Jobs: AI eng, ML eng, research, FDE, PM

Roles (practical, not HR)

Job titles in AI are inflated and overlapping. "AI Engineer" at a startup may mean "build RAG + FastAPI"; at a lab it may mean "pretrain infrastructure." Hiring managers look past titles for signals: shipped systems, eval discipline, cost awareness, failure-mode taste.

This lesson gives a practical map — not HR taxonomy — so you can aim the rest of ShipAI at a target and articulate your story in interviews.

Titles blur. A strong LLM engineer candidate often shows: a RAG or agent repo with JSON evals, notes on cost per query, and clear "what broke." A strong ML engineer shows: reproducible training, metrics on held-out data, and serving path. Research roles want depth on one technical wedge (attention implementation, fine-tuning method) plus publication or open replication.

Callout — hire signals beat buzzwords: "Built multi-agent swarm" without evals loses to "built support bot with 50-case golden set, 88% pass rate, $0.004/query." ShipAI trains the second portfolio.

What each role actually does day-to-day

AI / LLM engineer

• Design prompt + tool schemas; implement agent loops and retries.
• Wire retrieval (chunking, embeddings, reranking) and measure answer quality.
• Partner with PM on eval sets; block launches on regression.
• Debug production: hallucinations, tool parse errors, latency spikes.

Gap signals: never logged token costs, no structured eval JSON, cannot explain RAG failure modes.

ML engineer

• Own data pipelines, training jobs, model registry, batch and online inference.
• Maintain classical baselines alongside neural models.
• Monitor drift; retrain on schedule or triggers.

Gap signals: never ran train/val split honestly, no experience saving checkpoints, treats models as black boxes.

Research scientist

• Propose experiments, implement novel architectures or training objectives, write up results.
• Less emphasis on CRUD apps; more on compute-efficient training and benchmarks.

Gap signals: cannot implement baseline from paper; no ablation discipline.

Forward-deployed / solutions engineer

• Embed with customer; map messy requirements to shippable POC.
• Integrate with SSO, CRM, on-prem constraints; hand off to eng for hardening.

Gap signals: demos that break on fresh data; no documentation for handoff.

PM (AI)

• Define success metrics before model choice; prioritize eval cases with legal/compliance.
• Design UX for uncertainty (citations, confidence, escalation).
• Negotiate cost vs. quality with eng.

Gap signals: success = "felt smart"; no measurable eval definition.

Using ShipAI as a gap-closing plan

You are not expected to arrive complete. You are expected to leave with artifacts that prove closure on gaps you named honestly.

Examples of gap → module mapping:

Your gap list should be personal — copied from job posts you want, not generic.

Engineering problem (staff framing)

Titles blur. Maintain an honest gap list mapped to portfolio proof.

Diagram — Role map

flowchart LR
  AIE[AI Eng] --> Sys[Agents/RAG/evals/cost]
  MLE[ML Eng] --> TR[Train/data/serve]
  Sys --> Port[Portfolio]
  TR --> Port

Precise definitions & mental model

AI eng vs ML eng vs research vs FDE; gap list as living artifact.

Tradeoffs — when to use what

Breadth tour vs depth on 2 systems — pick consciously.

Failure modes (interview + on-call)

Claiming agents for one ChatCompletion; no metrics; stale gap list.

Production & OSS practices

Job hunt as product: ICP, positioning, proof, mock-interview iteration.

Deep dive (FAANG / OSS bar)

Push «jobs-and-your-gap-list» past tutorial depth: write the interface contract (inputs/outputs/invariants), list three measurable metrics, and name two degrade modes if the happy path fails. Add a short threat note: what an attacker or noisy tool result could do, and which layer catches it (schema, policy, HITL, or eval gate).

flowchart LR
  Contract[Interface contract] --> Metrics
  Metrics --> Degrade[Degrade modes]
  Degrade --> Threat[Threat + control]

Micro-project: Personal gap list

Create m1/gaps.md:

1. State your target role (primary) and optional secondary.
2. List 5–10 skills or experiences you lack for that role — specific ("never wrote pytest for tool outputs," not "need more AI").
3. Map each gap to a ShipAI module or milestone that closes it.
4. Mark top 3 priorities for the next 4 weeks of study.

Example entry:

- **Gap:** No experience serving a fine-tuned model behind an API
  **Closes in:** SLM module (adapter + serve) + deployment module
  **Priority:** P1

Keep it honest. Interviewers probe gaps you claim to have closed — your portfolio should show the work.

Milestone 1 — Field-map write-up

Publish m1/FIELD_MAP.md (blog-style, 800–1500 words) that stitches the module together:

1. Arc narrative — symbolic AI limits → classical ML loop → deep learning representations → transformers/GPT lineage → generative systems landscape.
2. Build vs. buy — summarize your open-weight vs. API matrix conclusion; name one driver that mattered most.
3. Your gap list — how the rest of ShipAI closes it (training, RAG, agents, evals, deploy).
4. Links — essay (symbolic-wall.md), timeline, sklearn metrics, small NN comparison, product landscape table, decision matrix.

Update root README: mark Milestone 1 complete with link to FIELD_MAP.md.

This document is interview gold: a coherent two-page story beats scattered repos.

Callout — write for a smart colleague, not a textbook: FIELD_MAP should sound like you explaining the field after doing the work — citations to your own artifacts, one vivid example from your essay, no buzzword soup.

Checklist

• m1/gaps.md committed with 5–10 mapped gaps
• m1/FIELD_MAP.md committed (800–1500 words, linked artifacts)
• Milestone 1 marked done in root README