AI fundamentals webinar session in progress

Webinar Materials & Announcements

AI fundamentals, session by session

Sessions you can attend or revisit

Each session runs 90 minutes and targets one specific concept — no broad overviews. You get recordings within 24 hours, so missing a live slot does not mean missing the material.

Upcoming — 14 Feb

Neural Networks from Scratch: Architecture and Training

90 min Online Olha Savchenko

We will build a 3-layer network using only NumPy — no PyTorch, no Keras. The session covers forward propagation, computing cross-entropy loss, and manual gradient descent across 6 training iterations. Participants who ran through the pre-session notebook reported spending about 40 minutes on it, which is the right amount of prep.

Weight initialization gets more attention than usual here, because random seed choices affect convergence in ways that beginner tutorials skip entirely. We will compare 3 initialization strategies on the same dataset and measure the difference in epochs needed to reach 88% accuracy.

Session details

Live — 21 Feb

Prompt Engineering for Production Systems

90 min

Structuring prompts for reliability in real applications — chain-of-thought, few-shot patterns, and output formatting that does not break downstream parsers.

Session details

Recorded

Attention Mechanisms Explained

85 min

How self-attention computes relevance scores, why positional encoding matters, and what the query-key-value pattern actually does when you trace through a real sentence.

Watch recording

Recorded

Overfitting and Regularization in Practice

80 min

Dropout, L2 penalty, and early stopping — tested on the same CIFAR-10 subset across 4 configurations. The gap between training and validation accuracy tells you more than any single metric.

Watch recording

Upcoming — 28 Feb

Evaluating Language Models Honestly

90 min

BLEU scores, perplexity, and human evaluation — what each one actually measures and where each one misleads you if you rely on it alone.

Session details

What participants reported after 8 weeks

After the first cohort completed the full session sequence, we collected structured feedback from 34 participants. These numbers reflect what they reported — not what we hoped for. Some improved faster, some slower, and a few dropped off partway through.

12% 67%

Concept confidence

Share of participants who rated themselves confident explaining gradient descent to a colleague — before vs. after the sequence.

4 hrs 55 min

Time to first working model

Median time to produce a working binary classifier from a blank notebook — measured at session 1 vs. session 7.

19 6

Avg. debug cycles per task

How many attempts participants needed to fix a broken training loop — the drop reflects pattern recognition, not luck.

3 28

Sessions attended per person

Average attendance across the cohort. Recordings helped participants who missed live slots catch up within 48 hours.

Session leads

Who runs these sessions

Olha Savchenko ML Systems Instructor

Olha has been working with neural network implementations since 2017, mostly in applied computer vision contexts. She spent 5 years building training pipelines for a logistics company before shifting to instruction. Her sessions tend to be dense — she covers about 30% more ground than average because she skips motivational framing and goes straight to the mechanism.

01 Neural Networks from Scratch Forward pass, backpropagation, weight initialization — built manually in NumPy across 90 minutes.
02 Overfitting and Regularization in Practice Four configurations tested on the same dataset so you can see the difference, not just read about it.

Daryna Kovalchuk NLP & Evaluation Specialist

Daryna focuses on language model behavior — specifically on the gap between benchmark performance and what actually happens when you deploy a model in a real pipeline. She has run evaluation frameworks for 3 different research groups and has a particular interest in when standard metrics mislead practitioners. Her sessions on prompt engineering and model evaluation are the most frequently rewatched in the archive.

03 Prompt Engineering for Production Systems Chain-of-thought, few-shot patterns, and output formatting — tested against real parser failures.
04 Evaluating Language Models Honestly BLEU, perplexity, and human evaluation — where each metric helps and where it actively misleads.