Generative specialist · Research

Computer Vision Engineer

Detector that never loaded weights

What this agent actually does

CV models in production loop.

A detector that never loaded weights. A CLIP sticker on a page that never ran a frame.

The pain — unique to this desk

A detector that never loaded weights. A CLIP sticker on a page that never ran a frame.

The need this desk closes

Visual QA goldens, drift vs baseline, Sobel analog here. ViT / CLIP attachable — never invented on this host.

Why a generic chat cannot fake this

Detector that never loaded weights

Sobel analog here. ViT/CLIP attachable. Goldens and drift before canary. A CLIP sticker is not vision ops.

Takes

  • Uploaded image
  • Metrics + samples

Returns

  • Edge overlay
  • Drift flag
  • Edge vs serverless pick

Modalities this desk actually handles

ImageVideo (ABR ladder)Edge topology

How this specialist fuses

Detector that never loaded weights

Early visual encode: Sobel gradient on the pixels (ViT-class patches if a VLM is attached). Language is a second desk. Late fusion happens when LLM / RAG sits on the same ticket. No joint CLIP space on this host.

Vision-language on this desk

Named family. Never assumed.

Dual-encoder (CLIP/SigLIP, InfoNCE cosine) for retrieve/rank. Fusion-encoder: BLIP-2 Q-Former, LLaVA MLP projector, Flamingo gated xattn, or a GPT-4V-class hop. Unattached, this desk goldens visual QA and Sobel-checks the frame. Detector weights are not invented.

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 new detector is proposed. caps.golden regresses visual QA prompts. caps.drift flags score vs baseline. deploy.topology picks edge when the constraint is offline so the latency budget is visible before a hardware buy. The public analog is a Sobel overlay on an uploaded image. No claimed detector weights. Canary quality, not an infra canary that ignores the box.

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 Research. 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.

  1. Computer Vision Engineer analog closeHuman bar: Name the ticket. Attach evidence. Second family grades. Analog: #vision. Kernel #/vision 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 the new detector. Hold if mAP vs last week's goldens drops. Prefer on-device if the plant is offline.

Sit keys vision / image / Sobel / detector / bbox. Ticket kinds pr and deploy. MLOps owns the canary state machine. This desk owns labels, drift, and the edge call.

Playbook

How this agent works the ticket.

  1. 01 GoldensLabel the visual QA set. Regression before weights move.
  2. 02 DriftScore vs baseline. Alert is a number, not a vibe.
  3. 03 TopologyEdge vs serverless. Offline constraint named.
  4. 04 Sobel analogEdge overlay on an uploaded image. Honest about what loaded.
  5. 05 CanaryQuality floor. Same gate family as LLM paths.
  6. 06 Update goldensNew failure modes join the suite. They are not deleted because they lost.

Acts like the role

Work it does. Work it will not fake.

Does

  • Check drift on a labeled set
  • Canary weights behind a quality floor
  • Pick edge topology when the plant is offline
  • Run a Sobel overlay so the page is honest about weights

Does not

  • Claim YOLO / Detectron weights are loaded on this host
  • Paint a detector that never loaded weights
  • Skip goldens because the demo image looks sharp
  • Bypass eval on a 'just an overlay' change

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.

P0

Golden dataset regression

Today. Prompt/model change ships without suite

This agent. Versioned golden set + offline gate before merge

P1

Production drift detect

Today. Model gets quietly worse

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

P0

Canary with quality SLO

Today. Infra canary ignores answer quality

This agent. Shadow traffic score vs baseline before promote

P1

Corpus / feature freshness

Today. Stale RAG chunks and features

This agent. Ingest pipeline status + reindex job

P1

Cost-aware model routing

Today. Single frontier model burns budget

This agent. Token budget + cascade cheap→frontier

P0

False completion graders

Today. Agent claims done while tests fail

This agent. Objective graders before done

P1

Local + frontier dual path

Today. All-or-nothing cloud dependency

This agent. Model vault with local fallback route

P1

Incident runbook agent

Today. Pager chaos without structured triage

This agent. Incident desk with severity + rollback path

Skill.md

CV model canary

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(computer-vision-engineer).

Trigger. New detector weights and a quality floor exist.

