Technical systems you can touch
Open the black box.
Stop reading about how LLMs work and start touching the systems themselves. Interactive playgrounds where you type real text and watch the internals respond — tokenization, context windows, sampling, and the full transformer architecture, one hands-on experiment at a time.
FDE@ProdAI · LABSMembership access
Learn free. Build with Pro. Ship in a Live Cohort.
Free explains production AI. Pro unlocks your own infrastructure and the complete structured path. Live Cohort adds mentors, reviews, deadlines, and a deployed capstone.
Showing 9 of 9 labs
Knowledge Graph RAG
When the answer lives between documents
Some questions can't be answered by any single passage — the answer is stitched across documents, in the relationships between entities. See where vector search runs out of road, how a corpus becomes a graph of entities and relationships, and the two ways to reason over that graph.
- ▸Where vector search runs out of road — multi-hop & global questions
- ▸How a corpus becomes a graph of entities and relationships
- ▸Two ways to reason: community summaries vs dual-level keywords
- ▸Interactive: match a question type to the right retrieval
The Evolution of Search
Keyword → Semantic → Hybrid → Agentic
Walk the four eras of search — from exact-match keywords (BM25) to meaning-based vectors, hybrid fusion, and autonomous agents. Compare sparse vs dense encoding, see what vectors are actually for, and why hybrid is state-of-the-art today.
- ▸Interactive timeline of the four search eras
- ▸Sparse vs dense encoding, side by side
- ▸What vectors are used for — text, images, audio
- ▸Why hybrid (keyword + semantic) wins today
Hybrid Search (Keyword + Semantic)
Fuse BM25 and vectors with a weight slider
Run keyword (BM25) and semantic (real Titan embeddings) search in parallel and fuse them with Reciprocal Rank Fusion. Drag the weight slider and watch exact-term matches and meaning-based matches trade places — hybrid is state-of-the-art search today.
- ▸Keyword BM25 + semantic vectors, live
- ▸Reciprocal Rank Fusion with a tunable weight slider
- ▸Per-doc keyword-rank vs semantic-rank vs fused score
- ▸See why neither method alone is enough
Semantic Search (Real Embeddings)
Search by meaning, not keywords
Type a question and rank documents by meaning using real Amazon Titan embeddings — server-side, with Amazon OpenSearch k-NN when configured. See semantic search actually work, not a simulation.
- ▸Real embeddings from Amazon Titan (Bedrock), server-side
- ▸Ranks documents by meaning, not keyword overlap
- ▸Amazon OpenSearch k-NN backend when configured
- ▸Credentials stay on the server — never in the browser
HNSW Tuning Workspace
Find the recall, latency, and memory operating point
Tune M, ef_construction, ef_search, and vector precision against a fixed 1M-vector benchmark. See the simulated Recall@10, P95 latency, memory, and index-build trade-offs before applying the same workflow to a real index.
- ▸Tune M, ef_construction, and ef_search together
- ▸Watch Recall@10 and P95 latency move on an operating curve
- ▸Compare FP32, FP16, and INT8 memory trade-offs
- ▸Use a repeatable exact-ground-truth tuning checklist
RAG Readiness Scan
Parse a PDF, page by page, before RAG
Parse a real PDF into page-level text, fragmentation, and extraction-order diagnostics before committing it to a RAG pipeline. See exactly which pages need OCR, layout recovery, or a manual look before you chunk.
- ▸Parse a real PDF in your browser — page by page
- ▸Flags pages that need OCR, layout recovery, or manual review
- ▸Fragmentation & reading-order diagnostics from text coordinates
- ▸Per-page verdict table + a document readiness score
Tokenizer & Context Window
See exactly what the model reads
Type any text and watch it split into the real tokens a model sees. Then map those tokens against real context-window limits (up to 1M tokens) to understand how much room a prompt actually uses.
- ▸See the real tokens behind any text you type
- ▸Color-coded token chips — toggle to reveal each token's ID
- ▸Compare how the same text tokenizes across models
- ▸Context-window meter against real limits (8K → 1M tokens)
Temperature & Sampling
How randomness shapes output
Visualize how temperature, top-p and top-k reshape a probability distribution over next tokens.
- ▸Live probability bar chart that reshapes as you drag temperature
- ▸Watch top-p / top-k trim the distribution in real time
- ▸Roll the dice — sample tokens and watch frequencies converge
- ▸Determinism meter from predictable to random
Context Budget Planner
Fit RAG chunks into a window
Plan a real prompt budget: system prompt + retrieved chunks + chat history + expected output, all measured with the real tokenizer against a chosen model's context limit.
- ▸Allocate tokens across prompt sections
- ▸Live overflow warnings against the model limit
- ▸Understand why long RAG contexts get truncated
Pairs with the Knowledge Lab
Intelligent Search — the written deep-dives
The Search labs above are the hands-on half. The Intelligent Search series in the Knowledge Lab covers the theory in depth — vector search, sparse vs dense encoding, similarity metrics, kNN & HNSW, quantization, hybrid search, and evaluation.
Core AI Intuitions
Build deep intuition for the core operations that power all of AI — dot products, softmax, broadcasting, and norms.
Complete Free anchor lessons appear throughout the track · the surrounding sequence is Pro
See the full Free and Pro roadmap →Similarity With Dot Product
The one operation behind embeddings, search, and attention.
Softmax Probabilities
Turn raw scores into probabilities that sum to one.
Tensor Broadcasting
How arrays of different shapes do math together.
L1 vs L2 Norms
Two ways to measure the size of a vector.
Matrix Multiplication as Transformation
How a matrix reshapes, rotates, and scales a vector.
Cosine vs Euclidean
When direction matters, and when distance does.
Gradients & Gradient Descent
How models learn by following the slope downhill.
Logits to Probabilities
From raw scores to probabilities and log-probs.
Vectors & Vector Spaces
The arithmetic behind king − man + woman = queen.
Attention as a Weighted Average
Softmax the scores, then blend the values.
Deep-dive course
Transformer Architecture
The full “Attention Is All You Need” architecture, built up one component at a time. Every chapter shows real computed numbers, tensor dimensions at each step, the paper figure it reproduces, and what breaks without the piece. Fourteen interactive chapters.
Tokenization (BPE)
Raw text → integer token IDs via Byte-Pair Encoding
Free preview
Token Embeddings
One-hot × Wₑ × √d_model — sparse IDs to dense vectors
Pro
Positional Encoding
Sinusoidal functions that inject sequence order
Pro
Scaled Dot-Product Attention
softmax(QKᵀ/√d_k)·V — the core operation
Free preview
Multi-Head Attention
h parallel heads, same cost, richer representation
Pro
Causal Masking
Blocking future tokens with −∞ before softmax
Pro
Layer Normalization
γ⊙(x−μ)/√(σ²+ε)+β — stabilizing activations
Pro
Feed-Forward Network
max(0, xW₁+b₁)W₂+b₂ — the per-position nonlinearity
Pro
Residual Connections
LayerNorm(x + Sublayer(x)) — the gradient highway
Pro
Encoder–Decoder & Cross-Attention
The full architecture and three attention types
Pro
Training
Cross-entropy, label smoothing, warmup schedule, dropout
Pro
Inference & Generation
Beam search, temperature, KV-cache, sampling
Free preview
End-to-End Trace
Complete forward pass — every tensor, every multiply
Live Cohort
Live Playground
Real embeddings — type text, see vectors
Live Cohort
Labs are how we teach production AI
Every FDE@ProdAI cohort pairs deep articles with hands-on labs like these. Understand the internals, then build systems that ship.