Operations specialist · Scale

AI Infrastructure Engineer

Finance finds the runaway agent first

What this agent actually does

Inference & platform plane.

Finance finding the runaway agent first. A cluster claim with no GPU, batch, or seq.

The pain — unique to this desk

Finance finding the runaway agent first. A cluster claim with no GPU, batch, or seq.

The need this desk closes

Token bucket → breaker → fallback. Cost ledger vs budget. Latency budget pass/fail. Cluster analog you set.

Why a generic chat cannot fake this

Finance finds the runaway agent first

Token bucket → breaker → fallback. Cluster analog you set. Cost ledger vs budget.

Takes

SLO + current stack

Returns

Route plan
Cost delta
Latency budget pass/fail

Modalities this desk actually handles

ServingCostAttachable cluster

How this specialist fuses

Finance finds the runaway agent first

Route cheap-then-reason. A VLM is a more expensive route, not a default. Breaker still fires.

Vision-language on this desk

Named family. Never assumed.

Vision-language serving is attachable (Bedrock / Azure / local). Unattached, this desk does not invent a GPU cluster.

Patch / tokenize → project → fuse → ground is the host mechanism. This desk fuses late under its own playbook — not a slogan copied across the roster. Host mechanism.

Burst traffic hits the plane. enforce.chain is primary → cheaper model → cache → 503. Models vault connects local and frontier (Ollama / vLLM / API — connected, not invented as capacity). caps.cost is spend vs budget. arch.latency is the p95 path. Cluster bubble is a labeled simulation — you set GPU count, batch, seq. Swarm sizing curve is a superpower. Exactly-once delivery sits with Data Engineer on retries.

Live missions run in Role Studio. Analog kernels stay in this tab until a hosted call is chosen. Authors never grade themselves. A hashed pack is not a letter. Hire bands are market replacement cost, not EngOS payroll. Isolation is probed, not assumed.

Reports up as Scale. Second family ai-trainer-evaluator. Authors never grade themselves. Isolation is probed, not assumed.

Five analog ticket kinds

How a hired human spends the week. What the analog kernel closes.

Percent splits stay human-week language. Never GPU-loop quotas. Each kind binds to the analog tool that closes it. Escalate only when the kernel cannot.

AI Infrastructure Engineer analog closeHuman bar: Name the ticket. Attach evidence. Second family grades. Analog: #cluster. Kernel #/cluster on this host. Arithmetic / Canvas 2D. Device none. Escalate when: The analog kernel cannot close, or a hosted connector is required as if it were present.

A typical ticket

What you say. What sits.

Cut p95 20% with routing + cache. If the frontier is down, fall back local. Halt if we burn the daily cap. Simulate 8×A10, batch 16, seq 2k.

Sit keys GPU / cluster / infra / Kubernetes. Ticket kinds deploy, incident, billing. MLOps owns canary. This desk owns the serving plane and the cost chain.

Playbook

How this agent works the ticket.

01 VaultLocal + frontier dual path. All-or-nothing cloud is the pain.
02 ChainToken bucket → breaker → fallback. Primary → cheaper → cache → 503.
03 CostSpend vs budget. USD estimate from numbers you supply.
04 LatencyFull-stack p95 flame path. Budget pass/fail.
05 Cluster analogYou set GPU count, batch, seq. Labeled simulated. Not a rented cluster.
06 ConsolidateDCE · microkernel · FFI · zero-copy inventory before the next ship.

Acts like the role

Work it does. Work it will not fake.

Does

Tune model routes and watch the cost budget
Contain a runaway agent with the enforce chain
Canary an inference build
Run latency fix-all on p95 misses
Simulate a cluster from GPU / batch / seq you type

Does not

Rent GPUs from this page
Invent CUDA capacity
Let finance discover the runaway first
Claim Kafka / ClickHouse as the shipped bus

Jobs on the board

Scenarios this desk was built to close.

From Role Studio's scenario bank — current pain without Helix, and the job the agent actually runs. Top eight of twenty.

Local + frontier dual path

Today. All-or-nothing cloud dependency

This agent. Model vault with local fallback route

Cost-aware model routing

Today. Single frontier model burns budget

This agent. Token budget + cascade cheap→frontier

Canary with quality SLO

Today. Infra canary ignores answer quality

This agent. Shadow traffic score vs baseline before promote

Incident runbook agent

Today. Pager chaos without structured triage

This agent. Incident desk with severity + rollback path

Typed MCP tool boundary

Today. N×M custom tool integrations

This agent. MCP registry + deny-list + audit per tool call

Agent loop with durable checkpoints

Today. Agents lose state mid-run; restart from scratch

This agent. Checkpointed harness with resume + budget per cycle

Production drift detect

Today. Model gets quietly worse

This agent. Drift monitor on labels/scores + retrain trigger

Secret / PII exfiltration

Today. Models echo keys from context

This agent. Secret scan on context + artifact write

Skill.md

Serving change

Durable analog execution standard. Trigger, five ticket kinds, analog tools, must / must-not, handoff (evidence not authority), spend.halt, second family. Same factory as PR Creator, Code Reviewer, Handoff Verifier — plus fromDesk(ai-infra-engineer).

