Superstar FDE: Inference Cloud, End to End

The forward deployed engineer at an inference and fine-tuning cloud (Fireworks, Together AI, Baseten style) is the one person who can follow a customer from the first discovery call to a tuned production deployment: run the call, size the deployment, benchmark it, read the customer’s model code, scope the fine-tune, take over from the ML engineer or the performance engineer when needed, and relay to research without losing the detail.

This path is that role, assembled from six parts. Each part is a path on its own for a narrower role, and anything you finish in a part counts here.

flowchart LR
    RB["0. Role brief"] --> LF["LLM Foundations"]
    LF --> FM["Frameworks and Models"]
    FM --> TR["Training"]
    FM --> IP["Inference Performance"]
    IP --> FS["AI Full Stack"]
    TR --> FE["Field Engineering"]
    IP --> FE
    FS --> FE
    FE --> CAP["Capstone: one customer, end to end"]

Part	Stages	Also the path for
LLM Foundations	history, transformer math, architecture variants	anyone new to LLMs
Frameworks and Models	PyTorch, JAX, Keras 3; loading from the Hub	ML engineer onboarding
Training	RL foundations; pretraining, post-training, LoRA, scoping	research-adjacent ML engineer
Inference Performance	engine internals, quantization, frameworks, serving and load, platforms	performance engineer (MTS)
AI Full Stack	streaming, retrieval, evaluation, routing	AI application engineer
Field Engineering	discovery, sizing, performance testing, POC, migration, commercials, escalation	sales engineer, solutions architect
Capstone	one mock customer through every part	this role

The technical parts build depth; Field Engineering turns it into customer outcomes, and the capstone proves both. If you already work in inference, start with Field Engineering to see what the depth is for, then fill the gaps it exposes.

How to use this path

practice/bin/ol learn superstar-fde                          # every stage, grouped by part, with progress
practice/bin/ol learn superstar-fde next                     # read the next unfinished stage
practice/bin/ol learn superstar-fde inference-performance:4  # read one stage of a part
practice/bin/ol learn superstar-fde --done inference-performance:4

A PDF of the whole path, one Part per section, is attached to every GitHub Release (supersource-superstar-fde.pdf).

What the job is

The company landscape (what Fireworks, Together AI, and Baseten sell), who owns what (FDE vs sales engineer, ML engineer, performance engineer, research, account executive), and the customer journey live in the Field Engineering track, because every customer-facing role needs them.

Which part answers which customer question

Customer says	Where the answer is
“Why is DeepSeek cheaper to serve than a dense model of the same size?”	LLM Foundations 3 (MoE, MLA), Inference Performance 1
“Our fine-tune works in Transformers but outputs garbage on your endpoint.”	Frameworks and Models 2 (chat templates, tokenizers), Training 2
“We wrote the model in JAX. Can you serve it?”	Frameworks and Models 1 and 2
“How many GPUs do we need for 50 req/s?”	LLM Foundations 2 (FLOPs, KV size), Inference Performance 4, Field Engineering 2
“Latency spikes every few minutes.”	Inference Performance 1 and 4 (chunked prefill, preemption)
“Prove you are faster than our current provider.”	Field Engineering 3, Inference Performance 4
“Should we do LoRA or full fine-tuning? Or RL?”	Training 2
“Did the switch make our product worse?”	AI Full Stack 3, Field Engineering 5
“What does this cost us next year, and can our data stay in the EU?”	Field Engineering 6

Done when for stage 0: without notes, you can say which product surface (serverless, dedicated, batch, fine-tune, BYOC) you would propose for three workloads: a 5 req/s internal chatbot, a nightly 200M-token classification job, and a regulated healthcare assistant that cannot leave the customer’s VPC.

Stages

Generated from path.tsv. Track progress locally with practice/bin/ol learn.

Stage	Read	Done when
0	Role brief Superstar FDE: Inference Cloud, End to End	You can say what an inference cloud sells, what an FDE owns, and which part of this path answers each customer question.
Part	LLM Foundations	Every stage of that path
Part	Frameworks and Models	Every stage of that path
Part	Training and Post-Training	Every stage of that path
Part	Inference Performance	Every stage of that path
Part	AI Full Stack	Every stage of that path
Part	Field Engineering	Every stage of that path
1	Capstone: one customer, end to end Capstone: One Customer, End to End	You deliver every capstone artifact for one mock customer, from discovery notes to a migration plan, and each one passes its stated check.