  1. 01 StepConfirm goldens
  2. 02 StepScore drift vs baseline
  3. 03 StepChoose topology (edge if offline)
  4. 04 StepCanary with a quality floor
  5. 05 StepHold if goldens fail

Must

  • Drift check
  • Canary decision recorded

Must not

  • Claim weights that did not load
  • Ignore the offline constraint
# Computer Vision Engineer

id: computer-vision-engineer
layer: research (Research)
second family: ai-trainer-evaluator
tools: #/vision

## Pain
The hired Computer Vision 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
- vision-pass · Computer Vision Engineer analog close · analog #/vision · 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.

Sit keys

Talk Route matches these words.

vision

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

RBAC

Pass bar 78. C1.

  • Ceiling. C1 propose and gated read. Deploy and credential reveal denied.
  • Deny. credentials.reveal · deploy.canary
  • Scopes. workspace_read
  • Swarm seats. Not named on a ticket-kind composition. Talk Route can still sit it.

Daily missions

Run against live engines. Not slides.

deploy.topology

Simulate edge inference path

Quantized model on-device topology

Accept: Edge selected with offline constraint. Pain removed: Latency budget visible before hardware buy. Surfaces: Video Stream · Systems Lab.

Superpowers

  • ABR media lab
  • Edge deploy path
  • Classifier lab

Daily jobs

  • Check drift
  • Canary weights
  • Update goldens

Skills

  • Detection
  • Recognition
  • Vision models

Toolkit

Command Center becomes this desk.

Eval Gate

Label goldens

Regression on visual QA prompts

Outcome: Suite · caps.golden.

Command Center

Score drift

Metric vs baseline

Outcome: Alert · caps.drift.

Work templates

pr

CV model canary

Canary new detector with quality floor

Paste: Metrics + samples. Accept: Drift check · Canary.

Surfaces

Nav this desk actually opens.

Command CenterModels & KeysRole StudioEval GateKubernetesEvent AnalyticsIntelligence LabCertify & OperateUpdates

Local analog

Edge overlay

Sobel on an uploaded image. No claimed detector weights.

Edge overlay: Sobel on an uploaded image. No claimed detector weights. SAMPLE.

Run Edge overlay

Replacement cost

$150,000$215,000 cash

Year-1 loaded + recruiting $270,100. KORE1 mid. Senior cash $220k–$310k+.. Not EngOS payroll. Not ARR.

Pilot $0 / Team $79 sits this analog desk. Year-1 hire is $270,100 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 RouteAll twenty desks

AI Trainer and EvaluatorNLP Engineer

Generative neighbors

Generative specialist

Generative AI Engineer

Playground demo sold as a generation system

A playground demo sold as a generation system. The style guide lives in chat history.

Skill.md, goldens, and a cost route. A pretty sample is not a system.

TextImage (lexical latent analog)Video (ABR lab)
  • Version a prompt and keep the score with it
  • Run a golden suite before a generative UI ships
  • Package a repetitive style guide as Skill.md
Analog · Lexical latent explorer · bar 78 · C1

Generative specialist

Prompt Engineer

System prompt that lives in Slack

A system prompt that lives in Slack. Two variants, no goldens, a judge that is the author.

Registry, STE, A/B through the gate. Chat history is not a prompt product.

Text
  • Tighten a flaky system prompt against goldens
  • STE-clean a model reply for a production UI
  • Run a regression suite on multi-turn tasks
Analog · Tokenizer & priority · bar 78 · C1

Generative specialist

RAG Engineer

Hallucinated SKU in the FAQ — no chunk named

A RAG claim with no split, no overlap, no token count. Hallucinated product answers in the FAQ.

Named split, overlap, token count, 11-node graph. A vector demo is not retrieval.

DocumentsText
  • Fix a hallucinated product answer with a confidence gate
  • Add metadata filters for a tenant
  • Compare hybrid vs vectorless on hard queries
Analog · Chunk strategy playground · bar 78 · C1

Generative specialist

NLP Engineer

Silent truncation on a 32k chat

Silent truncation on a long chat. Extraction F1 that nobody can reproduce.

Context pack under a named token budget. The triad scores the extract.

Text
  • Score the triad on hard queries
  • Version extraction prompts against goldens
  • Compress long chat under a named token budget
Analog · Embedding distance · bar 78 · C1

EngOS