# g factor technologies

> g factor is a private Generative AI post-training, evaluation, and deployment platform. It enables organizations to post-train and serve open-weight language models using Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) inside customer-controlled private cloud or on-premises environments without sending sensitive data to third-party model APIs.

## Overview

g factor turns proprietary enterprise data, expert feedback, edge cases, and engineering workflows into owned model assets. Rather than relying on public model APIs with external data leakage risks and unpredictable pricing, organizations use g factor to train, qualify, export, and privately serve specialized AI agents.

Key capabilities:
- Supervised Fine-Tuning (SFT): Distill trusted examples and domain traces into portable LoRA adapters before reinforcement learning.
- Reinforcement Learning with GRPO: Train agents directly inside stateful execution environments with typed actions and deterministic reward verifiers.
- 500+ Open Models Supported: Train on open architectures including Qwen, DeepSeek, Mistral, Llama, and more.
- Commercial OpenEnv Marketplace: Pre-built and custom executable task gyms covering native CAD, CAM, Verilog, Modelica, CFD, analog IC design, and quantitative finance.
- Reproducible Qualification: Frozen held-out benchmarks, side-by-side base vs. adapted comparisons, promotion gates, and immutable provenance.
- Private Cloud & On-Premises Serving: Deploy qualified models to customer-governed infrastructure with full data residency and privacy.

## Solutions

- [LLM Fine-Tuning & GRPO Post-Training](https://www.g-ftech.com/solutions/llm-post-training): Adapt open models into specialized skills using SFT, GRPO, governed datasets, distributed GPU topologies, and portable LoRA adapters.
- [Private AI Training Environments](https://www.g-ftech.com/solutions/private-training-environments): Train specialized agents in executable OpenEnv environments with hidden deterministic checks, native tools, and machine-checkable outcomes.
- [Reproducible LLM Evaluation](https://www.g-ftech.com/solutions/model-evaluation): Qualify model capabilities before promotion using frozen held-out benchmark plans, matched policy comparisons, and immutable artifact provenance.
- [Private & On-Premises LLM Deployment](https://www.g-ftech.com/solutions/private-model-deployment): Operate approved adapters and checkpoints inside customer-controlled private cloud or on-premises boundaries for sensitive and regulated industries.
- [All LLM Solutions](https://www.g-ftech.com/solutions): Overview of the complete post-training, evaluation, and private deployment lifecycle.

