Topic 17.10
Where to Go Next: AI Engineer, Forward Deployed Engineer, AI Scientist
In one line
What three AI-era roles do day to day, which skills from this course each one uses most, and what to learn after it, with the Hectal courses that already exist and the ones that are planned.
Think of it like this
Three people who all say 'I work with hospitals'. One builds the machines (an engineer), one visits each hospital to set the machines up and fix what is unique to that building (a field engineer), and one studies new treatments in a laboratory (a scientist). They share a language and some tools, but their days look very different. The same is true of AI roles: the titles overlap, and each company uses them a little differently, so read the job description, not just the title.
Words you'll meet
New words in this topic, in plain English. Come back here whenever one feels fuzzy.
- AI engineer
- A software engineer who builds products and services using existing AI models.
- Forward deployed engineer
- An engineer who works with customers to make a product work in their own setting.
- AI scientist
- A person who researches and trains models and designs experiments to improve them.
- Fine-tuning
- Training an existing model further on your own data.
- Evaluation (eval)
- A repeatable test that scores how well an AI system does on example tasks.
- MLOps
- The practice of deploying, monitoring and updating machine learning systems reliably.
- Ablation
- An experiment that removes one part of a system to see how much it mattered.
Step by step
01Roles are bundles of skills
A job title is a bundle of skills in different amounts. Instead of asking 'which role is best', ask 'which skills do I enjoy using every day, and which am I willing to build up?' The first runnable example is a small self-assessment: it stores what each role leans on, rates your skills from 0 to 3, and prints how many points of each role's profile you already cover and the biggest gaps.
The numbers in the example are invented for the exercise and are not a measurement of the job market. Change them to match the job descriptions you actually read and your honest self-rating.
02Day-to-day, in plain words
An AI engineer might spend Monday improving how a retrieval step splits documents, Tuesday adding retries and logging to a model call, Wednesday writing an evaluation set, and Thursday fixing a deployment. A forward deployed engineer might spend a week at a customer cleaning messy data, building a workflow around the product, and presenting the result. An AI scientist might spend the week running three experiments, reading two papers and writing up what failed. All three write Python every day.
03Plan a learning order with dependencies
Topics depend on each other. The second runnable example uses graphlib.TopologicalSorter from the standard library to turn 'what I need first' into an order. Planned courses are marked as planned. If you accidentally create a cycle (A needs B needs A), graphlib raises CycleError: a useful reminder that a plan with a circular dependency cannot be followed.
04What to do this week
Pick one of three: (1) extend docqa with a real embedding model and a 20-question evaluation set, (2) solve ten DSA problems in Python and explain each aloud, (3) turn a messy CSV into a clean report with pandas and a script that anyone can run. Each takes a weekend, and each is something you can show in an interview.
Try it yourself
- 1
Rate yourself honestly
Edit the
minedictionary in the first example. Give a 3 only for things you could teach. Then look at the biggest gaps for the role you are most curious about, and match each gap to a topic in this course or a course on Hectal. - 2
Add your own goals
In the second example, add a node for something you want to learn (for example 'cloud basics' needing 'docker') and see where it lands in the order.
Code & diagrams
Edit the `mine` ratings to match yourself and rerun. The role profiles are illustrative, not data.
Expected output
AI Engineer: 10/20 points covered (50%)
biggest gaps: async (+2), prompts (+2), systems (+2)
Forward Deployed Engineer: 8/17 points covered (47%)
biggest gaps: communication (+3), systems (+2), apis (+1)
AI Scientist: 7/16 points covered (44%)
biggest gaps: experiments (+3), data (+2), math (+2)Courses marked planned do not exist yet. The order shown is one valid order, not the only one.
Expected output
1. python (this course)
2. dsa
3. databases
4. docker
5. machine learning (planned)
6. system design
7. kubernetes
8. deep learning (planned)
9. llm engineering (planned)
CycleError: nodes are in a cycleBreak it on purpose
Errors are the best teachers. Make each change, read the error, guess what went wrong, then reveal the answer.
Break #1
Misspell a key in a lookup table
Look up a role name with a trailing space.
ROLES = {"AI Engineer": ["python", "apis"]}
print(ROLES["AI Engineer "])Myth vs fact
Myth
I must pick one role before I start learning.
Fact
The first 80 percent of the skills overlap. Learn the shared base, build projects, and let your preferences guide the last stretch.
Myth
AI scientists do not need software engineering, and AI engineers do not need maths.
Fact
Both need some of each. Research code that nobody can rerun is hard to trust, and engineers who cannot read an evaluation or reason about data make poor decisions.
Myth
A model API replaces the need to learn programming.
Fact
Models help you write code faster, but you still have to read it, test it and own it. The people who get the most from assistants are the ones who can already judge the output.
Pro corner
Extra depth for experienced readers. New to this? Skip it for now and come back later.
- ▸
Hiring bars and titles change quickly. Read recent job descriptions in the area you want, collect the skills that repeat, and compare them with your project history, not your course list.
