Project OpenCoder: Strategic AI Independence Initiative

Prepared for: Comfac Technology Group & Cornersteel Systems Corp.

Objective: Transition from renting intelligence ($1,200/yr/seat) to manufacturing intelligence using a self-hosted, open-source ecosystem that powers both software development and physical production.

Executive Summary

The Pivot: Moving beyond simple "Chatbot" replacements to establish the Comfac Matrix Model (CMM) Business Unit—a proprietary engine that automates not just software code, but the physical design and manufacturing processes (via FreeCAD/Python) for the group and partners.

The Strategy

Open Source Core (The Stable Release): Release "Stable" models to the public, building a community of students and developers who learn the stack, effectively becoming a distributed R&D department.
Partner Premium (The Cutting Edge): Reserve "Cutting Edge" models—trained on latest data and capable of faster iteration—for Comfac and paying partners (e.g., Cornersteel, Government).
Funding Mechanism: Partners receive Exclusive Access, allowing their teams to be trained on and utilize "Cutting Edge" resources.

Financial Model

Operational Expenditure (OpEx): ₱15M per year upon release of Open Source and Cutting Edge models
Partner Commitment: ₱5M/year in Partner Access fees (less than $100,000 USD per year)
Partner ROI: Access to high-end unified memory AI PCs capable of running full-sized models, with training and customization for specific corporate requirements, plus highly optimized tiny to small AIs for departmental usage
Overage Billing: Resources beyond initial commitment billed at ₱5,000 per hour
Reinvestment: Revenue funds project growth: hiring, infrastructure, marketing, community training

Target Outcome: Reduce per-seat AI cost to $200/year while establishing a new revenue-generating Business Unit that creates intelligence for future Comfac robotics.

The New Business Unit: CMM (Comfac Matrix Models)

Mission: "Dev, Design, and Manufacture."

The CMM Unit's mandate is to enhance the ability to create the intelligence for the robotics and systems Comfac will build.

Core Products (2026 - 2027 Strategy)

These foundational products fuel revenue growth required to develop advanced Robotics and VLA models (Stage 3), which require 1-2 years to reach market readiness.

1. The "Coder Matrix" (3B-16B Coding LoRA)

The Concept: A fine-tuned coding assistant replicating "Claude Code Premium" capabilities without per-seat subscription cost.

Immediate Impact (2026): Enables Comfac to rapidly turn around software solutions, directly aiding in capturing sales.

Key Beneficiaries & Applications:

CSC Facilities: Rapid deployment of management services and custom integrations (2026)
Frappe Ecosystem: Rapid creation of Apps and ERPNext customizations (2026)
NextCloud: Custom App development for partner network (2027)
TrueNAS: Creation of apps and middleware integrations (2027)
"Steward": Philippine home-grown Home Assistant for facility control—tracking all solar panels in CGG with real-time energy usage visible up to every panel board
Secada: Enhanced local-made system (like PaperlessNG, open source) with improved image recognition, running on smaller GPUs until reaching sub-50k PHP PC with 8GB model. Used by all CGG departments.
Synopsis: Comprehensive OpenCLAW version consolidating all messaging systems, allowing custom tools to organize/search messages and create reports. Used by CGG to remove need for constant Viber checking—ensuring emails reach everyone across platforms with singular indexed source of truth.
Synx: Scheduling software for workers, manpower organization, work breakdown documents, and workload prediction. Used by schools for class scheduling and businesses for monitoring work schedules.
Data Cleaning Tools: Secada and spreadsheet importation tools for Frappe app migration. Comfac Accounting migrated to ERPNext with tools automating data structuring, extraction, and encoding—processing 2-3 companies per month per specialist.

2. The "FreeCAD Matrix" (CAD Automation)

The Concept: Parametric models designed to Draw and Create 3D assets automatically.

Timeline: Full benefits by End of 2026 as workflows migrate to FreeCAD.

Key Innovations:

Path Procedural Design Workbench: Engineers draw a simple path; AI automates generation of pipes, ducts, cabling, including intersections, mounting brackets, and supports.
Civil Works Automation: Procedural room design and layout generation.
Output: Generates 3D Models, 2D Shop Drawings, and BOM (Bill of Materials) instantly.

The Proprietary Asset: Not just AI model weights, but the Technique: The Order of Operations, Business Processes, and Internal Workflows that automate detailing of Fitouts, Engineering, and Container Customization (SBC).

Technical Architecture (The Stack)

Hardware (The Engine)

GPU Strategy: Consumer AMD Radeon cards and next-generation Unified Memory architectures (High Memory per Dollar).

