MANUFACTURING INTELLIGENCE · HMLV PRECISION SHOPS

Every drawing,
decoded.
Every job,
remembered.

Contract manufacturers lose hours interpreting client drawings and lose years of process knowledge when senior engineers leave. VOXXL solves both — an AI-first system built for the shop floor to decode manufacturability, suggest exact process routing, and permanently archive plant knowledge.

CLIENT DRAWING — MB-BKT-2291-R3
ASME Y14.5 · EN19 STEEL
What's the tightest tolerance and will it be difficult to hold?
Tightest: ±0.02mm on the 45mm bore (datum A–B).
Requires finish boring pass. JB-1847 had identical tolerance — chatter on R2 pass fixed by -15% feed.
Suggest exact process routing
1. Rough turn (EN19 billet)
2. Heat treat (32–36 HRC)
3. Finish bore (Mazak, CBN insert)
40min

Average time a junior engineer spends interpreting one drawing

16K+

MSMEs in India's defence supply chain receiving new design files

0

Purpose-built tools for manufacturer-side drawing interpretation globally

5%

Annual revenue leaked by precision manufacturers through slow quoting

High-mix manufacturing runs on institutional knowledge nobody has captured

Every time a new client drawing arrives, someone must read it, understand it, and recall whether the plant has done anything like it before. Right now, that someone is one overloaded senior engineer.

01

Drawing Interpretation is Slow

A junior engineer reading a GD&T-heavy drawing from a foreign client can take 40+ minutes — and still miss critical callouts. Every misread tolerance is a scrapped part or a client complaint.

02

Process Knowledge Walks Out the Door

The setup tricks, the feed rate adjustments, the tooling choices that made a difficult job work — none of it is written down. When senior engineers leave, the plant starts from zero on every similar job.

03

12–18 Month Onboarding

New hires take over a year to become useful. Onboarding means shadowing one person and hoping to absorb enough. Senior engineers spend half their time answering questions instead of doing their jobs.

04

Critical Feature Blindness

Not all tolerances are equally difficult. Seniors scan and flag the features that determine feasibility. Juniors read every dimension with equal attention — and miss the ones that actually matter.

05

Specification Cross-Referencing

A single aerospace drawing can reference 10–15 external documents — AMS, ASTM, MIL-SPEC, Boeing BPS. Each requires manual lookup, reading, and compliance verification. This invisible workload consumes hours per job.

06

Format Incompatibility

Clients send CATIA, SolidWorks, NX, STEP, IGES, PDFs — sometimes split across multiple files in different revision states. Reconciling these manually eats 15–30 minutes per job before any engineering even begins.

Interpretation and memory, in one layer

VOXXL sits between receiving a client file and beginning process planning. Upload the drawing, ask anything, and surface every relevant job your plant has ever done.

CORE

Drawing Interpretation & Manufacturability

Upload a STEP file and 2D PDF. VOXXL extracts tolerances, GD&T callouts, surface finish requirements, and material specs. It flags what's difficult to machine and suggests the exact process routing required to build it — in seconds, not 40 minutes.

CORE

Natural Language Q&A

Ask the drawing anything in plain English. "What's the tightest tolerance?" "What does this GD&T symbol require?" "Which surfaces need finishing?" Answers cite the exact drawing section.

MOAT

Plant Memory — Job Archive

Every job processed builds a searchable, structured archive. Process notes, tooling choices, problems encountered, solutions found — all attached to the drawing, all retrievable by any engineer at any time.

MOAT

Similar Job Retrieval

When a new file arrives, VOXXL surfaces the most similar past jobs automatically. Same material, similar tolerances, related geometry — with process notes from those jobs immediately available.

voxxl · JOB INTELLIGENCE
voxxl› load MB-BKT-2291-R3.step + .pdf
Title block extracted
47 dimensions parsed
8 GD&T callouts identified
Material: EN19 · HT: 32–36 HRC
Critical: Ø0.02 true position (datum A–B)
Ra 1.6 on bore surface — finish pass req.
voxxl› find similar-jobs
Match 94% · JB-1847 · March 2024
EN19, ±0.02 bore, same datum structure
Note: chatter on R2 pass → -15% feed
Match 87% · JB-1621 · Oct 2023
EN19 hardened, Ra 1.6 bore
CBN insert used for post-HT finish
voxxl›

Eight systems. One intelligence.

VOXXL doesn't add another tool to your stack. It replaces the slowest, most error-prone parts of your feasibility workflow with a precision-engineered AI pipeline.

Multi-Format Intelligent Ingestion

Accepts every format — STEP, CATIA, SolidWorks, NX, PDF. Automatically reconciles split packages and flags revision conflicts.

OpenCASCADE · PyMuPDF

Feasibility Intelligence

Tolerance criticality maps, GD&T decoded to plain English, surface finish requirements, and manufacturability scoring.

YOLOv11 · Florence-2

Plant Capability Scoring

Every drawing is scored against your plant's true machine capabilities. Green, amber, or red — with the specific constraint identified.

Llama-3 · Matrix

GraphRAG Institutional Memory

A structured knowledge graph connecting jobs, materials, tolerances, machines, tools, problems, and solutions — traversed to answer multi-hop questions.

Neo4j · GraphRAG

Specification Cross-Reference

Identifies external specs from notes, links to your local library, and extracts the exact clause — displayed inline next to the feature.

Local RAG · Qdrant

Confidence-Gated Review

Low-confidence items are routed to a human review queue. Every correction retrains the model on your plant's drawing conventions.

Active Learning

Voice-Captured Process Notes

Engineers speak for 30 seconds after a job. Whisper transcribes locally, extracts structured entities, and writes them to the graph.

