PERSONAL MASTER PRESENTATION / V0.5
Infrastructure becomes real when software meets the machine.
Pengyi Peng / PINF / software, hardware, AI serving, quant systems, and reliability.
01 / THE PROMISE
Infrastructure makes every higher-level promise possible.
Capacity, reproducibility, latency, reliability, and safe change are properties of the whole system.
02 / TWO SURFACES
Software and hardware are one decision surface.
Software shapes the workload. Hardware sets the physical envelope. Measurement keeps both honest.
- Runtime / compute, memory, storage, network, power
- Platform / environment, CI, release, telemetry, recovery
- Interface / profile, benchmark, decision, rollback
03 / DEPENDENCY MAP
Six governed layers carry research into execution.
Dependencies point downward. Evidence, telemetry, and feedback move both ways.
- L0-L1 / hardware, runtime, software platform
- L2-L3 / data and research workloads
- L4-L5 / risk and execution
04 / SOFTWARE INFRA
Reliable software begins before application code and continues after release.
Environment, contract, build, test, deployment, telemetry, rollback, and recovery form one operating path.
- Reproducible environments
- Typed ownership boundaries
- Observable release and recovery
05 / HARDWARE INFRA
AI capability has a physical boundary.
Accelerators, memory, storage, bandwidth, power, thermals, and continuity constrain every software promise.
- Placement and capacity
- Latency and throughput
- Contention and failure
06 / DECISION LOOP
Profile the workload before choosing the machine.
Workload -> baseline -> profile -> bottleneck -> candidate -> benchmark -> rollout -> recovery.
- Fix one representative workload
- Change one constrained dimension
- Record cost, trade-off, rollback, regression
07 / AI SERVING
A model API is the visible edge of a deeper machine.
Accelerator -> memory -> interconnect -> runtime -> scheduler -> serving -> Agent workload.
- Kernels, precision, memory movement
- KV cache, batching, admission
- Tail latency, health, recovery
08 / QUANT INFRA
Research needs data, risk, and execution boundaries to survive contact with markets.
PINF governs compatibility and evidence without absorbing the systems it connects.
- PDAT / freshness, schema, lineage
- PRT / deterministic risk and halt state
- PET / lifecycle, reconciliation, recovery
09 / RELIABILITY
Infrastructure is tested when a dependency fails.
Failure must be visible, bounded, recoverable, and auditable.
- What is the SLO and owner?
- What detects degradation?
- What proves rollback or failover?
10 / CONTROL LOOP
Observe first. Normalize state. Evaluate policy. Escalate with evidence.
Collectors -> typed snapshot -> doctor -> policy checks -> dashboard -> Human decision.
- Read-only observation by default
- Private telemetry stays local
- Destructive change requires approval
11 / CURRENT EVIDENCE
The foundation is real. The strongest runtime claims are still ahead.
Current proof covers observation, policy checks, typed registries, contract validation, public-safe presentation, and baselines.
- Hardware and runtime collectors
- Topology and infra validators
- Promotion model and privacy boundary
12 / EVIDENCE GAP
Architecture documents are not production infrastructure evidence.
PINF does not yet prove production inference, distributed training, cluster operations, or data-center engineering.
- Measure latency, throughput, concurrency, memory
- Add alerts, restart policy, incident recovery
- Prove single-node behavior before multi-node claims
13 / ROADMAP
Build infrastructure depth one measured layer at a time.
One fixed workload should force software, hardware, network, and recovery evidence to meet.
- 01 / device capacity
- 02 / serving benchmark
- 03 / reliability and recovery
- 04 / scheduling and admission
- 05 / distributed systems
14 / NEXT CONVERSATION
Give me the workload. I want to find the boundary.
Software infrastructure / hardware systems / AI serving / quant platforms.
- pengyi-pinf.pages.dev
- github.com/pengpengyi92
- pengpengyi92@gmail.com