Level	Hardware	Price	Purpose
Entry	Radeon RX 7600 XT (16GB)	~₱22,000	Good for Inference
Production	Radeon RX 7800 XT (16GB)	~₱35,000	Faster token/sec
Next-Gen AI PCs	AMD APUs with integrated TPU/NPU	—	Unified Memory sharing system RAM with AI models

Next-Gen Advantage: Unified Memory eliminates strict VRAM limitations, empowering deployment of massive, full-sized models natively on consumer-grade hardware.

Infrastructure:

The Solar Data Center (Cabuyao): Utilizing Cornersteel's solar capacity to run inference at near-zero OpEx.

Software (The Brain)

Component	Technology
Base Model	Qwen 2.5 Coder 3B/32B (The Engine)
"Matrix" LoRA	Custom adapter trained on: Cornersteel manufacturing SOPs, FreeCAD Python Documentation, Comfac Legacy Codebases (Laravel/Node)
Deployment	Ollama + OpenCode CLI (for Devs) / Custom WebUI (for CAD operators)

Implementation Roadmap

Stage 1: Benchmarking & Setup (Weeks 1-4)

Goal: Prove Qwen 2.5 can handle Code + Basic CAD scripting.

Hardware: Purchase 1x RX 7600 XT (~₱22,000)
Task 1 (Software): Benchmark Qwen against Claude Sonnet on Laravel Controller task
Task 2 (Hardware/CAD): Benchmark Qwen on generating simple FreeCAD script ("Draw a cube with a hole")
Deliverable: Report identifying "Weakness Vocabulary" (what the model fails to understand about specific workflows)

Stage 2: POC - The "Matrix" Prototype (Weeks 5-8)

Goal: Create first "Comfac Matrix Model" (LoRA).

Training Data:

Export Git logs from Comfac (Software patterns)
Create dataset of "Text Description -> FreeCAD Script" pairs
Fine-tune Qwen 2.5 3B using Unsloth
RAG Setup: Ingest Cornersteel machine manuals and SOPs into Knowledge Base

Stage 3: Internal Deployment & OJT Swarm (Month 3-6)

Goal: The "Open Source Ecology" in action—developing ecosystem of specialized models and tools.

The OJT Army: 10-20 Students/OJTs given access to Stable model release. Exchange: They get free, unlimited AI; We get their error logs and patches to improve Cutting Edge model. If we have enough success we can roll it out to school partners who will help us improve the LoRA and create better models.

Crowdsourced R&D (Model Expansion):

Visual Models: Facility inspection, defect recognition, site monitoring
CAD Models (The "Matrix" Core): Hybrid of Coder and Vision Language Action (VLA) models trained to:
- Control FreeCAD API
- Interpret uploaded PDFs/images and "draw" in FreeCAD
- Convert 2D sketches into 3D parametric models
- Generate BOM and Spreadsheets automatically
- Navigate various Workbenches (PartDesign, Arch, Draft)
Vision Language Action (VLA) Models: Connecting visual input directly to digital execution
Action Models (Robotics): Trained in Godot game engine to simulate physical movement, hardware interaction, kinematics before real-world deployment
Audio Models: Voice-commanded site logging, machine diagnostics, hands-free interface
Expert Models: Specialized knowledge bases as "Expert Systems" for niche domains (HVAC troubleshooting, Electrical load calculation)
Engineering Design Models: Orchestrating CAD models, interpreting design outputs, organizing unstructured data, comparing designs against Building Codes (NBCP)

The Tooling Expansion (Software Ecosystem):

Comfac Custom Workbenches: Proprietary FreeCAD environments
FreeCAD-Blender Bridge: Plugins automating import/export, preserving metadata bi-directionally
PandaPower App: FreeCAD/Blender integration for power system analysis and grid simulation
Geometry Node Interoperability: Migrate geometry nodes between Blender and FreeCAD
Godot "Tri-App" Workflow: Plugins for seamless transfer between FreeCAD (Design), Blender (Visualization), and Godot (Simulation)

The Hardware Cascade (Upgrade Cycle):

Senior Staff Upgrade: 8-12GB VRAM → 16-20GB VRAM workstations
Junior Inheritance: 8-12GB hardware trickled down to Junior staff and Trainees

Stage 4: Commercialization & Partnerships (Month 6+)

Goal: Revenue Generation.

Partner Offer: "Sustainable AI Integration" (₱5M/Year ROI)

For less than $100,000 USD per year, partners receive:

Dedicated High-End Compute: Unified memory AI PCs with exclusive access to proprietary LoRA-trained models
Bespoke Full-Size Training: Customization to partner's proprietary requirements
Distilled Edge Solutions: "Tiny to small" AIs optimized for localized, niche usages

Service Example: Cornersteel Facilities Management utilizing deployed "CMM Nodes" for Vision Model maintenance scanning, cost-auditing apps, on-demand embedded system drafting.

