[ CURRICULUM ARCHITECTURE ]

Learn

Three ways we teach. None of them are a lecture.

Track 01 Hands-On • 2-4 Hours

Build Sessions

You show up, you build something real, you leave with it working. No slides-only sessions — every Build Session ends with something you can point to and say "I made that."

Topic 01 • System Architecture

Autonomous Tool Loops From Scratch

Writing prompt parsers, schema definitions, and shell-execution dispatchers in pure Python with zero LangChain or external framework abstractions.

Topic 02 • Hardware Optimization

Weight Quantization for Consumer GPUs

Compiling llama.cpp with Metal/CUDA kernels and quantizing Mistral 7B to 4-bit precision so it runs on 8GB RAM laptops at 30+ tok/sec.

Topic 03 • Search & Retrieval

Hybrid Keyword + Vector Embedding Search

Building BM25 sparse index combined with dense vector retrieval using Reciprocal Rank Fusion (RRF) for accurate technical documentation search.

Topic 04 • Scheduled Next

Edge SLMs on Embedded Systems

Stripping overhead from 1B-3B parameter models to run local inference loops on Raspberry Pi and low-wattage developer devices.

Topic 05 • Pipeline Design

Structured JSON Extraction Without Halts

Enforcing strict CFG (Context-Free Grammar) state machines on raw token sampling logits to guarantee valid schema JSON output 100% of the time.

Topic 06 • Verification

Deterministic Evals for LLM Applications

Replacing human "vibes" with reproducible unit tests, semantic assertion matrices, and synthetic benchmark suites before shipping to users.

Track 02 Failure Analysis • Live Code

Post-Mortems

Most tech content only shows you the finished, working version. We do the opposite — we break down what actually went wrong on a real project, live, so you learn what the polished tutorials skip.

Post-Mortem 01

The $1,400 Infinite Agent Loop

How a recursive tool-calling agent hallucinated error codes, retried itself indefinitely, and burned 80 million tokens overnight — and the circuit breaker pattern that stops it.

Post-Mortem 02

Why 90% of RAG Demos Break in Production

Chunking 500-token paragraphs destroys table hierarchies and tabular data. We look at real failing queries, dissect embedding cluster collisions, and rebuild the pipeline.

Post-Mortem 03

Prompt Injection in Multi-Tenant Apps

How user-uploaded resumes hijacked system prompts to leak other users' evaluations, and why simple regex filters fail against Unicode-encoded payloads.

Post-Mortem 04

OOM Crashes Under Concurrent Streaming

When 50 concurrent SSE connections exhaust server memory buffers during token generation. Profiling async coroutines, backpressure, and socket teardowns in Python.

Post-Mortem 05

Vector Drift in Long-Running Indexes

Why semantic cosine distances degrade when domain terminology evolves over six months, and how to detect stale embedding clusters before search quality crashes.

Post-Mortem 06

The Synthetic Fine-Tuning Collapse

How training a smaller model on unfiltered model outputs caused repetitive output loops and mode collapse after epoch 3 — and how data curation fixed it.

Track 03 Emerging Frontier • Non-Mainstream

Ahead-of-the-Curve

By the time something's "officially supported," it's already mainstream. We cover what's two steps ahead — the tools, models, and techniques worth knowing before everyone else catches up.

Emerging Tech 01

Speculative Decoding Engines

Using small draft models to predict token sequences ahead of large verifier models, achieving 2x to 3x throughput speedups without losing mathematical accuracy.

Emerging Tech 02

State-Space Models & Mamba Architectures

Linear-time sequence modeling that eliminates quadratic attention bottlenecks for ultra-long context windows on hardware with constrained VRAM.

Emerging Tech 03

BitNet b1.58 1-Bit LLM Architectures

Ternary weights {-1, 0, 1} that replace matrix multiplication with pure integer additions, dropping energy consumption and enabling local execution on phones.

Emerging Tech 04

WebAssembly + WebGPU In-Browser Inference

Running vision and language models directly inside client web browsers via WebGPU compute shaders with zero server infrastructure overhead.

Emerging Tech 05

Direct Preference Optimization (DPO) Mechanics

Fine-tuning models on human preference pairs without training unstable reward models or reinforcement learning value heads.

Emerging Tech 06

Graph RAG & Knowledge Triplet Extraction

Connecting entities across document corpora using graph databases and Cypher queries to answer high-level thematic queries that vector search misses.

Ready to build one of these live?

We implement these tracks during our in-person sessions. View what we're running next on our Events page, or ask builders questions in our Community channels.