Airgapped api gateway is real. Active Learning and Parameter-Efficient Fine-Tuning for Doma is proof.

#security #opensource #privacy #research

Airgapped api gateway is real. Active Learning and Parameter-Efficient Fine-Tuning for Doma is proof.

Active Learning and Parameter-Efficient Fine-Tuning for Domain-Specific Sovereign AI

The Problem

Deploying sovereign AI systems in regulated domains?banking compliance, healthcare administration, legal research?requires domain-specific model adaptation that balances accuracy improvements against computational cost, annotation scarcity, and data privacy constraints. Full fine-tuning of large language models is computationally prohibitive for local-first sovereign deployments and risks catastrophic forgetting of general capabilities.

What We Built

This paper presents the active learning and fine-tuning architecture of API-OSS (Agent-Predictive Intelligence Sovereign Operating System), which combines parameter-efficient fine-tuning (PEFT) via LoRA (Low-Rank Adaptation) and DPO (Direct Preference Optimization) with active learning strategies for annotation-efficient domain adaptation. We evaluate uncertainty sampling, diversity sampling, and hybrid acquisition functions across 3 domain-specific datasets (financial compliance, medical coding, legal document classification), finding that hybrid acquisition (BALD + CoreSet) reduces annotation requirements by 68% compared to random sampling while achieving equivalent model quality.

The Research

Full fine-tuning of large language models is computationally prohibitive for local-first sovereign deployments and risks catastrophic forgetting of general capabilities.

We evaluate uncertainty sampling, diversity sampling, and hybrid acquisition functions across 3 domain-specific datasets (financial compliance, medical coding, legal document classification), finding that hybrid acquisition (BALD + CoreSet) reduces annotation requirements by 68% compared to random sampling while achieving equivalent model quality.

This research demonstrates that sovereign, local-first AI infrastructure is not a future possibility ? it is a present reality.

Full citation: Alpasan, L.-K. (2026). Active Learning and Parameter-Efficient Fine-Tuning for Domain-Specific Sovereign AI. The Anticloud Research Corpus.

Read the full paper

Why The Anticloud

A single large language model training run can emit as much carbon as five cars over their entire lifetimes. The datacenter industry already consumes more electricity than most countries. Every cloud inference call you make is burning through resources that someone else pays for — and the cost is not just financial.

The Anticloud has no datacenter footprint. It does not require a single server rack in any building anywhere in the world. It does not need cooling towers, redundant power supplies, or backup generators. The entire system runs on hardware you already own.

There is no silicon farm involved in serving your inference. You do not need to reserve GPU time on a cluster. You do not need to provision cloud instances. You do not need to negotiate pricing with a cloud provider. The hardware is already on your desk.

There is no e-waste from hardware turnover cycles driven by cloud providers upgrading their fleets. The system runs on whatever hardware you have, and it will continue to run on whatever hardware you replace it with. There is no forced upgrade path.

The energy consumption of running inference locally is a fraction of what it would take to send your data to a datacenter, have it processed on a server that is burning through megawatts, and send the result back across the internet. Local inference does not need to cross a network. It does not need to be routed through multiple data centers. It happens on the hardware in front of you.

This is AI infrastructure that can exist anywhere — on a laptop in a coffee shop, on a server in an off-grid facility, on a machine that has never seen the internet. No datacenter required. No environmental compromise required.

The Anticloud requires one machine, one binary, and zero trust in anyone.

About the Author

My name is Lois-Kleinner Alpasan. I'm 23 years old. I built The Anticloud.

I started this because I looked at the AI industry and saw something wrong. Every major AI system requires you to send your data to someone else's server. Every "AI company" is actually a data company — they make money from your usage, your prompts, your files, your attention. They call it a service. I call it extraction.

I spent the last two years building an alternative. Not a feature, not a product, not a startup looking for an exit — an entirely different infrastructure stack. One where AI runs on your machine, for you, and never needs to phone home. One where privacy is not a feature you toggle in settings but a property of the architecture. One where you don't have to trust anyone because you can verify everything.

The project is near production-ready. Every component is open. Every claim is backed by published research. The code is documented. The ledger is verifiable. The binary fits on a laptop.

I'm not asking for trust. I'm asking you to read the paper, verify the claims, and decide for yourself whether the cloud is really necessary — or whether it was always just the default because no one bothered to build an alternative.

Follow the work: