Operations specialist · Scale

MLOps Engineer

Silent regression that reached production

What this agent actually does

Ship models safely.

A silent regression that reached production. Infra canary that ignores answer quality.

The pain — unique to this desk

A silent regression that reached production. Infra canary that ignores answer quality.

The need this desk closes

Canary only if eval merges. Topology named. spend.halt armed. Probe trail on the release.

Why a generic chat cannot fake this

Silent regression that reached production

Canary only if eval merges. spend.halt armed. Quality SLO, not only infra green.

Takes

Version, SLOs, rollback triggers

Returns

Release history
Topology recommendation
Prod-ready score

Modalities this desk actually handles

Release stateProbesVRAM plan

How this specialist fuses

Silent regression that reached production

Quality SLO + infra probe. A vision model canary still needs eval merge — pixels do not skip the gate.

Vision-language on this desk

Named family. Never assumed.

VRAM for a VLM is weights + vision encoder + KV. The allocator is P × bytes and GQA-labeled. A vendor screenshot is not a capacity plan.

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.

A model or prompt bundle wants prod. eval.run_gate must merge before k8s.deploy. deploy.advise picks serverless / on-prem / edge. Circuit breakers and spend.halt bound blast radius — finance does not find out first. VRAM allocator is weights = P × bytes, KV = 2·L·n_kv·d_h·S·B·bytes, GQA labeled. Prod-readiness fix-all is CI/CD · workers · Docker · latency. Isolation is probed, not assumed.

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.

MLOps Engineer analog closeHuman bar: Name the ticket. Attach evidence. Second family grades. Analog: #vram. Kernel #/vram 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.

Canary prompt bundle v12. Rollback if groundedness drops 3 points. VRAM plan for 7B Q4 on 24GB. spend.halt if token burn exceeds the daily cap.

Sit keys VRAM / deploy / MLOps / canary. Ticket kinds deploy, incident, chore. Security final-checks. Trainer / Evaluator is the merge family. This desk owns the state machine.

Playbook

How this agent works the ticket.

01 Eval firstBad gate blocks ship. No silent regressions to prod.
02 VRAM planWeights and KV named. GQA labeled. Vendor screenshot is not a plan.
03 TopologyServerless vs on-prem vs edge with cold-start and residency named.
04 CanaryReal state machine with probes. Quality SLO, not only infra health.
05 spend.haltBreaker + halt. Blast radius bounded. Not a 'kill' slogan.
06 Fix-allProd-readiness chain before the next ship.

Acts like the role

Work it does. Work it will not fake.

Does

Canary a new prompt/model bundle with rollback criteria
Investigate probe failures
Audit rate limits and breakers
Run production-readiness fix-all before ship
Print a VRAM plan from P, L, n_kv, d_h, S, B, bytes

Does not

Promote on infra green if quality SLO is red
Treat a vendor screenshot as capacity
Leave spend uncapped
Invent Kubeflow as live on this host

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.

Canary with quality SLO

Today. Infra canary ignores answer quality

This agent. Shadow traffic score vs baseline before promote

Production drift detect

Today. Model gets quietly worse

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

Cost-aware model routing

Today. Single frontier model burns budget

This agent. Token budget + cascade cheap→frontier

Incident runbook agent

Today. Pager chaos without structured triage

This agent. Incident desk with severity + rollback path

Golden dataset regression

Today. Prompt/model change ships without suite

This agent. Versioned golden set + offline gate before merge

Local + frontier dual path

Today. All-or-nothing cloud dependency

This agent. Model vault with local fallback route

False completion graders

Today. Agent claims done while tests fail

This agent. Objective graders before done

Corpus / feature freshness

Today. Stale RAG chunks and features

This agent. Ingest pipeline status + reindex job

Skill.md

Model rollout

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(mlops-engineer).

Trigger. A new model or prompt bundle has SLOs and rollback triggers.

01 StepRun eval gate
02 StepName VRAM and topology
03 StepCanary with probes
04 StepWatch quality vs baseline
05 StepPromote, hold, or roll back
06 StepLog events. Arm spend.halt

Must

Canary healthy or rolled back
Events logged
Eval merge

Must not

Skip eval
Ignore quality SLO

# MLOps Engineer

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

## Pain
The hired MLOps 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
- vram-pass · MLOps Engineer analog close · analog #/vram · 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.

incident · bar 88
AI Security Specialist · Generative AI Engineer
Triage, root-cause, patch, certify, hold canary until independent merge
chore · bar 82
Generative AI Engineer · Data Engineer for AI
Low-risk maintenance under standard gates
billing · bar 82
Forward Deployed Engineer · AI Product Manager
Cost caps, plan, usage ledger — admin only paths
deploy · bar 88
AI Security Specialist
Canary only after moat + hardening green

Engineering Agents 'Crucible Twin' analog on the infra side: DST-style debug + spend.halt. We do not claim SEV-SNP.

