LLM X-Ray.

No pre-baked demos — every visual is driven by the user's own prompt, captured from a live PyTorch forward pass and streamed to the browser as it happens. The model answers the question (with an optional <think> reasoning trace first), so what's on screen is the model actually thinking, not a canned autocomplete animation.

Every view reads from the same captured pass — nothing is recomputed per tab.

The core (xray_engine.py) registers forward hooks on every decoder block to capture the residual stream and attention weights, then projects each layer's hidden state to a top-k prediction — a batched logit lens: all layers' last-position vectors through final_norm + lm_head in one matmul, not one pass per layer.

Architecture-specific module paths are resolved by a ModelAdapter, so swapping MODEL_NAME to gpt2 or a different Qwen3 size needs no engine changes. Attention rollout — the cross-layer 'what actually drove this token' attribution — is computed client-side, lazily, only for the selected step.

It runs on CPU at roughly 0.13s/token, but is modern and instruction-tuned enough to actually answer — including an optional reasoning trace — rather than just autocomplete. Its logit lens sharpens into coherent words across 28 layers, where GPT-2 small often stayed noisy the whole way through.