We do the research so the product can be trusted.
Our mission is to make technology accessible to everyone. That starts by fixing what's broken in today's AI: applications that fail silently, models that only run in someone else's cloud, and systems nobody is accountable for. We take what this research proves and put it into our products, to make them more reliable, more local, and more accountable.
This is what the research is for.
Every result on this page feeds one goal: answers you can hold to account. The model's own output is never deterministic — what's deterministic is the layer that checks it before a human ever sees the result.
- 1
Grounded
Answers are drawn from your own documents and data — retrieval, not memory.
- 2
Checked
Deterministic rules, lookups, and reconciliation with tolerances check the result before anyone sees it.
- 3
Held for a human
Anything uncertain, disputed, or out of policy waits for a person instead of shipping unapproved.
- 4
Recorded
An audit-ready record of what was decided, and why — so it can be checked later, not just trusted now.
What every product is held to
Our R&D program exists to earn these three commitments, in this order — not as slogans, but as research we hold ourselves to with measured, honest results.
AI applications that fail silently — a confident, wrong answer looks identical to a correct one, and nobody is told the difference.
We make AI-native applications reliable at the core. Our verification-anchored research cut silent errors from 36% to 4% in our research.
“Fails loudly, never silently.”
- Orchestrator Studio
A second-opinion reviewer model holds a run for humans whenever it disagrees; confidence-gated review queues catch what a single model would have missed.
- Vexilo
Tested against a real persona task set with zero silent wrong answers.
AI that only runs in someone else's cloud — so the teams with the most sensitive data are locked out of using it at all.
We make AI run on consumer-grade hardware people already own. Our products run on everyday hardware — a laptop GPU, an Apple Silicon Mac, or one on-prem workstation.
“Runs on hardware you own.”
- Orchestrator Studio
Runs on your hardware, from a laptop GPU up to a shared team server.
- CrestBid AI
Runs fully local on GPU or Apple Silicon — or points at a model you choose.
- Vexilo
A local 27B model via llama.cpp. No account, no sync, no telemetry.
AI nobody is accountable for — it acts, and no human ever confirmed the action was right.
We make AI accountable to its human owners: systems deeply aligned with humans, with human-in-the-loop confirmation. Confidence-gated review queues, approval-only learning, cell-level diffs, and audit trails by default.
“A human confirms when it matters.”
- Vexilo
One Ctrl+Z undoes a whole reply. It shows a diff of every cell it touched and asks before destructive actions.
- Orchestrator Studio
Policy-based HITL review queue, durable resume, full run history.
- CrestBid AI
Cognitive flywheel: only approved answers teach the system. It never fine-tunes weights.
Verification-Anchored Federation (VAF)
The research program behind "reliable at the core." A router hands each request to a small specialist model, and nothing ships until it has been verified and calibrated.
- 1
Input
A request arrives — a question, a document, a spreadsheet task.
- 2
Router
The router reads it and decides which kind of work this is.
- 3
Four specialists, one backbone
Four LoRA specialists — tool, RAG, soft-skill, and extract — sit on a single shared 1.2B-parameter backbone (LFM2-1.2B). No 7–8B-scale model required.
- 4
Tiered verifier
Before an answer ships, a tiered verifier checks it. Self-reported model confidence is discarded — it isn't trusted as evidence.
- 5
Conformal abstention gate
Calibration comes from conformal prediction, tuned for 95% coverage — not from the model grading its own work.
- 6
Out-of-distribution rejection
Requests that fall outside what the system was calibrated on are rejected rather than guessed at.
- 7
Outputs
Every request ends one of three ways: an answer, an abstention, or an escalation to a human.
What the research produced
Measured on our own hardware, not projected from a slide.
36% → 4%
Silent-error rate, down from an unverified 7–8B baseline, on our internal benchmark.
2.2×
Tokens per second vs. baseline on the same host (19.5 vs. 9.0 tok/s).
4.5 GB
Active memory, vs. 13 GB for the baseline.
Q4_K_M
No measurable quality loss vs. F16 in our quantisation sweep.
We are driving abstention down without raising silent errors — we do not claim this is solved.
How we measure
We measure honestly, including where the result falls short of our target. A model's own confidence in its answer is never treated as evidence — every number here is checked against held-out data, not the model's opinion of itself.
What didn't work
We publish what doesn't work, not only what does. A negative result, measured honestly, is still research.
We tried a training-free "context compiler" — turning a document directly into a LoRA adapter, with no training step at all. We tested it four different ways, and it didn't clear the bar we set for it in any of them.
The reason lines up with a simple trilemma: quality, zero training, and zero context tokens don't come as a set of three. A system can have any two.
Open for collaboration
Our mission is to make technology accessible to everyone — that's easier with partners who share the problem.
Universities & labs
Joint work on verification, conformal calibration, and evaluation methodology — including the negative results we publish internally.
Email usHardware & model partners
We build for consumer-grade hardware. If you make GPUs, NPUs, or open-weight models, there's a shared interest in what runs well where.
Email usDesign partners (pilots)
Teams running Orchestrator Studio or Vexilo pilots shape what ships next. Early access, direct line to the people who built it.
Email usResearch you can hold the product to.
Talk to the team doing the research, or see what it has already shipped.