Its real source was being fetched and displayed on the public Space.
Test Bench now only references the public victor/gemma-avatar Space.
Add Preview/Code toggle: real source files from the reference targets
Claude
Unverified signature
0f54cd7
For a pipeline we didn't design and have no connectivity history on,
the README summary in the Project tab isn't enough to actually work on
it. The new CODE tab (fixed to the right edge) fetches real entrypoint
files (package.json, Dockerfile, index.*, etc.) from both reference
targets via the GitHub/HF Space APIs and shows their real content in a
file browser, so surgery or an autopsy on someone else's pipeline is
actually possible instead of guessing from a description.
Mark which stages are actually responsible for perceived realism
Claude
Unverified signature
6b5c34f
Adds a realismCritical flag to PipelineNode, set on Avatar/TTS/Video
Quality (the stages whose own output quality directly drives how
human this looks/sounds) but not Vision/ASR/LLM/Orchestration (affect
correctness/latency, not realism itself). Shown as a fuchsia REALISM
badge in the full graph and a compact eye icon in the Test Bench grid.
Also added to the pipeline-generation schema so grounded/generated
pipelines can flag it too.
Give Jessica a Play button: real LLM line + real speech, audio-reactive puppet
Claude
Unverified signature
6553b8c
/api/avatar-speak asks the same NIM/Anthropic model already wired up
to generate one short first-person line as the avatar. The browser's
real speech synthesis speaks it, and its actual boundary/playback
events drive a mouth-open pulse and periodic blink overlaid on
Jessica's photo. This is real audio driving real motion — not neural
lip-sync, which would need a GPU-hosted video model instead.
Add Test Bench: grid-inspect pipeline stages against real reference implementations
Claude
Unverified signature
e1e33d1
Grid layout of compact stage cards next to an Artifact panel with Project
(live-fetched HF Space + GitHub context), Console (scripted, honestly-labeled
simulated GPU test run under a 3-min budget), and Samantha tabs. DEEP DIVE
seeds Samantha with a compare-against-reference question. Extracted repo/space
fetch helpers into src/lib/repo-fetch.ts for reuse with generate-pipeline.
Switch Samantha to NVIDIA NIM (Nemotron 3 Ultra) as primary model
Claude
Unverified signature
fc3642a
A Claude subscription doesn't include Anthropic API access, so both
API routes now go through NGC_API_KEY -> NVIDIA's hosted NIM catalog
first, falling back to ANTHROPIC_API_KEY if that's set instead.
Verified directly against the real endpoint before wiring this in:
integrate.api.nvidia.com speaks OpenAI's /v1/chat/completions format,
not Anthropic's /v1/messages (that path 404s on the hosted service —
the Anthropic-compatible mode is apparently a self-hosted NIM
container feature, not available on the shared API). New
src/lib/llm-client.ts normalizes both backends behind one
callChatModel() interface so the two routes don't need to know which
one is active.
Fix pasted-source bug: actually generate a pipeline instead of relabeling the demo
Claude
Unverified signature
9dc9b9a
Previously, pasting a GitHub URL just slapped the URL on as a label
while showing the same static Beryl demo underneath. New
/api/generate-pipeline route fetches real repo metadata/README when
given a github.com URL and has Samantha synthesize a grounded
Pipeline via forced tool-use, marking every node confirmed/inferred
honestly. Falls back to the demo pipeline with a visible error if
generation fails (e.g. missing API key) instead of silently
mislabeling.
Also: node cards are larger with full (non-clamped) 200+ char
summaries covering who/what/when; data edges render in a vivid gold
with a glow; nodes animate in on a staggered mid-speed landing with
a brief gold sunburst flash; each node gets a gold DEEP DIVE button
that jumps to Samantha with a seeded question about that node.
Ground the expert panel as Samantha: 5-failure-mode audit + tiering
Claude
Unverified signature
11c5c3c
Every node now carries a tier (A/B/C) and per-failure-mode verdicts
(Uncanny Valley, End-of-Speech Pop, Cascaded Latency, VRAM Death,
VASA-1 Problem) instead of a blanket pass. Added a pipeline-level
audit banner for architecture-wide findings that don't belong to any
single node (the ASR->LLM->TTS->Avatar cascade is a literal Failure 3
regardless of per-stage latency, and several model names need
verification against real NVIDIA documentation). The chat panel is
now Samantha, whose system prompt carries the rubric verbatim and
defaults to skeptical verdicts over blind passes.
Next.js + React Flow app that turns a pasted pipeline into an
interactive circular node graph (deep sub-component expansion,
Local/NIM implementation toggles, live latency/VRAM metrics), plus
a Claude-powered side panel that explains model tradeoffs and
handoff contracts. Ships with the NVIDIA avatar pipeline as the
first working example.
