Ezora
AI Lab
Working demos of what we build
Six small models running live in your browser on sample data. Ask, generate, forecast, classify, delegate and monitor: the same building blocks we ship for clients, at demo scale.
Runs in your browserSample data onlyNothing you type leaves this page
Ask the sample handbook
Answer
Pick a question above or type your own.
Real retrieval: TF-IDF vectors, cosine similarity and BM25 over 12 passages from a fictional company's handbook. The answer is assembled from the best matching sentences; in client systems an LLM writes it from the same passages.
Vector space
Top passages
Ranked passages appear here.
A tiny language model, trained in your browser
Output
A word-level trigram model built from a short text when this page loads. It runs the same loop as a large model: score every possible next token, apply temperature, sample one, repeat. Large models do it with billions of learned weights instead of counts.
Next token: candidates
Vocabulary0distinct tokens
Trigrams learned0from the training text
Tokens generated0
Perplexity-lower is more certain
Weekly demand, sample stores
HistoryBacktest on the last 12 weeksForecast80% range
Holt-Winters exponential smoothing (level, trend and yearly season) fitted in your browser by searching the smoothing parameters. It is tested on 12 weeks it never saw before it forecasts. The range widens with the horizon because the future is uncertain.
Backtest error (MAPE)-on 12 unseen weeks
Next week-
Trend per week-units
Season peak-week of year
Fitted model
Alpha, beta and gamma control how fast the level, trend and season react to new data. They are chosen by grid search on the training weeks.
Route a support ticket
Decision
The routing decision appears here.
Why: the words that moved the decision
Probability per team
Training tickets04 teams
Cross-validated accuracy-5-fold, on sample data
Multinomial naive Bayes trained on sample tickets when this page loads. Each word adds evidence for a team; the chips show the strongest evidence. Production classifiers use embeddings or a fine-tuned model, with the same kind of explanation.
Give the agent a goal
The agent's plan, tool calls and results appear here, step by step.
Tools the agent can call
The tools are real functions over a sample sales dataset (five fictional stores, twelve weeks). The planner here is rule-based; in production an LLM plans with the same tools, a step budget and a human approval before anything is sent.
Requests per minute-live
Latency p95-milliseconds
Accuracy (rolling)-labelled sample
Feature drift (PSI)-basket size
Latency p95 and anomalies
p95anomaly (z > 3)
Accuracy
rolling accuracyalert threshold
Alerts
Drift: basket size, reference vs live
Want one of these on your own data?
We scope it with you, build it, and run it in your cloud or ours.