Shirin Khosravi Jam

You don't always need LLM!
The best AI Engineers know when to use what
Here's a simple roadmap to guide you 👇

If you’re serious about NLP (basic → LLMs), 
these are the only books you need to backup your theory.

NLP (basic to LLM) roadmap:

1️⃣ 𝗛𝗮𝗻𝗱𝘀-𝗢𝗻 𝗠𝗮𝗰𝗵𝗶𝗻𝗲 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗦𝗰𝗶𝗸𝗶𝘁-𝗟𝗲𝗮𝗿𝗻, 𝗞𝗲𝗿𝗮𝘀 & 𝗧𝗲𝗻𝘀𝗼𝗿𝗙𝗹𝗼𝘄 - Aurélien Géron
↳ Covers all core ML: regression, trees, SVM, neural nets.
↳ End-to-end projects in TensorFlow & Keras.
(Note: PyTorch is more popular now - but the concepts are same)

2️⃣ 𝗣𝘆𝘁𝗵𝗼𝗻 𝗡𝗮𝘁𝘂𝗿𝗮𝗹 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗣𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 𝗖𝗼𝗼𝗸𝗯𝗼𝗼𝗸 - Zhenya Antić, PhD & Saurabh Chakravarty, Ph.D.
↳ Tokenization, n‑grams, stop‑words, stemming, TF‑IDF, BM25, NER.
↳ Hands-on with NLTK, spaCy, PyTorch, Hugging Face.
↳ word2vec, GloVe, FastText (pre-transformer must-knows).

3️⃣ 𝗡𝗮𝘁𝘂𝗿𝗮𝗹 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗣𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗲𝗿𝘀 - Lewis Tunstall, Leandro von Werra, Thomas Wolf
↳ From BERT & DistilBERT to sentence-transformers.
↳ Hugging Face workflows for classification, QA.
↳ Super clear on transformer internals.

4️⃣ 𝗛𝗮𝗻𝗱𝘀-𝗢𝗻 𝗟𝗮𝗿𝗴𝗲 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 - Jay Alammar & Maarten Grootendorst
↳ Build a transformer from scratch.
↳ Prompting, embeddings, RAG, eval.
↳ Function calling, modern agent pipelines.

Possibly you could also add -
Build a Large Language Model (From Scratch) by Raschka, PhD

That's really it!
(there would be some overlap of topics across the books)

These books are really enough for your theoretical backing - further try to incorporate these principles in your hands-on projects.

--
♻️ Repost to if you found it helpful 💚
➕ Join 29.000+ AI/ML builders here: https://lnkd.in/ds_SzEUH

city2graph is a Python library that converts geospatial datasets into graphs (networks) - streets, buildings, transit networks, mobility flows, etc.

I found this library this week and it could save you weeks of work.

city2graph by Yuta Sato handles the entire pipeline:
GeoPandas → NetworkX → PyTorch Geometric

One library. Three integrations. Zero friction.

Real use cases:
1️⃣ Transportation networks
  ↳ GTFS data → GNN-ready graphs
  ↳ Analyze bus/tram/train systems at scale

2️⃣ Urban planning & digital twins
  ↳ Building/street morphology → graph representations
  ↳ POI proximity → spatial relationships

3️⃣ Mobility analysis
  ↳ Bike-sharing flows → temporal graphs
  ↳ Migration patterns → network models

Why it matters:
Most geospatial tools ignore modern ML.
Most ML tools ignore spatial relationships.

city2graph bridges that gap.
Check it out: https://lnkd.in/d3xatTZc

Disclaimer: Personally I am not using it, but think it could be helpful for many.

--
♻️ Repost to if you found it helpful 💚
➕ Join 28.000+ AI/ML builders here: https://lnkd.in/ds_SzEUH

5 NLP techniques you should know

before Transformers and LLMs... these ruled the world (and still do!)

