PENGYI INFRASTRUCTURE

Requirements become
reliable systems.

One infrastructure map connecting quant workloads, software contracts, runtime evidence, and justified hardware decisions.

Public capability profile
CPUIntel i7-13700KF16 cores / 24 threads
GPURTX 407012 GB VRAM
Memory32 GBSystem RAM
Local modelQwen3-8BQ4_K_M

INFRASTRUCTURE OPERATING MODEL

Research stands on maintained infrastructure.

GOVERNED CONTRACT
L0Hardware InfraPINF

Compute, storage, network, displays, power, thermals, and continuity.

L1SWE InfraCoding OS / System Design OS / PORC

Environment, configuration, tests, CI, deployment, observability, and recovery.

L2Data InfraPDAT

Freshness, schemas, lineage, point-in-time correctness, storage, and serving.

L3Research InfraPAT / PAAT / PCPT

Reproducible experiments, factors, backtests, portfolios, and release evidence.

L4Risk InfraPRT

Versioned policies, deterministic decisions, conditions, rejection, and halt state.

L5Execution InfraPET

Orders, fills, reconciliation, monitoring, recovery, and controlled broker state.

PINF governs cross-system contracts and hardware/software evidence. PDAT, PRT, and PET retain independent implementation and operational authority.

REQUIREMENT-DRIVEN INFRASTRUCTURE

Profile the workload before choosing the machine.

Every request moves from a declared problem and fixed workload through baseline, profiling, benchmark, decision, rollout, and regression review. Unknown results remain UNMEASURED.

TARGET WORKSPACE / NOT CURRENT INVENTORY

Two hosts. Five stable operating surfaces.

TARGET DESIGN
HOST AWindows compute planeBuild, inference, data, tests, deployment
HOST BMacBook continuity planeReview, research, communication, failover
  1. 01BuildIDE, terminal, testsHost A primary
  2. 02ApproveCommand and Human GateHost A primary
  3. 03ObserveRuntime and healthHost A primary
  4. 04ResearchPapers, docs, evidenceHost B primary
  5. 05CommunicatePMAIL, meetings, operationsHost B primary

Purchase remains gated on verified display ports, MacBook external-display limits, dock and cable requirements, desk ergonomics, thermals, and a tested failover path.

WHY PINF EXISTS

AI capability has a physical boundary.

PINF connects device capacity, model runtime, and workload behavior. It turns CPU, GPU, memory, storage, quantization, context, and concurrency into an explicit operating model for local AI.

COMPUTE BOUNDARY

Default local workload

Model class7B-9B Q4

Suitable for local tool use, structured extraction, document triage, RAG assistance, and low-risk Agent work.

Default context8K tokens

Context and concurrency are controlled because KV cache, desktop workloads, and runtime overhead share the same device.

AI INFRA CAREER LENS

Build evidence one infrastructure layer at a time.

01DeviceInventory, placement, capacity, policy
02ServingLatency, throughput, context, concurrency
03ReliabilityTelemetry, SLO, incidents, recovery
04ScaleContainers, scheduling, multi-node systems

CONTROL DOMAINS

One normalized device snapshot

01ObserveWindows CIM, NVIDIA, Ollama, Docker
02NormalizeTyped hardware and runtime snapshot
03EvaluateCapacity model and policy thresholds
04PresentDoctor output and local dashboard

LOCAL MODEL POLICY

System disk is not model storage.

  • Runtime, model weights, and temporary model files use a dedicated data volume.
  • Local models are unloaded from VRAM when a task finishes.
  • Cloud escalation remains available for difficult reasoning.
  • Deletion, migration, and runtime shutdown require human approval.