## Engineering Blog & Technical Notes

- [Engineering Blog](https://www.g-ftech.com/blog): Deep-dive technical articles, systems research, and benchmarks from g factor technologies on post-training and inference systems.
- [Benchmarking Qwen-27B Across Inference Providers: Together, Fireworks, Nebius, and g factor](https://www.g-ftech.com/blog/qwen-27b-inference-providers-benchmark): Empirical systems comparison of Qwen 3.8 27B across Together AI, Fireworks AI, Nebius, Doubleword, and g factor on 2x H100 and B200 setups. Examines why single-node Tensor Parallelism hits 191 tok/s at c1 while cross-node TP stalls, how MTP4 pushes dual-H100 throughput to 769 tok/s, why MTP8 hits diminishing returns, and how prefix caching unlocks 1,140 tok/s on structured prompts.
- [The Limits of AI: Induction, Deduction, and Why Models Can't Jump](https://www.g-ftech.com/blog/ai-reasoning-limits-induction-deduction-abduction): Why LLMs master pattern recognition (induction) and logical proof (deduction) but fail at hypothesis invention (abduction). Insights from DeepMind's Tom Zahavy, with historic breakthroughs from Einstein, Kepler, and Semmelweis.
- [Physical AI: Why the Next Big Frontier Is Giving Software Agents Hands](https://www.g-ftech.com/blog/physical-ai-embodied-agents-mhs-robotics): How reasoning models connect to real hardware, treating robots and wet labs as stateful APIs, Anthropic's Model Hardware Standard (MHS), OpenAI's humanoid push, and an interactive Isaac Lab gear assembly case study.
- [Distributed Training & Inference: From CPUs and GPUs to a Cluster](https://www.g-ftech.com/blog/distributed-training-inference-cpu-gpu-clusters): An illustrated guide to CPUs, GPUs, CUDA, clusters, data, tensor and pipeline parallelism, FSDP, DeepSpeed ZeRO, Ray, and inference.
- [From RLHF to RLVR: The Evolution of Reward Signals and the Battle Against Reward Hacking](https://www.g-ftech.com/blog/rlhf-to-rlvr-reward-hacking): The journey from human preference (RLHF) and soft LLM judges to deterministic verifiable rewards (RLVR) with Lean 4, real-world reward hacking exploits, and an engineering guide to defending against gaming.
- [AsyncGRPO: Eliminating GPU Idle Bubbles in Environment-Heavy RL Post-Training](https://www.g-ftech.com/blog/async-grpo-gym-rollout-scaling): How AsyncGRPO decouples rollout generation, CPU gym execution, and policy updates with bounded staleness and host node colocation to eliminate 80% GPU idle bubbles in environment-heavy agent reinforcement learning.
- [Owning vs. Renting Intelligence: Why Enterprises Are Building Sovereign AI](https://www.g-ftech.com/blog/owning-vs-renting-intelligence): Strategic analysis of API vendor lock-in, data sovereignty for legal, medical, and financial sectors, token economics at enterprise scale ($45k/mo API vs $5k H100 node), and environment-specific post-training.
- [LoRA & DoRA: The Math, Memory, and Trade-offs](https://www.g-ftech.com/blog/lora-dora-parameter-efficient-fine-tuning): Deep dive into parameter-efficient fine-tuning, training VRAM math (324 GB vs 54 GB on 27B), weight decomposition into magnitude and direction (ICML 2024), and why DoRA achieves full fine-tuning parity.
- [Multi-Reward Reinforcement Learning for LLM Agents: Comparing PPO, GRPO, DAPO, and GDPO](https://www.g-ftech.com/blog/multi-reward-rl-ppo-grpo-dapo-gdpo): Deep dive into multi-objective policy optimization, scale dominance, reward collapse, and benchmarks showing why decoupled normalization (GDPO) outperforms vanilla GRPO in multi-reward agent environments.
- [SFT vs. RL: What Changes Inside the Model?](https://www.g-ftech.com/blog/sft-vs-rl-spectral-reasoning): An intuitive SVD guide to demonstrations, verifiable rewards, and ISO. Interactive 3D transformations and 2x2 matrix geometry distinguish nearly unchanged RLVR spectra from ISO’s explicit constraint.
- [Latent-GRPO: Reinforcement Learning in Continuous Thought Space](https://www.g-ftech.com/blog/latent-grpo-deep-dive): Deep dive into continuous thought recurrence, stochastic exploration in embedding space, literature taxonomy (Coconut, SofT-GRPO, CoLaR, SLPO, Switch), and real-world 4-arm benchmarks on Qwen3.6-27B.
- [High-Throughput LLM Inference & Training: A Deep Dive into vLLM](https://www.g-ftech.com/blog/vllm-throughput-deep-dive): Deep dive into PagedAttention, continuous batching, chunked prefill, CUDA graphs (2.95x-4.32x observed speedup on H200), and empirical GRPO rollout benchmarks comparing Hugging Face vs. vLLM (3.59x engine speedup on SQL smoke).