Sit keys

Talk Route matches these words.

vram

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. incident · chore · billing · deploy

Daily missions

Run against live engines. Not slides.

eval.run_gate · k8s.deploy

Canary only if eval merges

Promote release when orthogonal + eval pass

Accept: Bad gate blocks ship. Pain removed: No more silent regressions to prod. Surfaces: Eval Gate · Kubernetes.

Superpowers

Canary + spend.halt
Eval merge gate
Circuit breakers

Daily jobs

Canary a new prompt/model bundle
Investigate probe failures
Audit rate limits and breakers
Run production-readiness fix-all before ship

Skills

CI/CD
Monitoring
Automation
Rollbacks

Toolkit

Command Center becomes this desk.

Systems Lab

Deploy advisor

Serverless vs on-prem vs edge for your constraints

Outcome: Topology recommendation · deploy.advise.

Kubernetes

Canary + rollback

Real state machine with probes

Outcome: Release history.

Kernel Guard

Prod readiness fix-all

CI/CD · workers · Docker · latency chains

Outcome: Prod-ready gate score · findings.fix_all.

Kernel Guard

Arch consolidate

DCE · microkernel · FFI · latency inventory

Outcome: Consolidation report · prompts.consolidate.

Work templates

incident

Model rollout

Canary new model version with rollback criteria

Paste: Version, SLOs, rollback triggers. Accept: Canary healthy or rolled back · Events logged.

Surfaces

Nav this desk actually opens.

Command CenterModels & KeysRole StudioKubernetesEvent AnalyticsKernel GuardSystems LabEval GateCertify & OperateUpdates

Local analog

VRAM allocator

Weights = P × bytes. KV = 2·L·n_kv·d_h·S·B·bytes. GQA labeled.

VRAM allocator. Weights = P × bytes. KV = 2·L·n_kv·d_h·S·B·bytes. GQA labeled. SAMPLE.

Run VRAM allocator

Replacement cost

$145,000–$200,000 cash

Year-1 loaded + recruiting $255,300. KORE1 mid. GPU class extra not included.. Not EngOS payroll. Not ARR.

Pilot $0 / Team $79 sits this analog desk. Year-1 hire is $255,300 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

AI Infrastructure Engineer

Finance finds the runaway agent first

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

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

ServingCostAttachable cluster

Tune model routes and watch the cost budget
Contain a runaway agent with the enforce chain
Canary an inference build

Analog · Cluster bubble sandbox · bar 82 · C2

Operations specialist

Data Engineer for AI

Retry storm that double-charges

A retry storm that double-charges. Near-dup lines in the corpus nobody counted.

Exactly-once (uniqueEffects == 1). SHA-256 ingest. Near-dups counted before anyone trains.

DocumentsTablesImages as files

Refresh a corpus and verify the RAG triad
Prove exactly-once on a retry storm
Count near-dups before anyone trains

Analog · Corpus cleaner · bar 78 · C1

MLOps Engineer

Ship models safely.

A silent regression that reached production. Infra canary that ignores answer quality.

Canary only if eval merges. Topology named. spend.halt armed. Probe trail on the release.

Silent regression that reached production

Silent regression that reached production

Named family. Never assumed.

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

What you say. What sits.

How this agent works the ticket.

Work it does. Work it will not fake.

Scenarios this desk was built to close.

Canary with quality SLO

Production drift detect

Cost-aware model routing

Incident runbook agent

Golden dataset regression

Local + frontier dual path

False completion graders

Corpus / feature freshness

Model rollout

Swarm compose by ticket kind.

AI Security Specialist · Generative AI Engineer

Generative AI Engineer · Data Engineer for AI

Forward Deployed Engineer · AI Product Manager

AI Security Specialist

Talk Route matches these words.

Pass bar 82. C2.

Run against live engines. Not slides.

Canary only if eval merges

Command Center becomes this desk.

Deploy advisor

Canary + rollback

Prod readiness fix-all

Arch consolidate

Model rollout

Nav this desk actually opens.

VRAM allocator

$145,000–$200,000 cash

AI Infrastructure Engineer

Data Engineer for AI

EngOS