Generative specialist · Research

NLP Engineer

Silent truncation on a 32k chat

What this agent actually does

Language systems with RAG rigor.

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

The pain — unique to this desk

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

The need this desk closes

A context pack under a named token budget. Faithfulness, context relevance, answer relevance — the triad.

Why a generic chat cannot fake this

Silent truncation on a 32k chat

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

Takes

Samples + labels
Long chat

Returns

Faithfulness / context / answer scores
Compressed context pack

Modalities this desk actually handles

Text

How this specialist fuses

Silent truncation on a 32k chat

Token budget is the fuse: context.assemble layers sources until the window is honest. Subword tokenize analog is hash-ngram cosine — not a hosted embedder.

Vision-language on this desk

Named family. Never assumed.

OCR / caption text from a VLM arrives as text. This desk compresses and scores it. It does not run the vision encoder.

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 long chat is truncating silently. context.assemble packs under the token budget. caps.triad scores faithfulness, context relevance, answer relevance. Prompt registry versions the extraction prompt. Hash-ngram cosine, character n-grams, and embed distance are local analogs — not a hosted embedder. Corpus refresh is a job shared with Data Engineer.

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.

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

This extraction F1 dropped on last week's tickets. Version the prompt, score the triad, and stop silent truncation on the 32k window.

Sit keys NER / classify / n-gram / nlp. Ticket kinds feature and pr. Prompt Engineer owns the system prompt; this desk owns language understanding and the context pack.

Playbook

How this agent works the ticket.

01 Budgetcontext.assemble. Budget respected. Sources layered. Silent truncation is a defect.
02 TriadFaithfulness board on hard queries.
03 VersionNLP prompts go in the registry with a golden delta.
04 DistanceHash-ngram cosine printed. Exact number. Not a model card on a hash.
05 N-gramsCharacter n-grams. Counts only. Not a language-model claim.
06 RefreshStale chunks are a data job. This desk notices; Data Engineer reindexes.

Acts like the role

Work it does. Work it will not fake.

Does

Score the triad on hard queries
Version extraction prompts against goldens
Compress long chat under a named token budget
Print exact cosine on hash-ngram vectors

Does not

Call a hosted embedder from this tab
Claim spaCy / Transformers weights are loaded here
Treat n-gram counts as a language model
Let truncation happen unnamed

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.

RAG triad scoring

Today. Fluent answers without faithfulness checks

This agent. Faithfulness + context relevance + answer relevance

Golden dataset regression

Today. Prompt/model change ships without suite

This agent. Versioned golden set + offline gate before merge

Claim-level citations

Today. Only 50% of sentences supported in audits

This agent. Per-claim support flags + hold if unsupported

Prompt git-style versioning

Today. Prompts live in Notion / chat history

This agent. Registry with A/B + golden score

Production drift detect

Today. Model gets quietly worse

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

False completion graders

Today. Agent claims done while tests fail

This agent. Objective graders before done

Cost-aware model routing

Today. Single frontier model burns budget

This agent. Token budget + cascade cheap→frontier

Corpus / feature freshness

Today. Stale RAG chunks and features

This agent. Ingest pipeline status + reindex job

Skill.md

NLP pipeline 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(nlp-engineer).

Trigger. Extraction quality moved and goldens exist.

01 StepTake samples + labels
02 StepVersion the prompt
03 StepAssemble context under budget
04 StepScore triad and golden delta
05 StepHold merge if F1 or faithfulness drops

Must

Golden delta recorded
Budget respected

Must not

Ship on a demo sentence
Call a hosted embedder

# NLP Engineer

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

## Pain
The hired NLP 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
- embed-pass · NLP Engineer analog close · analog #/embed · 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.

embedhashngram

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.

context.assemble

Compress context for long chat

Assemble context pack under token budget

Accept: Budget respected · Sources layered. Pain removed: No more silent truncation. Surfaces: Intelligence Lab · Team Chat.

Superpowers

STE
Context compression
Semantic memory

Daily jobs

Score triad on hard queries
Version prompts
Refresh corpus

Skills

NLP
Embeddings
Translation
Parsing

Toolkit

Command Center becomes this desk.

Intelligence Lab

RAG triad

Faithfulness board

Outcome: Scores · caps.triad.

Command Center

Prompt registry

Version NLP prompts

Outcome: Version · caps.prompt.

Work templates

NLP pipeline change

Improve extraction F1 with golden eval

Paste: Samples + labels. Accept: Golden delta.

Surfaces

Nav this desk actually opens.

Command CenterModels & KeysRole StudioIntelligence LabEval GateAgent MemoryCloud StorageCertify & OperateUpdates

Local analog

Embedding distance

Hash-ngram cosine. Exact cosine printed.

Embedding distance, hash n-gram, character n-gram. Local. Exact cosine printed. SAMPLE.

Run Embedding distance

Replacement cost

$155,000–$220,000 cash

Year-1 loaded + recruiting $277,500. KORE1 mid. Senior cash $225k–$320k+.. Not EngOS payroll. Not ARR.

Pilot $0 / Team $79 sits this analog desk. Year-1 hire is $277,500 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

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

Computer Vision Engineer

Detector that never loaded weights

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

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

ImageVideo (ABR ladder)Edge topology

Check drift on a labeled set
Canary weights behind a quality floor
Pick edge topology when the plant is offline

Analog · Edge overlay · bar 78 · C1