LABORATORY / WORKING INSTRUMENTS

Touch the model.

These are small tools built alongside the reports—not a project gallery. Each one makes a mechanism, assumption, or decision boundary easier to inspect.

On the bench

06 LIVE
06

INTERACTIVE EXPLAINER

Constrained Decoding

Watch a schema mask invalid tokens during generation and replay three malformed JSON failures.

IllustrativeLLMsStructured output
→
05

EXPERIMENT LAB

Medication Resolution

Read, replay, and inspect a frozen medication-search benchmark where always-on medical-LLM reranking made retrieval worse.

MeasuredReplayLive GPU
→
04

INTERACTIVE EXPLAINER

Prefill, Decode, and the KV Cache

Derive the cache behind autoregressive generation, then scale its memory cost by context length and active sequences.

CapacityLLMsCalculator
→
03

INTERACTIVE EXPLAINER

How an LLM Generates One Token

Follow a prompt through embeddings, attention, and MLP to the selection of Paris, one concept at a time.

ConceptualLLMsStep by step
→
02

LIVE SIMULATION

Arrival-Model Simulator

Hold the server fixed and watch closed-loop and open-loop request generation diverge during the same stall.

IllustrativeQueueingMethodology
→
01

LIVE MODEL

Mix-Axis Explorer

Change the share of writes in a Cassandra workload and inspect the capacity model against three measured runs.

Measured pointsModelled curveOperator decision
→

An interaction earns its place only when it reveals something that static prose or a chart cannot reveal as clearly.

Every instrument identifies what is measured, what is modelled, and where the claim stops. Read the standing Methods →