- ▸
Whatever the title, expect interviews to test fundamentals before AI topics: a Python coding round (Topics 17.1 to 17.4 and the
/dsacourse), a design discussion (/system-design) and questions about something you built. Practise explaining your docqa choices aloud (why overlap, why a Protocol, why a fake model, what you would measure next), because "walk me through a project" is asked in nearly every loop. - ▸
A portfolio of two or three small, finished, well-documented projects (with tests and a README, as in this capstone) says more than a long list of technologies.
- ▸
Evaluation is the shared skill across all three roles: whoever can measure whether a change helped is the most valuable person in the room. Keep practising it from Topic 16.11.
Remember this
- 1
AI engineer. Builds products and services on top of models that already exist: calling model APIs, writing prompts and tools, retrieval systems like docqa, evaluation, caching, cost and latency control, and the service around it (APIs, queues, databases, deployment, monitoring). The day is mostly software engineering with AI components. Most-used skills from this course: functions and classes (Phases 5 and 8), async and concurrency (12), type hints (10), testing (11), Pydantic, APIs and services (15), the AI-era toolkit (16), and this phase's capstone. To learn next: LLM engineering (RAG quality, agents, evaluation, and the basics of fine-tuning, which means training an existing model a bit more on your own examples), system design, databases and deployment. System design (
/system-design), databases (/database-design), Docker (/docker) and Kubernetes (/kubernetes) already exist on Hectal. - 2
Forward deployed engineer (FDE). Works directly with customers to make a product succeed in their real environment: understanding their data and process, building integrations and custom workflows, debugging in situ, and feeding what they learn back to the product team. The day mixes coding, talking to people, data wrangling and troubleshooting. Most-used skills: fast, readable Python, files and data formats (Phase 7), pandas and APIs (14 and 15), HTTP clients, SQL (the language for asking databases questions), debugging, writing clearly, and the ability to scope a problem and ship a working version quickly. To learn next: the cloud basics (
/aws), containers (/docker), databases (/database-design), security basics, and the communication habits of consulting work. - 3
AI scientist or researcher (ML scientist). Studies and builds models: designing experiments, training and evaluating models, reading papers, running ablations, and improving data. The day is mostly experiments, analysis and reading. Most-used skills: NumPy and pandas (Phase 14), the scientific workflow (16.2), classic machine learning (16.12), tensors and PyTorch (16.13), maths (linear algebra, probability, calculus), careful evaluation and reproducibility. To learn next: machine learning and deep learning courses (planned on Hectal), statistics, and how to read a paper end to end.
- 4
Skills that every role shares. Reading and writing Python with confidence; testing and debugging; Git and a terminal; knowing the data (shapes, missing values, bias); estimating cost and speed; and clear writing. Data structures and algorithms (
/dsa, with a Python solution for every problem) matter for interviews in all three roles, and for writing code that scales. Interview loops differ: AI engineer loops look like software engineering loops plus an AI design round, FDE loops add customer scenarios, and scientist loops add maths, ML theory and research discussion. - 5
Which Hectal courses exist today and which are planned. Existing now: Data Structures and Algorithms (
/dsa), System Design (/system-design), Docker (/docker), Database Design (/database-design), Kubernetes (/kubernetes) and AWS (/aws). Dedicated courses for LLM engineering, machine learning and deep learning are planned but not available yet, and they are meant to build on this course. Until then, use Phase 16 as the map and the official documentation of the libraries as the territory. - 6
A realistic plan. Pick a role as a direction, not a contract. Spend the next month building one more project end to end (docqa with your own documents is a good start, then add a real embedding model and an evaluation set). Learn one area deeply rather than five shallowly. Read other people's code. Write about what you built. Roles and tools change quickly; the habits of testing, measuring and communicating do not.
Explain it without notes
Describe the day-to-day work of an AI engineer and which Python skills it uses most.
How does a forward deployed engineer's work differ from an AI engineer's?
What do AI scientists do that engineers usually do not, and what should they learn next?
Which skills are shared by all three roles?
Practice
Write a function coverage(profile, mine) returning the percentage (0 to 100, rounded to one decimal) of a role's required points that you cover.
Write a function biggest_gaps(profile, mine, n) returning the n skills with the largest shortfall, largest first, ties alphabetical.
Use graphlib to print a valid learning order for three topics where b needs a and c needs b.
Trade-offs
- ↔
Breadth versus depth. Breadth keeps options open early; depth is what gets you hired and trusted. Build a broad base, then go deep in one area.
- ↔
Research versus product. Research roles reward novelty and rigour, product roles reward reliability and delivery. Choose the feedback loop you enjoy: long experiments or short shipping cycles.
Done when you can
Done when you can describe the three roles in your own words.
Done when you can name the Python skills each role uses most and your own gaps.
Done when you can say which Hectal courses exist now and which are planned.
Done when you have picked one project to build next.