The Marketplace:

Rent out spare compute from Cabuyao Solar Data Center to students/startups
"Plug & Play" AI Rentals for schools

Transition to Model Customization

The Evolution: Moving beyond simple LoRAs to building Cutting Edge Custom Models from scratch using Deepseek and Qwen Open Weights.

Financial Shift:

Model-building operational costs increase by ₱3M - ₱4M per year
Why: Deep customization requires higher compute density and data cleaning

Workforce Specialization:

MM Scientists: Focused on architecture, training, and weights of AI models
MM Designers: Focused on parametric logic, "Technique," and UX of AI-controlled tools

Financial Impact & Efficiency Analysis

Current Baseline:

Design & Engineering: ₱10M - ₱15M per year on Engineering and Design talent (~30 engineers)
Facilities & Maintenance: ₱10M per year on 20 engineering and maintenance staff

A. Design Automation (The 10x Lever)

CMM equips 1 designer to generate production elements at 10x speed of industry standard.

Metric	Traditional Workflow	CMM Powered Workflow
Annual Talent Spend	₱16M (30 Engineers)	₱16M (Same Team, Supercharged)
Output Equivalent	₱16M Value	₱50M Value (3.1x Realized Gain)
Production Speed	1x (Manual Logistics, Redesign)	10x (Parametric Gen, Auto-Paperwork)
Cost Per Revision	High (Man-hours)	Near Zero (Compute-seconds)

Note: Conservative ~3x realized gain due to human communication bottleneck between production and construction teams.

B. Facilities & Maintenance (The 2x Lever)

AI impacts the management layer—design, monitoring, scanning, vendor coordination—rather than physical execution.

Metric	Traditional Workflow	CMM Powered Workflow
Annual Talent Spend	₱10M (20 Staff)	₱10M (Same Team, Optimized)
Output Equivalent	₱10M Value	₱20M Value (2x Realized Gain)
Key Efficiency	Manual Inspection	Vision Model Scanning & Auto-Ticketing

C. The Robotics Frontier (Future CapEx Reduction)

Ability to create LoRAs and modify models = ability to create Robotic Intelligence.

The Godot Division: Operating at ~₱2M per year, simulating specific work environments for CGG and partners.

Environment LoRAs: Simulations generate specialized LoRAs allowing robots to "know" a facility before physically entering.

Phase 1: Navigation, scanning, object retrieval
Phase 2: Toxic/hazardous tasks (bioreactor maintenance, sewage/pipe infrastructure)

Execution: Combined with Design Automation, enables Comfac to build its own robotics to further automate Cornersteel's manufacturing lines—moving from "Software Automation" to "Physical Automation."

D. Organizational & Strategic Impact

Retention & Work-Life Balance: Automating "grunt work" eliminates overtime and burnout → exceptionally high retention rates
Flipped Bottlenecks: Sales becomes the new bottleneck as design throughput increases dramatically
Design-Led Sales: Designers join Sales in lead generation with real-time generative tools
Multidisciplinary Expansion: Absorb new requirements—Hydraulic & Sanitary (Water), Electrical (Power), Structural, Certifications
The New Challenge: Constraint shifts to "Hiring Certified Professionals" to sign off on AI-generated work volume
Unique Market Position: Only Philippine engineering firm balancing Engineering and IT, manufacturing its own intelligence

Immediate Action Items

Strategic Briefing: Present full "Project OpenCoder" roadmap to core leadership and technical leads
Operational Planning & Resource Allocation:
- Audit current engineering workloads
- Select internal leads (MMDs) for Data, Coding, and Manufacturing/CAD tracks
- Detail plan and assign roles before spending
Hardware Determination: Finalize hardware outlay based on identified personnel (CAD-focused vs. Code-focused setups)
Procurement: Purchase initial hardware batch (Pilot: AMD Radeon RX 7600 XT) once team is ready
Execution & Monitoring: Launch Stage 1 Benchmarking with weekly reporting cadence

Appendix: The Matrix Model Assets

Reference Links

FreeCAD Matrix Models: https://sites.comfac.net/freecad.html
Description: Models designed to Draw and Create 3D Parametric models

Proprietary Definition

The "Secret Sauce" of CMM Business Unit is not open-source code, but the Technique:

Order of Operations: Specific sequence for manufacturing steps
Business Processes: How design moves from prompt to physical Bill of Materials
Internal Workflow: Automation of detailing for:
- Fitout Construction
- Engineering Schematics
- Container Customization (SBC - Smart Building Components)
- Open Source Ecology Production

Drafted by the Comfac AI Strategy Team