AI

Explorations in applied machine intelligence for real-world sensing, classification, and play.

AI-000

AI Zoo

A growing in-browser museum of generative adversarial networks with live demos that run entirely on CPU power.

  • Every exhibit ships its full generator logic inline so guests can study and remix the models.
  • Searchable catalog designed to host dozens of GAN experiments with preview thumbnails.
  • Launch lineup features lightweight handwriting and chromatic pattern generators.

Enter the exhibit hall →

AI-001

Foot Traffic Intelligence

Multi-modal counting that fuses Wi-Fi probing, camera feeds, and door sensors to quantify movement across venues in near real time.

  • On-device processing keeps raw imagery private while streaming anonymized vectors to the edge.
  • Forecast module predicts surges so staff can stage resources before congestion starts.
  • Output dashboards plug into the analytics stack already deployed in the lab.

Open project dossier →

AI-002

Mycology Classification

A fine-tuned vision transformer that learns from herbarium scans, macro photography, and spore print imagery to identify fungal species.

  • Active learning loop requests new labels for ambiguous samples gathered in the field.
  • Outputs include toxicity risk bands, look-alike warnings, and habitat metadata.
  • Packaging targets a lightweight mobile inference bundle for offline use in forests.

Open project dossier →

AI-003

Phone Agent

Voice-first assistant that triages inbound calls, schedules follow-ups, and writes CRM notes using speech-to-text, LLM reasoning, and calendar sync.

  • Conversational state machine guards tone, compliance, and fallback routing to humans.
  • Integrates with Twilio SIP endpoints and the lab’s analytics warehouse.
  • Fine-tunes prompts from anonymized transcripts to raise resolution rates.

Open project dossier →

AI-004

Demo Arcade Intelligence

Competitive arcade sandbox where reinforcement learners iterate strategies while a player can drop in at any time to challenge the live leaderboard.

  • Best-performing agent trains continuously, publishing policy notes after each improvement.
  • Human mode loads the canonical game without assists so guests can attempt to dethrone the agent.
  • Results feed a shared scoreboard with telemetry explaining how each run was achieved.

Open project dossier →

AI-005

Mushroom Forecasting Alerts

Probabilistic growth model that ingests NOAA weather feeds, soil sensors, and phenology logs to forecast which species will fruit within a radius.

  • Users set distance thresholds and species lists; alerts trigger when humidity and heat align with a species’ fruiting profile.
  • Spatial tiles combine habitat suitability with recent mycology classification sightings.
  • Designed as a “weather-style” notification system that surfaces safe-foraging windows.

Open project dossier →