The RedMark Engine

A multimodal AI system, purpose-built for M&E drawings.

Generic OCR fails on technical drawings. Generic LLMs hallucinate on engineering tables. RedMark is built differently — a custom OCR pipeline, computer vision trained on tens of thousands of industry drawings, and reasoning models tuned for engineering rules. Together, they read drawings the way your senior engineers do.

INPUT Raw PDF M-201 LAYER 01 Custom OCR rotated · layered · dense LAYER 02 Domain CV VAVs · panels · symbols LAYER 03 Reasoning engineering rules OUTPUT Ranked issues VAV-1 keynote Panel MCA Panel overload Feeder mismatch Text over linework + 7 more → RedMark Report
How we built it

Three core principles.

Security, reliability, and explainability — the three things we found hard to achieve with a pure LLM approach, so we didn't build one.

01

Security

Computer vision, OCR and deterministic logic extract and structure the drawing locally, before an LLM is involved. Only minimal, context-free metadata is sent to it — your original drawings, title blocks, company names and project details are never transmitted.

See the security model
02

Reliability

Engineering rules are applied deterministically, not inferred. RedMark recalculates values from first principles — heater sizes from airflow and ΔT, MOCP against panel schedules — so a result either computes or it doesn't. The LLM is never the system of record.

03

Explainability

Every check produces a coordination view showing the two sources RedMark compared — detected issues in red, identified equipment in green. Rather than asking you to trust a black-box decision, we show exactly how each conclusion was reached.

See it on a drawing
The problem

Why off-the-shelf AI can't read your drawings.

Generic AI fails because
RedMark solves it by
OCR engines built for office documents miss rotated, low-contrast, or layered CAD text.
Custom OCR pipeline tuned for dense technical notation, layered output, and engineering-table layouts.
Generic computer vision doesn't recognise VAVs, panel schedules, P&ID symbols, or one-line diagrams.
CV models trained on tens of thousands of M&E drawings — discipline-specific symbols, not generic shapes.
LLMs hallucinate on engineering schedules and apply consumer-grade reasoning to safety-critical rules.
Fine-tuned reasoning that applies engineering rules deterministically — LLMs only handle narrow, structured tasks.
Sends your drawings to third-party LLMs that train on your data.
Your drawings, title blocks and company names are never sent to an LLM. It only ever sees pre-extracted, context-free metadata.
The engine

How the engine works.

01

Custom OCR

Reads what others miss: rotated text, dense schedules, layered CAD output, tables. Tuned on tens of thousands of M&E drawings — built where generic engines fall over on rotated, low-contrast, and layered text.

02

Domain-trained computer vision

Recognises VAVs, panel schedules, diffusers, valves, instruments, symbols, and callouts native to MEP and P&ID. Not generic shape detection — discipline-specific.

03

Engineering-rule reasoning

Cross-validates against rules — keynote completeness, schedule-vs-plan coordination, panel-power-plan integrity, panel load limits, feeder-to-breaker integrity. LLMs handle narrow text tasks; the system of record is deterministic rule logic.

Accuracy

Built for engineering precision.

Tuned on real drawings

Custom OCR + computer vision tuned on tens of thousands of real M&E drawings — built where generic engines fall over on rotated, low-contrast, and layered CAD text.

Engineering rules encoded

Rules across mechanical, electrical, and P&ID, extended with customer-requested checks.

Validated across regions

Validated against drawings from MEP firms across the US, Canada, UK, and Western Europe.

The team

20 years of production AI. 25+ years inside MEP.

KS

Shrikrishna Srinivasan

CTO

20 years building production AI. Texas Instruments. AutonomIQ (acquired by Sauce Labs). Most recently led a Vision-Language AI platform for manufacturing.

AV

Arvind Viswanath, PE

Founder

25+ years across MEP and manufacturing. 3M. National Laboratories. Hundreds of complex projects. Knows what your senior engineers actually check for — because he is one.

Security

How does this all stay secure?

CV, OCR and deterministic logic structure the drawing first. Only minimal, context-free metadata reaches an LLM. Enterprise deployments with contractual no-training guarantees.

See our security model
AES-256 at rest TLS 1.3 in transit No model training SOC 2 — in progress GDPR-ready
Talk to engineering

Want the technical deep-dive?

Book a 30-minute session with our CTO. We'll walk you through the OCR pipeline, the CV training set, and the engineering-rule reasoning layer — and answer the questions you'd ask any AI vendor before signing.