OPTIVERSE SYSTEMS JOURNAL

Make sense of AI.
From the inside out.

A field guide for people who want to understand, build, and run AI systems. Follow the evidence, experiment with the diagrams, and learn where systems succeed—or fail.

25 guidesBilingualPractical exercises

Start with a question

Full reading route

01 / Understand models4 guides incl. prerequisites

What is the model actually doing?

Start with tokens, context, and the limits of evidence.

Thehotel?

has62%offers25%is13%

Next-token candidates · illustrative

Explore this route

02 / Build applications17 guides incl. prerequisites

How do answers become reliable actions?

Connect evidence, validation, permissions, and recovery.

KYO-H12 / RATE$240 / nightValid during April

Supported answer$240 per night in April[RATE]

Explore this route

03 / Run inference10 guides incl. prerequisites

Where do time and memory go?

Explore scheduling, KV memory, and GPU placement.

825 tokens832 allocated slots

Useful KV dataTail waste

Last 8 blocks shown · 16 tokens per block

Explore this route

Ordered by prerequisites. Routes share some foundation guides.

THE LIBRARY

A closer look at every idea.

Explanations, interactive diagrams, and exercises in every guide. Explore a topic or follow the reading route.

25 guides

Model Foundations

4 guides

Understand what models predict, what context can hold, and where evidence and permissions matter.

Foundation10 min read
How Language Models Generate Text
Follow a prompt from tokens to predictions, and learn what a fluent answer can—and cannot—establish.
Diagram + practiceRead guide
Foundation10 min read
Prompts, Tokens, and Context
Build a clear request, budget its context, and handle missing information before asking for an answer.
Diagram + practiceRead guide
Foundation9 min read
Why Fluent Answers Can Be Wrong
Separate missing evidence, stale knowledge, retrieval errors, and unsupported claims so you can choose the right control.
Diagram + practiceRead guide
Foundation11 min read
From Model Output to Controlled Action
Turn a proposed tool call into a bounded operation with explicit permissions, approval, and recovery.
Diagram + practiceRead guide

Reliable AI Systems

7 guides

Give each component a clear responsibility. Validate outputs, manage state, and recover safely.

Core concepts19 min read
Designing an AI System
Map the responsibilities of models, tools, data, and application code before connecting them into a system.
Diagram + practiceRead guide
Core concepts23 min read
Reliable Pipelines and Validation
Design stage contracts that catch malformed data, unsupported interpretations, and broken business rules.
Diagram + practiceRead guide
Core concepts21 min read
Memory, State, and Knowledge
Give conversation context, task state, evidence, and durable records clear owners and lifetimes.
Diagram + practiceRead guide
Core concepts22 min read
Choosing Between Workflows and Agents
Choose where a model may decide, which tools it may use, and which actions still require human review.
Diagram + practiceRead guide
Core concepts25 min read
Evaluating AI Systems
Build a versioned evaluation set and a regression gate that can reveal where a system starts to fail.
Diagram + practiceRead guide
Core concepts20 min read
Identity and Authorization for Agents
Keep delegated authority and approval outside model-generated text, and test the boundary with adversarial inputs.
Diagram + practiceRead guide
Hands-on25 min read
Agent Loops and Tool Reliability
Bound a tool loop, distinguish failure from uncertainty, and recover without duplicating external effects.
Diagram + practiceRead guide

Retrieval & Evidence Systems

4 guides

Find relevant sources, reconstruct documents, and connect each answer to evidence you can inspect.

Core concepts29 min read
Grounding Answers with RAG
Retrieve permitted evidence, preserve its sources, and evaluate retrieval separately from the final answer.
Diagram + practiceRead guide
Advanced25 min read
Hybrid Retrieval and Evidence Packets
Combine lexical and dense rankings, then assemble a bounded evidence packet with explicit freshness rules.
Diagram + practiceRead guide
Core concepts26 min read
Reading Documents Beyond OCR
Reconstruct reading order, tables, and source relationships so a document becomes traceable evidence.
Diagram + practiceRead guide
Advanced26 min read
Grounding Multimodal Evidence
Connect visual claims to source pages and regions, and recognize what an image cannot establish.
Diagram + practiceRead guide

Applied Architecture Labs

2 guides

Bring the parts together in support and travel workflows with explicit boundaries and failure cases.

Hands-on15 min read
Lab: A Reliable Support Workflow
Resolve a support case with evidence, account permissions, clear escalation, and verified outcomes.
Diagram + practiceRead guide
Advanced37 min read
Lab: Build a Governed Travel Copilot
Bring evidence, state, approvals, and recovery together in an inspectable travel-planning system.
Diagram + practiceRead guide

Inference Systems

8 guides

Follow a request through scheduling, memory, and GPU placement. Measure capacity under real constraints.

Foundation11 min read
Inside an LLM Inference Request
Trace a request through prefill, decode, caching, and streaming—and reconcile its timing and memory costs.
Diagram + practiceRead guide
Core concepts12 min read
Continuous Batching
Explore token-budget scheduling, chunked prefill, preemption, and backpressure in a small runnable simulation.
Diagram + practiceRead guide
Core concepts11 min read
Managing the KV Cache
Calculate memory per token, allocate complete blocks, and distinguish snapshot capacity from room to finish.
Diagram + practiceRead guide
Core concepts14 min read
Distributing a Model Across GPUs
Compare data, tensor, pipeline, and context parallelism using memory, communication, and failure boundaries.
Diagram + practiceRead guide
Hands-on11 min read
Prefix Caching and Request Routing
Balance prefix reuse against queue load, compatibility, and tenant isolation when routing inference requests.
Diagram + practiceRead guide
Advanced12 min read
Separating Prefill and Decode
Decide when phase separation is worth the cost of KV transfer, extra coordination, and a more complex deployment.
Diagram + practiceRead guide
Advanced13 min read
Deploying Mixture-of-Experts Models
Separate model routing from expert placement, and calculate the memory and communication each layout needs.
Diagram + practiceRead guide
Hands-on20 min read
Benchmarking Inference Capacity
Measure useful capacity with a frozen workload, per-class latency goals, queue evidence, and failure tests.
Diagram + practiceRead guide