Whisper.cpp · Local

Tolerance Boundary Verification

Machine manuals are parsed once. Plant managers annotate real-world limits. New drawings are checked against earned experience.

Manual Ingestion
1

Ingest

All file formats accepted, unified, and reconciled automatically.

OpenCASCADE · PyMuPDF
2

Extract

GD&T, dimensions, tolerances, surface finishes, and specs decoded.

YOLOv11 · Florence-2
3

Retrieve

Similar past jobs and applicable specs retrieved from memory.

Neo4j · Qdrant
4

Score

Drawing checked against plant capability — green, amber, or red.

Llama-3 · Matrix
5

Dashboard

Structured analytical view with everything you need in one screen.

React · Next.js

Built on research.
Measured by results.

+0%

Precision increase over GPT-4o on GD&T extraction

+0%

F1-score improvement on drawing understanding

under 0s

From RFQ ingestion to feasibility verdict

0%

Data stays local — zero cloud exposure

An air-gapped intelligence engine.
No data leaves your building.

VOXXL runs entirely on a local server inside your plant's intranet. Every model, every database, every computation — local. Updates arrive as signed packages through a one-way tunnel.

  • Air-Gapped Deployment
    A dedicated local server plugged into your intranet. No outbound connections. No cloud APIs. Zero data exposure.
    Docker · NVIDIA RTX · One-Way Update Tunnel
  • Two-Stage Vision Pipeline
    YOLOv11 detects and bounds symbols. Florence-2 interprets semantics. Separation of detection from interpretation eliminates hallucination.
    PyTorch · Ultralytics YOLOv11 · HuggingFace VLM
  • GraphRAG Memory Engine
    Neo4j stores structured relationships. Qdrant stores unstructured spec text. LangChain orchestrates multi-hop retrieval.
    Neo4j · Qdrant · LangChain · Microsoft GraphRAG
  • Local LLM Reasoning
    Llama-3 8B quantized to 4-bit runs on local GPU. Smart enough to follow strict JSON schemas, small enough for edge deployment.
    Ollama · Llama-3-8B-Instruct · 4-bit Quantization
Frontend & Dashboard

Interactive analytical view with overlay detection boxes and structured feasibility brief.

Next.jsReactTailwindCSSSocket.io
AI & Vision Pipeline

Symbol detection, semantic interpretation, and process routing generation.

YOLOv11Florence-2Llama-3 8BWhisper.cpp
Memory & Retrieval

Structured knowledge graph and vector store for historical job retrieval.

Neo4jQdrantLangChainGraphRAG
Ingestion & Parsing

Multi-format file translation, reconciliation, and entity extraction.

OpenCASCADEPyMuPDFFastAPI
Air-Gapped Local Infrastructure

All layers run on local hardware. One-way update tunnel only.

DockerNVIDIA RTXLocal Intranet

Three forces making this problem worse every year

India's precision manufacturing sector is growing — but the knowledge infrastructure to support that growth doesn't exist yet. VOXXL is that infrastructure.

Defence indigenisation

16,000 MSMEs receiving new design files. India's defence production hit ₹1.27 lakh crore in FY24. Every indigenised component is a drawing that arrives at an Indian job shop for the first time — often in a format never seen before.

PLI scheme

New plants, zero institutional memory. PLI investments of ₹1.76 lakh crore are creating hundreds of newly certified precision manufacturers. These plants are young, have no senior engineers yet, and are immediately receiving complex high-spec client work.

China-plus-one

New global clients, unfamiliar standards. Global OEMs are actively moving precision sourcing to India. Indian job shops are winning contracts with US and European clients who send drawings to ASME or ISO standards the plant's engineers have never worked with before.

Workforce attrition

Senior engineers are retiring. The generation of engineers who built India's precision manufacturing capability is leaving. Their knowledge — accumulated over 20–30 years — has no home. VOXXL is that home.

Built specifically for High-Mix, Low-Volume precision contract manufacturers

Not generic job shops. Not Tier-1 auto-component clusters. The plants that receive a different part from a different client almost every week — and must interpret each drawing from scratch.

Aerospace & Defence

AS9100-certified shops supplying HAL, DRDO, global primes (Boeing, Airbus, GE). Complex first-article parts, tight tolerances, new designs constantly. First-article inspection required on every new job.

JAIVEL AEROSPACE · CIM TOOLS · GEMSONS

Industrial & Export Job Shops

Precision machining for oil & gas, semiconductor tooling, industrial automation — often for US or European clients sending drawings to ASME standards. Multiple clients, no two jobs alike.

PUNE · COIMBATORE · RAJKOT CLUSTERS

Medical Devices

PLI-driven growth, export-oriented, ISO 13485 certified. High-spec components — implants, instruments, housings — where each order is a different part, a different material, a different tolerance stack.

CHENNAI · BANGALORE · HYDERABAD

Why VOXXL isn't
another quoting tool.

Existing tools produce a price or a markup. VOXXL produces a structured feasibility intelligence brief — the thing senior engineers actually need.

Capability Standard ERP / MES CADDi / CoLab VOXXL
2D Drawing GD&T Extraction None Partial Full (YOLO + VLM)
Manufacturability & Process Routing None None Automated, Plant-Specific
Institutional Memory (GraphRAG) None Vector Search Only Structured Knowledge Graph
Plant-Specific Capability Scoring None None True Machine Limits
Air-Gapped / Local Deployment Yes Cloud Only 100% Local
Voice-Captured Process Notes None None Whisper, Local

Your plant's knowledge
deserves a permanent home.

We're onboarding a small number of precision job shops for early access. If your engineers interpret new client drawings regularly — VOXXL is built for you.

CURRENTLY IN VALIDATION · INDIA