SirusAI takes on AI engagements that other vendors quote and walk away from. We do not write proposals; we run a paid two-week probe on your data and return a written answer. If the probe says yes, we build. If it says no, we say so. Either way, you get a real number.
Six shapes account for nearly every project we run. The shape is decided in the scoping conversation, not by a price list. Pick the one that sounds like your room; we will tell you in the first call if it fits.
Defect classification on production lines, count and posture detection on a shop floor, PPE compliance at gates, gauge reading from analog dials. We have shipped four of these. Each is a model trained on the room it lives in, not a generic vision API.
Invoices on carbon paper, hand-marked forms, regional-language receipts, bills of lading scanned at angles. We extract structured fields with provenance, so your accounts team can audit a row back to a pixel. Built for the documents your generic OCR vendor refused.
Failure forecasting for irrigation pumps, demand forecasting for a regional retailer, churn risk for a subscription product. We refuse to forecast on synthetic data; if you can hand us two years of clean history, we can hand you a model with a calibrated confidence band.
Internal copilots over your policy manuals, technical docs, or compliance archives. Every answer cites the source line. We build small, fast, retrieval-grounded systems that work on a dataset of 50,000 documents without making things up.
Anomaly detection from a microphone or vibration sensor on a machine. Bird-call classification for a conservation project. Anything where the input is a waveform and the output is a label. Niche, but we have shipped two.
A retainer model where you bring us the question of the week and we return a written experiment by Friday. For teams that have a research budget but not the headcount. Minimum three months, capped at two questions a week.
A specific timeline. Specific exits. Each phase ends with a document we both sign before the next phase starts. There is no surprise at week eight—because we agreed at week zero what week eight would look like.
Three things, every time. Other consultancies hand you a slide deck and a Slack channel. We hand you a system, a method, and a contract that says you can extend both without us.
The model, packaged for your environment. Containerized, on-prem by default, optionally on your cloud of choice.
Everything needed to retrain, extend, and audit the system. The reason your engineer of record can run the second site without us.
Plain language. No vendor lock-in. The IP terms most customers ask for; the audit terms most regulators ask for.
We would rather lose a deal at the probe than win one at the demo. The first is a research decision. The second is a sales one.
Names redacted at the customer's request. Methodology, datasets, and unrounded numbers available on signed NDA. We will not show these in a sales deck.
Eight defect classes, three cameras, one shop floor. Replaced a two-person visual inspection station that could not keep up at line speed. Probe in two weeks; build in nine.
Forty-two field types, hand-marked carbon-paper originals, scanned at angle. Replaced a manual data-entry team of nine. Provenance baked into every extracted field.
Forty-two pumps, two years of vibration and current data. The model predicts failures two weeks before they happen with a calibrated confidence band the operations team trusts enough to act on.
We negotiate exact numbers against your room, but the shapes are these three. Most customers start with a probe, convert to a build, and keep the audit retainer. Some only need the first.
Two weeks. A written report with the metric we promised. The cheapest experiment you will run this quarter.
Eight to twelve weeks. A working system, a written method, and a clean hand-off. The default engagement.
One day a quarter. A re-run of the metrics on fresh data, a one-page review, a written drift flag. Optional, retained.
