tillGlyph translates plain-language questions into SQL, graph, and search queries — then a review ensemble elects one, or refuses. Only the elected query runs. Answers stream back live. Inference can stay on your infrastructure, so the question never has to leave your network.
The name: till — "before," and the work of cultivating a field — and glyph, the symbol. tillGlyph tills the symbols: it works the raw data layer before it hardens into glyphs on a chart, so answers stream up in the language you asked in.
A question produces several candidate queries. Specialist reviewers check schema fit, permissions, cost, and dialect — in parallel. An aggregator elects one query, or refuses. Only a passing verdict touches a backend.
Context assembled from your bound schema — not a guess.
Several candidates generated - query review will decide what candidates advance
Four lenses review in parallel.
Answers the question at the right grain.
Read-only, no sensitive columns exposed.
Candidate 2 is an unbounded scan. Dropped.
Candidate 3 uses the wrong query language. Dropped.
tillGlyph treats models as named, queryable streams — features in, predictions out. Start with simple rules; promote to trained models only when evaluation beats what is already live.
Rule-based heads on live data — thresholds, z-scores, moving averages. No training required.
Pre-trained forecast models score the same features without per-tenant training.
When you have enough labels, a small trained head can replace the baseline — but only if it wins on held-out evaluation.
tillGlyph picks the right query language for wherever your data actually lives.
Real-time dashboards, continuous queries, event-driven results.
Joins, aggregations, and analytics on structured data.
Traversal and exploration across connected records.
Log analysis, text search, and large aggregations.
Product newsletter
Product news, rollout notes, and when beta access opens. No waitlist queue — just the list.
Product updates only. Unsubscribe anytime.
tillGlyph