PERSONAL MASTER PRESENTATION / V0.4
Infrastructure Enthusiast.
Pengyi Peng / learning to connect software behavior, runtime systems, and physical capacity.
01 / DEFINITION
Infrastructure is the system that makes every higher-level promise possible.
PINF owns cross-system architecture, requirements, interfaces, benchmarks, resource constraints, and hardware/software decisions.
02 / TWO FIRST-CLASS SURFACES
Software and hardware infrastructure must be reasoned about together.
Software determines workload shape; hardware determines the physical envelope; measurement connects them.
- Hardware/runtime / compute, memory, storage, network, OS, power, thermals
- Software / environments, CI, packaging, release, observability, recovery
- Interface / workload, profile, benchmark, decision, rollback
03 / COMPLETE DEPENDENCY MAP
Research and trading stand on six governed infrastructure layers.
Dependencies point downward; evidence and feedback move in both directions.
- L0 Hardware and runtime / PINF
- L1 Software engineering platform / PINF contracts
- L2 Data / PDAT
- L3 Research workload / PAT, PAAT, PCPT
- L4 Risk / PRT
- L5 Execution / PET
04 / SOFTWARE INFRASTRUCTURE
Reliable software begins before application code and continues after deployment.
Environment, contracts, build, test, release, telemetry, rollback, and recovery form one operating path.
- Reproducible environments and dependency control
- Typed interfaces and independent ownership
- CI, deployment, rollback, and recovery
- Observability, SLOs, incidents, and capacity
05 / HARDWARE INFRASTRUCTURE
AI capability has a physical boundary.
CPU, accelerators, VRAM, RAM, storage, bandwidth, network, power, thermals, and continuity constrain what software can promise.
- Capacity and placement
- Latency and throughput
- Contention and admission
- Failure, continuity, and safe teardown
06 / DECISION LOOP
Hardware purchase is an output of profiling, not a starting preference.
Workload -> Baseline -> Profile -> Bottleneck -> Candidate -> Benchmark -> Decision -> Rollout -> Recovery check
- Fix the representative workload
- Measure software and hardware state
- Change one constrained dimension
- Record cost, trade-off, rollback, and regression
07 / AI INFRASTRUCTURE
A model API is the visible edge of a much deeper machine.
Accelerator -> memory and interconnect -> runtime -> serving -> scheduling -> reliability -> Agent workload
- Precision, quantization, kernels, and memory movement
- KV cache, batching, concurrency, and admission
- Tail latency, throughput, health, and recovery
- Local evidence informing cloud-model judgment
08 / QUANT AND TRADING INFRA
Correct research is still fragile without data, risk, and execution boundaries.
PINF governs compatibility and evidence between independent systems without centralizing their implementation.
- PDAT / freshness, schema, lineage, point-in-time correctness
- Research / reproducibility, evaluation, portfolio computation
- PRT / deterministic risk decisions and halt state
- PET / order lifecycle, reconciliation, monitoring, recovery
09 / RELIABILITY
The real test of infrastructure begins when a dependency fails.
Healthy systems make failure visible, contain blast radius, recover predictably, and preserve an audit trail.
- What is the SLO and who owns it?
- What telemetry detects degradation?
- What is the rollback or failover path?
- What evidence proves recovery?
10 / PINF CONTROL LOOP
Observe first. Normalize state. Evaluate policy. Present bounded evidence.
Collectors -> Typed snapshot -> Doctor and policy checks -> Local dashboard -> Human decision
- Read-only observation by default
- Local telemetry stays local
- Destructive operations require approval
- Public surfaces expose capability, not machine state
11 / CURRENT EVIDENCE
PINF has a real foundation, but the strongest runtime claims are still ahead.
Current proof covers device observation, policy checks, typed registries, contract validation, public-safe presentation, and timestamped baselines.
- Hardware and local-runtime collectors
- Doctor, topology, infra, and software-infra validators
- Six-layer ownership and promotion model
- Public-safe capability profile and private telemetry boundary
12 / EVIDENCE GAP
Documentation is not production infrastructure evidence.
PINF does not yet prove production inference, distributed training, cluster operations, or data-center engineering.
- Measure latency, throughput, context, concurrency, and memory
- Add telemetry, alerts, restart policy, and incident recovery
- Containerize with resource limits and safe teardown
- Study multi-node systems only after single-node evidence is defensible
13 / SUPER INFRA ROADMAP
Become an infrastructure engineer one measured layer at a time.
The next move is not more architecture prose; it is one fixed workload that forces software and hardware evidence to meet.
- Stage 1 / Single-device inventory and capacity
- Stage 2 / Controlled serving benchmark
- Stage 3 / Reliability and recovery evidence
- Stage 4 / Containers, scheduling, and admission
- Stage 5 / Distributed compute, storage, and network systems
14 / NEXT CONVERSATION
Give me the workload. I want to find the boundary.
Software infrastructure / hardware systems / AI serving / quant and trading platforms
- pengyi-pinf.pages.dev
- github.com/pengpengyi92
- pengpengyi92@gmail.com