BY PERGENCE
An intelligence layer that lives inside the OS.
Pri — the Persistent Reinforcement Interface — is an agent built by Pergence. It learns how a specific machine is used and manages its resources autonomously: offline reinforcement learning that runs for years. Old hardware made responsive again. Nothing leaves the device.
It stops the machine from repeating its mistakes.
Every machine develops habits — the same pressure at the same hours, the same slowdowns under the same load. Pri watches those patterns, remembers them, and manages resources ahead of them. Autonomously, through a safety separation layer that keeps every action deliberate and reversible.
PRI · LEARNING LAYER
OFFLINE · A FEW MB
TELEMETRY UP · TUNING DOWN · THE OS IS NEVER MODIFIED
We solve resource contention at the root.
I plan on replacing static, universal OS thresholds (like earlyoom) with per-machine learning using native Linux kernel-level actuators like cgroups and PSI. We solve resource contention at the root. I have trained the RL engine for almost 2 months now and I'm getting closer to the benchmark.
System design
PRI is a single threaded C11 daemon that runs a closed control loop on the host it manages. A learning core proposes one intervention per cycle; an independent safety layer masks, gates or vetoes it; an execution layer applies it through standard Linux primitives; and every change is registered for automatic reversal.
PRI DAEMON · SINGLE-THREADED C11
memory composition · swap · cpu · load · per-process cpu · thermal
setpriority · sched_setscheduler (SCHED_BATCH) · SIGSTOP/SIGCONT · page cache reclaim
ONE INTERVENTION PER CYCLE · EVERY CHANGE IS A LEASE · SHADOW MODE STOPS BEFORE STAGE 04
Subsystems
Operating modes
Full decision loop, no system calls.
Inference only, learning and writes disabled, for validated deployment.
Full closed loop.
Pri's core architecture is patent pending with the Indian Patent Office. Application No. 202621085200 is published on IP India and awaiting examination. We follow the filing guidelines, so detailed specifications are published here only as they clear.
No action without graduation.
Every deployment starts silent. Pri earns autonomy on your machines with evidence from your machines — and every rung of the ladder is reversible.
Pri never collects process names, file paths, or user identifiers — they are excluded by design, not by policy.
Observes and learns. Takes no actions. Proves it understands the machine before it's allowed to touch it.
Recommends actions with its reasoning. A human approves every change. Pri earns trust decision by decision.
Manages resources autonomously — always inside the safety separation layer, always deliberate, always reversible.
The economic engine.
For deep-tech infrastructure, here are the market mechanics and the economic engine behind the project.
The Market Size & Immediate ROI
The buyer is whoever manages fleets of constrained Linux machines and bleeds money on tuning labor or hardware churn.
The Linux server OS market segment alone is valued at $17.2 billion in 2026, with enterprises driving the majority of demand. MSPs charge flat fees; every L1 "slow server" or contention ticket destroys their margin. By autonomously squashing PSI spikes, we reduce ticket volume and deliver immediate, measurable margin expansion.
In dense Linux environments, efficiency is capital. If our per-machine learning squeezes 10–15% more density out of a server farm or extends hardware lifecycles, we become a multi-million dollar CapEx deferral line item.
The Strategic Acquisition Potential
Once deployed at scale, this fleet-learning flywheel becomes a highly strategic asset. We are building the exact offline architecture that enterprise incumbents are desperate for:
Providers need verifiable, offline-only agent architectures to take automated actions on Linux endpoints without triggering compliance alarms.
Platforms need to transition from passively monitoring infrastructure to actively managing resource contention at the Linux kernel level.
The biggest barrier to deploying AI in the enterprise is InfoSec — the architecture is built on auditable autonomy that is local-only, deterministic, with no network exfiltration. Every action is backed by signed policies and an instant kill switch. We turn compliance from a roadblock into our primary sales weapon.
How I Build: Evidence-First Engineering for Autonomous Systems
When a system makes its own decisions, the hardest problem isn't building it — it's knowing whether it actually works. Autonomous systems produce results that look convincing and are wrong: a metric moves favorably for an unrelated reason, a policy appears to improve when it's really just overfitting the test, a safety property seems intact because a cleanup step masks its failure. Intuition is the least trustworthy instrument you have in this setting, and it fails quietly.
I develop against a discipline built to surface those failures early, borrowed from how empirical science guards against self-deception:
Success criteria are specified in full and sealed before any evaluation is run. Once sealed, they can't be edited. This removes the most common source of false results — quietly moving the goalposts after seeing the data, usually without even realizing you're doing it.
Every claim must point to a concrete, hash-locked artifact that can be regenerated. "It works" is not a statement; it's a pointer to a specific, verifiable run. Anything that can't be reproduced isn't claimed.
Results are graded against a fixed taxonomy of verdicts rather than a binary pass/fail. A result has to survive classification — is this a causal effect, a correlation, a null result, an instrumentation artifact? Forcing every outcome through that lens prevents weak evidence from being promoted to a strong claim.
The default assumption is that a good-looking result is wrong until the evidence trail proves otherwise. This has repeatedly caught real errors that felt correct at the time — including a trusted benchmark that had run against the wrong stored artifact, and a safety mechanism that had been silently non-functional for months, invisible because a shutdown path always cleaned up before anyone inspected it. Both were found through the evidence trail, not through intuition.
Persistent intelligence.
Pergence builds software that learns and stays. Pri — the Persistent Reinforcement Interface — is our first agent, built on a simple belief: the machines you already own have years of life left in them. They just need software that pays attention.
"Operating systems still manage resources with fixed rules, decades after machines stopped being used in fixed ways. Pri replaces those rules with learning. I've been building it for the past year, and it's in the last phase of training before it's fleet-deployment ready."
Old hardware.
New intelligence.
Pri is in development and rolling out to its first fleets soon. Add your name and we'll reach out as early access opens.