Learn by Building. Master AI from the Inside.

Build a small linear algebra library using pure Python implementing vectors, matrices, dimension validation, dot products, transposes, and matrix multiplication, verified numerically against NumPy.

Train a regression model without using a predefined machine-learning estimator—implementing linear predictions, Mean Squared Error loss, manual analytical gradients, gradient descent optimization, loss curve visualization, and held-out evaluation.

Create a supervised classifier that identifies spam messages—implementing text cleaning, Bag-of-Words / TF-IDF feature extraction, Multinomial Naive Bayes or Logistic Regression from scratch, and full classification error analysis.

Build a micrograd-style scalar automatic differentiation engine in pure Python that dynamically records computational graphs, performs topological sorting, backpropagates exact gradients via the chain rule, and trains a multilayer perceptron.

Implement and train a multilayer neural network in pure NumPy without high-level deep-learning frameworks—including dense layers, activation functions, forward caching, explicit matrix backpropagation, weight initialization, and mini-batch gradient descent.

Train a handwritten digit recognizer or image classification model in PyTorch—implementing dataset normalization, DataLoaders, network architecture, training/validation loops, confusion matrix diagnostics, failure inspection, and checkpoint save/load inference.

Train a character-level language model capable of generating short text—building a character vocabulary, context-target sliding windows, a bigram baseline, a neural language model, perplexity tracking, and temperature/top-k sampling.

Implement Byte Pair Encoding (BPE) tokenization from scratch in Python—starting from raw UTF-8 bytes, iteratively merging the most frequent adjacent token pairs, building a deterministic merge vocabulary, handling special tokens, and verifying 100% lossless encode/decode round trips.

Implement causal self-attention and multi-head attention from scratch in PyTorch—including Q, K, V projections, √d_k variance scaling, upper-triangular causal masking, attention softmax probabilities, head splitting/concatenation, and causality unit tests.

Build a complete decoder-only Transformer (Mini GPT) from scratch in PyTorch—including token and positional embeddings, Pre-Norm residual Transformer blocks, causal multi-head attention, MLP feed-forward networks, language modeling head, training loop, checkpointing, and autoregressive generation.

Train and evaluate a small GPT-style language model from randomly initialized weights using your own dataset pipeline, custom BPE tokenizer, AdamW + cosine decay pretraining loop, and evaluation harness.

Convert a base pretrained language model into an instruction-following conversational assistant using custom ChatML templates, prompt-masked Supervised Fine-Tuning (SFT), Low-Rank Adaptation (LoRA), and Direct Preference Optimization (DPO).

Build a production-grade LLM inference engine and streaming HTTP/SSE server featuring prefill/decode KV-cache management, INT8/INT4 weight quantization, speculative decoding, security guardrails, and TTFT/TPOT benchmarking.

Deconstruct a foundational AI/LLM research paper, formulate a falsifiable hypothesis, run a parameter- and FLOP-matched experimental study across multiple random seeds, and author a publication-grade technical report.