1️⃣ 𝗕𝗮𝗴 𝗼𝗳 𝗪𝗼𝗿𝗱𝘀 (𝗕𝗼𝗪)
The simplest way to turn text into features - word counts in a matrix.
Still used in spam detection, topic tagging, and fast prototyping.
✅ Fast
✅ Interpretable
✅ Surprisingly effective

2️⃣ 𝗧𝗙-𝗜𝗗𝗙 (𝗧𝗲𝗿𝗺 𝗙𝗿𝗲𝗾𝘂𝗲𝗻𝗰𝘆 - 𝗜𝗻𝘃𝗲𝗿𝘀𝗲 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁 𝗙𝗿𝗲𝗾𝘂𝗲𝗻𝗰𝘆)
BoW but smarter - gives importance to rare words.
Still the foundation behind search systems and BM25 scoring.
✅ Great for keyword-based retrieval
✅ Works well in low-data settings

3️⃣ 𝗪𝗼𝗿𝗱𝟮𝗩𝗲𝗰
Google changed NLP forever with this one in 2013.
Word embeddings that capture meaning.
👑 King - Man + Woman = Queen
✅ Semantic similarity
✅ Reusable pre-trained vectors
✅ Still used in RAG, search, and recsys!

4️⃣ 𝗕𝗠𝟮𝟱
The powerhouse of modern search engines.
An improved version of TF-IDF that adds saturation and document length normalization.
✅ Powers keyword search in multiple search engines
✅ Used in hybrid RAG pipelines
✅ Works great with small models too!

5️⃣ 𝗙𝗮𝘀𝘁𝗧𝗲𝘅𝘁
Word2Vec + Subword magic (by Facebook).
It captures meanings of unseen words using character n-grams.
✅ Robust for rare words, typos, or different languages
✅ Still a strong baseline for classification tasks

𝗠𝗮𝗶𝗻 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆:
You don’t always need a Transformer or LLM.
Sometimes, good old-school methods just work.
Especially when you want:
✅ Speed
✅ Simplicity
✅ Control

Do you agree?

--
♻️ Repost to if you found it helpful 💚
➕ Join 28.000+ AI/ML builders here: https://lnkd.in/ds_SzEUH

If I was starting with ML Engineering in 2026.

I will follow these simple steps and resources to get started
(no over complication)

ML Engineering still exists and is much needed!
These are simple go to resources.

You can replace resources by any.
Just don't get drowned into it.

1️⃣ 𝗣𝘆𝘁𝗵𝗼𝗻 𝗺𝗮𝘀𝘁𝗲𝗿𝘆:
a. Arjan Codes (Youtube) - youtube.com/arjancodes
↳ Real-world patterns, not toy examples

b. 100 Days of Code (Udemy) - https://lnkd.in/dt26wVzW
↳ Build while you learn

2️⃣ 𝗦𝗤𝗟 𝗳𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝘀:
a. SQL Fundamentals (DataCamp) - https://lnkd.in/dp9nRNcN
↳ Production queries from day 1

3️⃣ 𝗠𝗮𝘁𝗵𝗲𝗺𝗮𝘁𝗶𝗰𝘀 (𝗱𝗼𝗻'𝘁 𝘀𝗸𝗶𝗽, 𝗯𝘂𝘁 𝗮𝗹𝘀𝗼 𝗱𝗼𝗻’𝘁 𝗼𝘃𝗲𝗿𝗱𝗼 𝗶𝘁 ):
a. Linear Algebra (Youtube) -https://lnkd.in/dmCSmw4q
↳ Visual intuition over formulas

b. Mathematics of Machine Learning (Book) - amazon.com/dp/1837027870
↳ The one ML math book you need

4️⃣ 𝗗𝗲𝗲𝗽 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴:
a. Deep Learning Specialization (Coursera) - https://lnkd.in/d_MUvhqb
↳ Taught by Andrew Ng himself

5️⃣ 𝗦𝘆𝘀𝘁𝗲𝗺 𝗗𝗲𝘀𝗶𝗴𝗻:
a. Stanford System Design (Youtube) - https://lnkd.in/dcnzFYHU
↳ Architecture patterns that scale

6️⃣ 𝗠𝗟𝗢𝗽𝘀:
a. Designing Machine Learning Systems (Book) - amazon.com/dp/1098107969
↳ Production ML bible by Chip Huyen

b. AWS ML Specialty Certification (Udemy) - https://lnkd.in/dntE2dCx
↳ Hands-on cloud deployment

Total cost: <$500
Total time: 6-9 months (honestly, it takes time)
ROI: Career-changing

Check the steps added below 👇

--
♻️ Repost to if you found it helpful 💚
➕ Join 29.000+ AI/ML builders here: https://lnkd.in/ds_SzEUH