## Commercial OpenEnv Task Gyms

### Robotics & Physical AI
- Autonomous Driving (CARLA): Closed-loop sensor streams, trajectory tracking, collision avoidance, and traffic rule verifiers in CARLA 0.9.15.
- ManiSkill Articulation & Contact: Contact-rich articulated object manipulation (doors, drawers, gears) in SAPIEN / ManiSkill 3.
- LeRobot Imitation & Policy Learning: Dexterous manipulation policies across Hugging Face LeRobot and MetaWorld benchmarks.
- PX4 Drone Aerial Navigation: 3D waypoint navigation, wind disturbance rejection, and corridor bounds for quadcopters.
- MuJoCo Dynamic Balance & Locomotion: Dynamic gait stabilization and terrain traversal for bipedal and quadrupedal systems.

### Electronics & Semiconductors
- Verilog RTL Design & Verification: Synthesizable digital circuits validated with Icarus, Verilator, and Yosys/SKY130 simulation and formal timing checks.
- Chip Physical Design (OpenROAD): Automated floorplanning, placement, clock tree synthesis, and routing for open ASIC designs.
- SKY130 Analog IC Design: Transistor-level circuit sizing verified by ngspice simulation, DRC, LVS, and PEX.
- KiCad PCB Layout & Routing: Multi-layer component placement and differential pair trace routing verified by KiCad electrical DRC.
- Embedded Firmware & Drivers (QEMU): MISRA-compliant bare-metal firmware drivers running inside QEMU and Renode.

### Mechanical, Aerospace & Civil Engineering
- OpenFOAM CFD Optimization: Aerodynamic airfoil optimization verified on mesh validity, residual convergence, and lift-to-drag targets.
- FreeCAD CAM & CNC Planning: Collision-safe 3-axis milling toolpaths verified on reparsed ISO 6983 G-code.
- Parametric Mechanical CAD: Dimensioned mechanical solids generated from CadQuery / OpenCascade parametric source.
- Structural CAE & FEA (CalculiX): Finite element stress and deflection optimization under mechanical load cases.
- GNC & Orbital Mechanics (Orekit): Satellite orbit transfers, delta-v burns, and attitude maneuvers under strict fuel budgets.
- Physics & PDE Model Calibration (FEniCS): Calibrating partial differential equation parameters against empirical sensor data.
- Modelica Multiphysics Modeling: Multi-domain electro-hydraulic actuators verified by OpenModelica solver convergence.
- BIM & MEP Building Systems (IFC): HVAC and piping collision-free 3D routing through architectural building models.

### Energy & Industrial Systems
- Battery Charging & BMS (PyBaMM): Optimal fast-charging current profiles for Li-ion cells without lithium plating under PyBaMM DFN physics.
- Power Systems Grid Analysis: AC power flow and N-1 contingency limits under pandapower and MATPOWER simulation.
- Building HVAC Energy Optimization: Closed-loop supervisory control for multi-zone HVAC under BOPTEST and ASHRAE comfort bounds.
- Continuous Chemical Process Control: Stirred-tank reactor and distillation column control under FMI 2.0 / FMU dynamic models.
- Industrial Automation (OpenPLC): IEC 61131-3 PLC logic verified under automated plant fault injection.

### Software & Systems
- Lean 4 Theorem Proving: Mathlib theorems proved and verified by Lean's audited C++ microkernel with strict axiom checks.
- Kernel / Triton GPU Optimization: PyTorch operators optimized with Triton and CUDA gated by numerical precision and GPU kernel timing.
- SQL Transformation & Performance: Relational query refactoring and index optimization verified against DuckDB EXPLAIN plans.
- Compiler Optimization (LLVM): Peephole and loop passes verified for semantic equivalence via Alive2 and SMT solvers.
- Concurrency Race Repair: Multithreaded race and deadlock elimination verified under ThreadSanitizer and deterministic schedules.
- Desktop OS Agentic Computer Use: Multi-step enterprise workflows inside isolated QEMU desktop VMs verified by system states.
- Cloud Cost Optimization Under Strict SLOs: Cluster right-sizing reducing cloud spend while preserving p99 latency SLOs.
- Automated Software Migrations: Module refactoring across language and framework versions with automated verification of behavior preservation.

