live learning demo · v0.0.6 · released 2026-06-29 ↗

acuity. it learns while the page loads.

upload a csv or arff. acuity learns the relationship between your features and your target; the kind of problem that normally needs a trained multi-layer network, and returns held-out quality, a shuffled-target control, calibrated regression intervals, and answers for rows you left blank. no pretrained model, no stored state: every upload is learned from scratch, live.

fresh fit per upload · shuffled control on every run · 90% conformal intervals
ⓘ first time here? what am i looking at?

this page is a live demonstration of acuity; optrenium.ai's proprietary learning engine. where a neural network trains for minutes to hours, the production acuity arc learns the same class of relationship in milliseconds to seconds on a cpu, while your request is in flight. the method stays server-side; what you get back is the evidence.

and it is honest by construction: every run silently re-runs itself on shuffled targets and shows you that control next to the result. if your data carries no signal, acuity says so; it does not invent answers. predictions for blank rows carry 90% split-conformal intervals calibrated on data the fit never touched.

your file is processed in memory and not retained: see "what happens to your file" below.

try it

no data? download a sample ↓

the cannonball challenge

don't take our word for anything; verify acuity with high-school physics. download the cannonball dataset ↓: 1,400 shots of launch speed + angle → distance flown, with 16 rows left blank at round numbers (10–40 m/s at 15°, 30°, 45°, 60°).

upload it, then check the predictions yourself: distance = speed² × sin(2 × angle) ÷ 9.81, or any projectile calculator. acuity has never seen the formula; it learns it from the rows, live, in a fraction of a second. the answers land within centimeters.

the format

  • csv or dense/sparse arff, utf-8, 384 MB max, header row or arff attributes required; excel/json/parquet are politely rejected
  • numeric features and low-cardinality categorical features are accepted; categorical columns are encoded deterministically
  • common missing feature markers are accepted: blank, ?, na, n/a, nan, null, none
  • the target is the first column named target, class, label, y, or outcome; otherwise the last column
  • numeric continuous targets run as regression; integer or text targets with 2–64 classes run as classification
  • leave the target empty on rows you want predicted; you keep the answers, we never see them
  • at least 80 labeled rows; up to 75,000 rows, 5,000 prediction rows, 4,096 raw feature columns, and 8,192 encoded features

what happens to your file

  1. your browser posts it; a 384 MB cap is enforced before a single byte is parsed
  2. it is read in memory only and strictly parsed; any contract violation returns a precise error and nothing else happens
  3. a fresh acuity model is solved on your rows alone; no pretraining, no other data, no stored state
  4. it is graded on a held-out slice it never fit, and re-run on shuffled targets; the control shown beside every result
  5. the dossier comes back; your file and the model are discarded with the response; nothing is written to disk, and logs carry paths and sizes, never contents
ⓘ glossary · hover any underlined term for its definition