
Confirmed via their own rep that no existing tool solves this end-to-end today. No API. Still requires a manual component. Faster than Bluebeam, but not the automated fix Lutek needs. Worth knowing as a Bluebeam alternative — separate from what we're building.
Kreo
Tested directly against the real Sandoval file. Width-only per documentation (height never populated). Strong width accuracy (within 1–3") on detected shades. However, the precision/recall tradeoff requires a downstream review/filtering layer, not just a slider adjustment.
Feasibility: LLM (Claude) Alone vs. LLM (Claude) + Kreo
A clear comparison of what each piece of the architecture can and can't do on its own:
What This Means for the Build
Neither Claude nor Kreo solves the problem alone. The proposed system leverages their combined strengths: Kreo for raw geometric measurement and Claude for reading tags, filtering output, and flagging missing data.
Phase 1: Automated Tagging
Proven and ready for build: automated shade tag counting and classification. This eliminates manual tallying with zero hallucination risk.
Phase 2: Dimension Derivation
Validated approach, needing broader testing: dimension derivation via vector-geometry measurement against the drawing's own scale reference. This is technically sound but requires confirming across a wider sample of your real bids before full build.
No Bluebeam Dependency
We are explicitly not pursuing a Bluebeam-dependent architecture, aligning with your goal of removing Bluebeam from the workflow entirely.
Flag for Field Verification
Where dimensions are genuinely unrecoverable from the drawing (e.g., "VARIES" fields), the system will flag these for field verification, directly supporting your goal of freeing up time for site visits, not replacing them.
What This Means for the Build
Phase 1 — Ready to Build
Automated shade tag counting and classification. Eliminates manual tallying. Zero hallucination risk.
Phase 2 — Validated Approach
Dimension derivation via vector-geometry measurement against the drawing's own scale reference. Technically sound but needs broader sample confirmation.
Flagging, Not Guessing
Where dimensions aren't recoverable ("VARIES" fields), the system flags for field verification with additional context from site visits.
What Phase 2 Actually Includes
Phase 2 of the system will build upon the foundational elements to deliver a comprehensive solution, automating several critical steps in the estimation process:
Automatic Bid Intake
Pulling plan sets directly from BuildingConnected, eliminating manual file downloads and ensuring new bids enter the pipeline seamlessly.
Smart Page Sorting
Automatically identifying relevant sheets (e.g., shade plans, schedules) within a plan set to focus processing only on necessary pages.
Precise Width Detection
Measuring shade dimensions directly from drawing geometry with high accuracy, building on our validated testing methodology.
Missing-Info Flagging
Clearly flagging shades where size information is genuinely absent from the drawing for efficient field verification, rather than guessing.
Estimator Review Screen
A user-friendly interface displaying every result with a clear confidence level, empowering your team to trust or double-check with efficiency.
GC-Formatted Output
Delivering results in the precise layout and format your general contractors already expect, streamlining communication.
Ongoing Accuracy Monitoring
Continuously validating system accuracy against real bid outcomes over time to ensure proven, reliable performance.
Recommended Next Step
The Recommendation
A scoped, paid pilot phase to validate the dimension-extraction approach across a representative sample of simple-bucket bids — not just one file.
Why a Pilot First
One clean sample file tells us the approach can work. It doesn't tell us it will work across the variation you described: different title block formats, missing scale references, inconsistent drawing quality, et.
Pricing a full system off a single sample risks under-scoping or overpromising. A pilot protects both sides: you get a validated number before committing budget; I get real variance data before committing engineering time.
Pilot-Phase Proposal
Objective: validate that the vector-geometry measurement approach — proven on one file — generalizes reliably across a representative sample of Lutek's typical "simple" bids, before committing to the full build price.
Sample Files
Lutek provides 3-5 real "simple" bid packages (PDF plan sets + completed CSV takeoffs as ground truth), weighted toward typical volume, not best-case files.
Testing
Tag counting accuracy, scale-calibration feasibility per file, and whether the precision/recall tradeoff holds, and whether a rules-based filter can clean up Kreo's raw output reliably.
Deliverable
A written accuracy report per sample with confidence-tagged results: what can be automated now, what still needs field verification, and why.
Pilot Scope — What's In and Out
✅ In Scope
- Shade tag counting/classification accuracy across all samples
- Scale-calibration feasibility on each file
- Precision/recall tradeoff analysis across broader sample
- Rules-based filter testing (schedule cross-reference, tag-pattern matching)
- Written accuracy report with confidence-tagged results per file
🚫 Out of Scope
- No production system, dashboard, or GC output formatting
- No commitment to a specific third-party tool (Kreo or otherwise) until accuracy is confirmed on real files
- No in-house API integration. This remains a separate Phase 2/3 conversation, scoped once the core takeoff-assist piece proves out
Timeline & Investment
Timeline
1 week from receiving sample files. To be confirmed based on Lutek's bandwidth.
Pilot Investment
$3,500 — paid, non-refundable, but fully credited toward Phase 2 build if you proceed. Reflects genuine technical R&D beyond a standard audit.
Full Build (Post-Pilot)
$20,000–$35,000 upfront + $2,000–$4,000+/month ongoing
Covering: third-party API costs, review-dashboard maintenance, and confidence-tagging system upkeep.
If accuracy holds →
Proceed to full build proposal with a firm, defensible price
If accuracy doesn't hold on certain drawing types →
Report back honestly on which bid types remain human-assisted vs. automated, and adjust scope/price accordingly. No obligation to proceed until this is known.
A Note on These Findings
All findings above are based on one validated sample file (DPS_Sandoval). The pilot phase exists specifically to confirm these results generalize. This is presentation-ready as "strong initial findings," not as a validated production accuracy rate across Lutek's full simple-bid volume.
What's Proven
Tag extraction is deterministic and reliable. The dimension gap is real and confirmed. Generic AI tools have a measurable accuracy ceiling.
What the Pilot Confirms
Whether vector-geometry measurement generalizes across real bid variation, different formats, missing scale references, inconsistent drawing quality.
The Path Forward
A paid, scoped pilot protects both sides before committing to a full build. Strong initial findings. Clear next step. No guesswork.
Lutek Shading Systems Findings & Pilot Proposal