### Cybersecurity & Cryptography
- Memory-Safety Exploit Patching: Automated use-after-free and buffer overflow fixes verified under ASan, UBSan, and libFuzzer.
- Cryptographic Implementation Testing: Implementation validation against Project Wycheproof edge-case vectors, alongside separate constant-time timing leak analysis.
- Binary Reverse Engineering (Ghidra): Stripped binary decompilation and control-flow reconstruction verified on hidden oracle behavior.
- Threat Detection Engineering: Sigma and Suricata detection rule authoring evaluated on attack detection recall and false positive rates.
- Cloud IAM & Terraform Auditing: Least-privilege IAM and infrastructure security verified under OPA and Checkov.
- Solidity Smart Contract Verification: EVM smart contract vulnerability detection verified via Foundry invariant fuzzing and Slither AST checks.
- Zero-Knowledge Circuit Verification: Arithmetic R1CS circuit synthesis verified for constraint completeness in Circom and Halo2.

### Life Sciences & Materials
- Opentrons Bio-Lab Protocol Execution: Liquid-handling pipetting protocols for PCR and sample dilution verified on OT-2 geometry.
- XRD Materials Science Suite: Crystalline phase identification, Rietveld quantitative analysis, and residual stress quantification.
- CRISPR Guide RNA Design: Single-guide RNA design with systematic evaluation of off-target cleavage risks via Cas-OFFinder.
- Small Molecule Drug Discovery: High-affinity ligand design in AutoDock Vina satisfying Lipinski and ADMET properties.
- De Novo Protein Sequence Design: Candidate sequence generation verified for structural folding and stability via ColabFold and ESMFold.
- Clinical Quality Measures (FHIR & CQL): Healthcare quality measures computed across synthetic patient cohorts.
- Molecular Visualization (PyMOL): Structural 3D visualization highlighting active binding pockets and electrostatic surfaces.

### Finance, Accounting & Tax
- Cross-Sectional Alpha Generation: Alpha signal formulation tested under purged cross-validation, rank IC, and regime stress.
- HFT Market Microstructure: Execution algorithms optimized across L2/L3 order book feeds to minimize market impact and adverse selection.
- Financial Close & Revenue Waterfalls: Complex revenue recognition and equity waterfalls reconciled to the cent on double-entry ledgers.
- Statutory Tax & Benefits as Code: Tax code and payroll rule encoding verified to the cent under OpenFisca.

### Business Operations
- Combinatorial Scheduling (OR-Tools): Capacitated vehicle routing (CVRP) and job-shop scheduling under operational constraints.
- Multi-Echelon Inventory Replenishment: Dynamic inventory allocation across distribution networks under SimPy simulation.
- Supply Chain Carbon Accounting (LCA): Scope 1, 2, and 3 emission inventories verified on mass balance under Brightway2.
- E-Commerce Catalog & Order Management: Catalog synchronization and refund flows verified on simulated Shopify APIs.
- Enterprise Spreadsheets & Calculations: Multi-sheet formula dependency verification under headless LibreOffice.

### Graphics, Games & Digital Design
- Blender 3D Scene Engineering: Structured 3D scene construction with bpy; native scene state and exported artifacts determine success.
- Game & Rendering Fragment Shaders (GLSL): Shaders matching test scenes with a hard 16 ms frame budget under headless Mesa.
- Interactive 3D Graphics & Mechanics (Godot): Deterministic 3D scenes and rigid-body mechanics replayed across headless engines.
- Photoshop Raster Asset Automation: Image manipulation pipelines verified against SSIM perceptual metrics and channel diffs.

## Company & Access

- Website: https://www.g-ftech.com
- Organization: g factor technologies inc. (Delaware C-Corp)
- Contact: corporate@g-ftech.com
- Request a Platform Demo: https://www.g-ftech.com/#request-demo
- RSS / Atom Feed: /feed.xml
- Full LLM Knowledge Base: /llms-full.txt