ML / Research Engineer, Safeguards
When to Use
- Define research questions on harm detection, jailbreak resistance, or policy categories
- Curate or audit safety datasets — labeling guidelines, bias checks, version control
- Train or fine-tune classifiers, rankers, or small LLM judges for moderation
- Design benchmarks and eval suites — golden sets, adversarial slices, regression harnesses
- Run ablations — architecture, threshold, data mix, ensemble vs single model
- Analyze metrics — precision/recall, calibration, false positive/negative slices
- Write research memos — methods, results, limitations, production recommendation
- Specify promotion bar for a new safeguard model version
When NOT to Use
- Deploy gateways, GPU serving, canary routing →
ml-infrastructure-engineer-safeguards
- Execute structured red-team engagements on prod →
ai-redteam
- Draft acceptable-use policy or risk tiers →
ai-risk-governance
- Build customer-facing RAG/agents →
ai-engineer
- General literature survey unrelated to safety →
ai-researcher
- Token/context compression research →
research-engineer-scientist-tokens
- Product A/B and business metrics →
data-scientist
- PII detection benchmarks, memorization, logging minimization →
privacy-research-engineer-safeguards
Related skills
| Need |
Skill |
| Privacy research for safeguards |
privacy-research-engineer-safeguards |
| Production safeguard path and rollout |
ml-infrastructure-engineer-safeguards |
| Adversarial attack campaigns |
ai-redteam |
| Governance sign-off and model cards |
ai-risk-governance |
| Production eval harness in app |
ai-engineer |
| General research methodology |
ai-researcher |
| Classical ML and statistics |
data-scientist |
| Token efficiency ablations |
research-engineer-scientist-tokens |
| Release gates and ops cadence |
ai-lead-ops |
Core Workflows
1. Research framing (safety)
Hypotheses, harm taxonomy, success metrics.
See references/research_framing_safety.md.
2. Benchmarks and datasets
Golden sets, labeling, versioning.
See references/safety_benchmarks_datasets.md.
3. Model development
Training, fine-tuning, ensembles.
See references/classifier_model_development.md.
4. Evaluation and metrics
Slices, calibration, error analysis.
See references/evaluation_metrics_analysis.md.
5. Ablations and experiments
Controls, reproducibility.
See references/ablation_experiment_design.md.
6. Handoff to production
Promotion criteria, monitoring hooks.
See references/research_to_production_handoff.md.
Outputs
- Research brief — question, baseline, hypothesis, metrics
- Dataset card — sources, label schema, known limitations
- Benchmark spec — cases, categories, pass/fail rubric
- Results table — metrics by slice with confidence intervals where possible
- Error analysis — representative FP/FN clusters
- Promotion recommendation — go/no-go vs current production classifier
Principles
- Measure what policy cares about — category-level recall on high-severity harms
- Report failures honestly — FPs hurt UX; FNs hurt safety
- Hold out adversarial refresh — do not train on the only test set
- Reproducible — seeds, data version, model hash, eval script
- Separate research from ops — research proves lift; infra ships it
1---2name: ml-research-engineer-safeguards3description: Guides ML/research engineering for safeguards—safety classifier development, harm benchmarks and eval suites, labeled dataset design, fine-tuning and ablations, calibration and slice analysis, attack-surface research memos, and promotion criteria for new moderation models. Use when building or evaluating guardrail models, designing safety benchmarks, measuring precision/recall on policy categories, comparing mitigation techniques, or writing research reports on classifier improvements—not for production inference gateways (ml-infrastructure-engineer-safeguards), PII/leakage privacy research (privacy-research-engineer-safeguards), red-team attack campaigns (ai-redteam), AI governance policy (ai-risk-governance), general non-safety research (ai-researcher), or token-efficiency studies (research-engineer-scientist-tokens).4---56# ML / Research Engineer, Safeguards78## When to Use910- Define **research questions** on harm detection, jailbreak resistance, or policy categories11- Curate or audit **safety datasets** — labeling guidelines, bias checks, version control12- **Train or fine-tune** classifiers, rankers, or small LLM judges for moderation13- Design **benchmarks and eval suites** — golden sets, adversarial slices, regression harnesses14- Run **ablations** — architecture, threshold, data mix, ensemble vs single model15- Analyze **metrics** — precision/recall, calibration, false positive/negative slices16- Write **research memos** — methods, results, limitations, production recommendation17- Specify **promotion bar** for a new safeguard model version1819## When NOT to Use2021- Deploy gateways, GPU serving, canary routing → `ml-infrastructure-engineer-safeguards`22- Execute structured red-team engagements on prod → `ai-redteam`23- Draft acceptable-use policy or risk tiers → `ai-risk-governance`24- Build customer-facing RAG/agents → `ai-engineer`25- General literature survey unrelated to safety → `ai-researcher`26- Token/context compression research → `research-engineer-scientist-tokens`27- Product A/B and business metrics → `data-scientist`28- PII detection benchmarks, memorization, logging minimization → `privacy-research-engineer-safeguards`2930## Related skills3132| Need | Skill |33|---|---|34| Privacy research for safeguards | `privacy-research-engineer-safeguards` |35| Production safeguard path and rollout | `ml-infrastructure-engineer-safeguards` |36| Adversarial attack campaigns | `ai-redteam` |37| Governance sign-off and model cards | `ai-risk-governance` |38| Production eval harness in app | `ai-engineer` |39| General research methodology | `ai-researcher` |40| Classical ML and statistics | `data-scientist` |41| Token efficiency ablations | `research-engineer-scientist-tokens` |42| Release gates and ops cadence | `ai-lead-ops` |4344## Core Workflows4546### 1. Research framing (safety)4748Hypotheses, harm taxonomy, success metrics.4950**See `references/research_framing_safety.md`.**5152### 2. Benchmarks and datasets5354Golden sets, labeling, versioning.5556**See `references/safety_benchmarks_datasets.md`.**5758### 3. Model development5960Training, fine-tuning, ensembles.6162**See `references/classifier_model_development.md`.**6364### 4. Evaluation and metrics6566Slices, calibration, error analysis.6768**See `references/evaluation_metrics_analysis.md`.**6970### 5. Ablations and experiments7172Controls, reproducibility.7374**See `references/ablation_experiment_design.md`.**7576### 6. Handoff to production7778Promotion criteria, monitoring hooks.7980**See `references/research_to_production_handoff.md`.**8182## Outputs8384- **Research brief** — question, baseline, hypothesis, metrics85- **Dataset card** — sources, label schema, known limitations86- **Benchmark spec** — cases, categories, pass/fail rubric87- **Results table** — metrics by slice with confidence intervals where possible88- **Error analysis** — representative FP/FN clusters89- **Promotion recommendation** — go/no-go vs current production classifier9091## Principles9293- **Measure what policy cares about** — category-level recall on high-severity harms94- **Report failures honestly** — FPs hurt UX; FNs hurt safety95- **Hold out adversarial refresh** — do not train on the only test set96- **Reproducible** — seeds, data version, model hash, eval script97- **Separate research from ops** — research proves lift; infra ships it
daemon-blockint-tech