Animal communication
Vocal & behavioral sequences
Explore recurrent motifs and similarity structure in described calls, bursts, and temporal sequences.
Experimental ML · Pattern discovery
Hidden structure in real-world
signals — made legible
Modex is an experimental machine learning platform for discovering hidden patterns in real-world systems. Ingest signals, embed them, compare relationships, cluster similar structures, and generate early hypotheses — across domains from bioacoustics to fluids, materials, and unknown signal spaces.
Core flow
Signal → embedding → similarity → clusters → hypotheses
Modex is built as a research-facing experimentation engine — not a black-box decision product. The console runs ingestion, deterministic fallback embeddings when no external keys are present, similarity against stored signals when a database is configured, and local structural clustering for offline demos.
1
Ingest
Bring text-encoded signal descriptions and typed experiment contexts into the pipeline.
2
Embed
Project signals into a fixed embedding space — deterministic offline or upgradable to hosted models later.
3
Discover
Score similarity, group related structures, and surface concise hypotheses for further experiments.
Thesis
Pattern discovery, not opinion scoring
Many real-world systems emit signals that repeat, rhyme, or cluster in ways humans barely notice. Modex focuses on representation and comparison — helping teams explore structure before committing to a specific scientific story or product narrative.
Research tracks
Fluid / water dynamics
Flow & interaction signatures
Encode narratives of turbulence, waves, and coupling events to compare patterns across regimes.
Material interactions
Contact & response patterns
Capture how materials meet, wear, resonate, or transition — as structured text signals for embedding.
Unknown signals
Discovery-first workflows
When the domain is immature, Modex prioritizes clustering and hypotheses you can test — not premature labels.
How it works
An experiment console for signals
Each run walks the same scientific skeleton: represent the signal, compare it to peers, aggregate clusters, and phrase an early hypothesis string you can refine with domain experiments.
01
Typed experiments
Choose animal, fluid, material, sequence, or custom modes so embeddings and clusters inherit the right framing.
02
Similarity graph
When Supabase is configured, compare against stored signals; otherwise the UI still renders local similarity playgrounds.
03
Clusters
Group nearby embeddings with transparent thresholds — tuned lower for offline deterministic embeddings.
04
Hypothesis line
Surface a concise interpretation tier (strong / possible / weak pattern) to steer your next measurement — not to replace lab validation.
Technology
Composable ML primitives
Next.js API routes wrap a thin orchestration layer (`modex-core`) over embeddings, cosine similarity, clustering, and discovery helpers. Swap in richer models when you are ready — the pipeline stays the same.
Investor narrative
Modex sits upstream of vertical SaaS: cross-domain discovery infrastructure with a credible experiment UX and a path toward richer modalities (audio, time-series) when datasets attach.
Next step
Open the experiment console
Run a typed experiment, inspect embeddings and clusters, and export the session log from your browser — no auth required for the MVP shell.