About the practice
Chels.ai is a sovereign AI infrastructure practice: we design, build, fine-tune, host, and maintain on-premise deployments of flagship open-weight models for organizations that can’t send their data to someone else’s cloud — law firms, clinics, credit unions, and manufacturers across Canada.law firms, health networks, banks, and defense suppliers across the United States.
Why this firm exists
This practice wasn’t founded on a trend deck. It was founded on four dated events — and on what they mean for every organization that depends on AI it doesn’t own.
The US government ordered the leading American lab to block foreign access to its most advanced models within 48 hours, with no avenue for appeal. Organizations outside the US that had built workflows on rented frontier intelligence lost it overnight.
Two days later, Zhipu AI released GLM-5.2 — 744B parameters, 1M-token context — under the fully permissive MIT license, with no regional restrictions. A frontier-adjacent model anyone could download, audit, and run.
Moonshot AI released Kimi K3 — 2.8 trillion parameters, 1M-token context, native vision — beating the proprietary flagships of the previous generation on coding and agentic benchmarks at launch. For eleven days access ran only through Moonshot’s own servers — not an option for regulated data.
Moonshot published the 1.56 TB Kimi K3 checkpoint. A frontier-tier model became a static file you can checksum, air-gap, fine-tune, and own — with no US or Chinese provider anywhere in the chain. The organizations that benefited on day one were the ones whose infrastructure was already racked.
GLM-5.3-Flash (MIT, multimodal, single-node), Qwen3.8-Flash-Next, GLM-5.3, Tencent Hy4-preview (Apache 2.0) and DeepSeek V4-Flash-Vision (MIT) all shipped weights in one week. The capability premium for rented APIs collapsed; licensing became the differentiator. Getting regulated organizations ready for exactly this moment is the job this practice was built to do.
Operating principles
Residency clauses in a vendor contract are promises; data that physically never leaves your infrastructure is proof. Every deployment is designed to support our customers’ obligations — PIPEDA, Québec Law 25, PHIPA, OSFI Guideline E-23, law society confidentiality rulesHIPAA, GLBA, CMMC/ITAR, and state privacy acts— with zero data egress as the default posture, not an add-on.
Full-precision Kimi K3 is supercomputer-class, and we say so — most organizations are better served by a single-rack GLM-5.2 deployment. We publish real hardware requirements, quantization trade-offs, and cost curves, and the sovereignty assessment tells you which tier your workloads actually justify — whether or not you build with us.
We have no model to sell you. Kimi K3, GLM-5.2, DeepSeek, Qwen, Llama, Mistral — the bench changes monthly, and your infrastructure is built to outlast every entry on it. Where procurement policy requires Western-origin weights, we deploy them. Sovereignty means model origin is a governance choice you get to make.
Practice areas
Sovereign AI fails when it’s treated as an IT project. We run it as a complete practice — four disciplines under one engagement.
GPU cluster design and procurement sized to your models and concurrency — on premises, in-country colocation, or fully air-gapped.
Parameter-efficient tuning on your corpus, trained inside your environment and evaluated against a harness built from your real work.
Production in-country inference with SSO, role-based access, audit logging, and observability. Zero third-party API calls.
Open-weight releases arrive monthly. We evaluate them against your workloads, migrate fine-tunes, and upgrade serving stacks on your schedule.
Why we publish
Everything we learn deploying open-weight models — real hardware requirements, quantization trade-offs, compliance architectures, cost breakevens — gets written up and published in our guides and deployment blueprints. Regulated buyers deserve to check our reasoning before they trust us with their infrastructure, and honest technical writing is how a young practice earns that trust. If our numbers are wrong, tell us — we’ll correct them and date the correction.
The practice
Chels.ai’s practice combines AI-infrastructure engineering, model evaluation and fine-tuning, and regulated-industry compliance architecture under one roof — operating from Toronto, serving clients Canada-wideacross the United States. Engagements are led end-to-end by the people who design the architecture, not handed off to a rotating bench.
Team profiles are being added as the practice formalizes — for now, the fastest way to evaluate us is to read the published work or get us on a call.
Two weeks. We map your data obligations and workloads, size the hardware, and hand you a written architecture with a real cost model — whether or not you build with us.