Trigger. p95 or cost is off and an SLO exists.

01 StepRead SLO + current stack
02 StepPick route + cache policy
03 StepEstimate cost delta
04 StepArm breaker and spend.halt
05 StepCanary the inference build

Must

Route plan
Cost delta
Blast radius bounded

Must not

Leave the primary as the only path
Skip the 503 fallback

# AI Infrastructure Engineer

id: ai-infra-engineer
layer: scale (Scale)
second family: ai-trainer-evaluator
tools: #/cluster

## Pain
The hired AI Infrastructure Engineer seat does not exist yet, or the work has no named ticket.

## Need
A named ticket, an analog close, and a second family. Not a chat that grades itself.

## Isolation
Isolation is probed, not assumed. Unprobed stays unlabeled. Never certified-green from a desk.

## spend.halt
spend.halt on the ticket cap. Only the operator raises the ceiling.

## Ticket kinds
- cluster-pass · AI Infrastructure Engineer analog close · analog #/cluster · escalate when: The analog kernel cannot close, or a hosted connector is required as if it were present.

Who sits with this desk

Swarm compose by ticket kind.

Not named on a ticket-kind composition. Talk Route can still sit it from sit keys.

Engineering Agents 'Crucible Twin' analog: sandbox + fallback chain. We do not claim SEV-SNP or a live GPU fleet on this host.

Sit keys

Talk Route matches these words.

cluster

Keyword analog in this tab. Not a hosted model. Ticket id is the idempotency key.

RBAC

Pass bar 82. C2.

Ceiling. C2 gated act. Eval bar required. Secrets stay denied unless the profile lists them.
Deny. credentials.reveal
Scopes. deploy · queue · workspace_read · audit_read
Swarm seats. Not named on a ticket-kind composition. Talk Route can still sit it.

Daily missions

Run against live engines. Not slides.

enforce.chain · delivery.exactly_once

Contain a runaway agent

Burst traffic through token bucket → breaker → fallback models

Accept: Blast radius bounded · Primary→haiku→cache→503. Pain removed: Finance doesn't find out first. Surfaces: Systems Lab.

Superpowers

Token→breaker→fallback chain
Sandbox primitives
Swarm sizing curve

Daily jobs

Tune model routes
Watch cost budget
Canary inference build
Run prompts.consolidate + fix-all on latency misses

Skills

GPU
Distributed systems
Networking
Cost

Toolkit

Command Center becomes this desk.

Models & Keys

Model vault

Connect Ollama / vLLM / frontier

Outcome: Default route.

Command Center

Cost ledger

Spend vs budget

Outcome: USD estimate · caps.cost.

Systems Lab

Deploy topology

Serverless vs edge vs on-prem

Outcome: Recommendation.

Kernel Guard

Latency budgets

Full-stack p95 flame path

Outcome: Budget pass/fail · arch.latency.

Kernel Guard

Arch consolidate

DCE · microkernel · FFI · zero-copy

Outcome: Consolidation score 100 target · prompts.consolidate.

Work templates

Serving change

Cut p95 latency 20% with routing + cache policy

Paste: SLO + current stack. Accept: Route plan · Cost delta.

Surfaces

Nav this desk actually opens.

Command CenterModels & KeysRole StudioKubernetesSystems LabEvent AnalyticsKernel GuardCertify & OperateUpdates

Local analog

Cluster bubble sandbox

You set GPU count, batch, seq. Labeled simulated.

Cluster bubble sandbox. You set GPU count, batch, seq. Labeled simulated. SAMPLE.

Run Cluster bubble sandbox

Replacement cost

$155,000–$230,000 cash

Year-1 loaded + recruiting $284,900. Cluster desks. GPU $8k–$40k extra.. Not EngOS payroll. Not ARR.

Pilot $0 / Team $79 sits this analog desk. Year-1 hire is $284,900 loaded. That is not a replacement claim. The hired role remains the real thing. The analog desk reports up so one operator can run the ticket.

Open Talk Route All twenty desks

Operations neighbors

Operations specialist