FinnG is a local, privacy-first AI stack, model, retrieval, eventually custom hardware, that runs entirely on your own machine. Nothing rented. Nothing harvested. Built by the community, owned by the community, for the community.
It started with what I was reading and hearing.
People being laid off, sometimes in waves of thousands at a time, by companies that had spent the last year talking about how AI would augment workers rather than replace them. Big tech reporting record profits while the data centers powering all this AI consumed enough electricity and water to strain the communities they were built in. Energy bills rising for the people living near those data centers because the grid was being stretched. Subscriptions are creeping up and free tiers are shrinking. The pattern was hard to miss.
I'm not anti-AI. I think AI is genuinely useful, and I want to use it, and have been using it. What I don't trust is the current arrangement where using AI means sending my data to companies whose business model depends on that data, where I have to take their word that they won't use it or share it without my consent, where the whole industry is propped up by financing arrangements that look more fragile the closer you look. The market is booming, and I'm almost convinced the bubble will pop. The question is what's left standing when it does.
I want to build FinnG and I'm building FinnG because I want everyone to own their AI. Not rent it. Not subscribe to it. Own it in the way you own a laptop, calculator or a book. Run it on your own machine. Trust it because the architecture makes the trust verifiable, not because someone promises you in their privacy policy that they're being good. The data stays where it belongs: with you.
The reason I structured FinnG as a cooperative rather than a company is that I believe people are the power in building AI, owning AI, and shaping AI. Not investors, boards accountable to shareholders, but the people who actually contribute and use it. That's not a marketing slogan; it's a legal structure. AGPL-3.0 for software, CERN OHL-S v2 for hardware, both strongly reciprocal. Members vote on direction. The bylaws make acquisition impossible without unanimous member consent. The whole thing is designed so it can't be sold off later to whoever shows up with the largest check.
There are two things I won't compromise on, ever: cooperative ownership and privacy-first architecture. Everything else is open to discussion. Those two are the spine of the project. If I have to choose between speed and cooperative governance, I'll choose governance. If I have to choose between scale and local-first privacy, I'll choose privacy.
I know this is hard. I'm one person right now with two friends interested, a paused build, a job search in progress. Most projects like this fail. But I'd rather try than not try, and I'd rather build something that respects the people who use it than build something that doesn't.
If any of this resonates, this page tells you what's actually being built and how to help.
- Wirdani
The economics of cloud AI don't work not because AI doesn't work, but because centralized AI provision is structurally unprofitable. Inference costs eat the margins. Training costs grow 20× per generation. 97% of users don't pay. Hyperscalers take a cut. DeepSeek showed Chinese labs can deliver frontier-class models at a fraction of the cost.
Beyond economics, the structural risks compound. Privacy regulation is biting OpenAI has been investigated or banned under GDPR in four EU countries. The EU AI Act enters its enforcement phase this year. IP exposure is enormous Anthropic paid $1.5B to settle Bartz v. Anthropic for training on pirated books, and that was the cheap version. Circular financing Microsoft funds OpenAI which funds Microsoft. This means the whole arrangement wobbles together if any one piece slips.
FinnG isn't a bet that the bubble pops. It's a bet on something more boring and more durable: that a meaningful fraction of users; lawyers, doctors, retirees, anyone who works with sensitive material would prefer AI that runs on their own machine, owned by them, costing nothing per query, never sending their data anywhere. That market exists today and is not being served well by anyone.
FinnG Project is the parent initiative. FinnG AI is the flagship, a privacy-first local AI assistant for any work that can't be done with cloud AI. FinnG Retire and FinnG Transition are specialized variants built on the same core, focused on specific user needs.
A local-first AI assistant for any work that can't be done with cloud AI. Lawyers protecting client privilege, doctors handling PHI, accountants with sensitive financials, therapists with session notes, journalists protecting sources - all carry legal or ethical obligations that data not leave the device. FinnG AI is built for that work. Earlier prototype built on Illinois landlord-tenant law with ChromaDB and Mistral 7B; resuming when funding allows.
Beachhead market identified →A retirement calculator with a real niche; international workers. Models Social Security quarters of coverage, totalization agreements, NRA tax treatment, and 7 layers of risk. Runs locally in your browser. No data leaves the page. Available now under PRO+ tier.
finng-retire.pages.dev →A privacy-first conversational tool for people whose work is being reshaped by AI. Guided flow rather than dashboard. Designed to run locally with no career data ever leaving the device. Four-week build plan in progress.
Architecture in design →Structured as a Limited Cooperative Association. Contributors aren't donating to someone else's company. They hold a stake in what they help build. Voting rights, revenue share, and a legal guarantee against acquisition.
AGPL-3.0 for software. CERN OHL-S v2 for hardware. Both close the loopholes that let cloud companies strip-mine open projects. Anyone can use FinnG; nobody can enclose it. The same legal architecture that protected Linux, applied at both layers of the stack.
Models run on the user's machine: phase one on existing NPU hardware (Hailo, Kinara), eventually on community-designed boards. Queries never cross a network boundary unless the user explicitly chooses. Privacy isn't a feature; it's the architecture.
Fine-tune open-weight models (Mistral, Qwen, Gemma) on domain-specific data: retirement, career, legal, clinical. LoRA / QLoRA pipelines. Evaluation harnesses. Comfortable with Hugging Face and quantization.
For the daughterboard work. Real soldering iron needed. KiCad or Altium. Experience with NPU integration (Hailo, Kinara, or similar). Help us decide between custom PCB and integrating off-the-shelf modules.
Limited Cooperative Association formation, dual-licensing strategy, member contribution agreements. Bonus: comfort with open-source licensing edge cases and the strongly-reciprocal landscape (AGPL, OHL-S, SSPL, FSL).
FinnG Transition's guided conversational flow needs a designer who's done real conversational UX; not a chatbot, not a dashboard, something in between. Quiet aesthetic. Designed for trust, not engagement metrics.
International worker retirement is a real niche. Quarters of coverage, totalization, NRA tax. If you've worked with cross-border financial planning and want the calculator to actually be correct, we want you.
Cooperative-friendly funding sources exist, open-source foundations (NLnet, Sovereign Tech Fund, Mozilla MIECO), academic research grants, and civic-tech philanthropy. Help us write proposals that explain why a community-owned local-first AI cooperative is worth funding. Bonus if you've worked with non-profits or academic grant cycles before.
Help design FinnG's revenue mix: dual licensing, member contributions, hardware bundles, paid services, sponsored research. The goal isn't venture-scale returns; it's sustainable income that pays contributors fairly without compromising the cooperative thesis. Comfort with co-op finance, mission-driven fundraising, and modest-scale sustainability planning matters more than corporate fundraising experience.
The FinnG documentation, distilled from real research conversations, is the recruiting document for the next phase. Help us turn it into something readable. Also: documentation, blog posts, member onboarding materials.
FinnG isn't a company with employees. It's a cooperative being built by its contributors. Every member of record holds voting rights on project governance, license decisions, and revenue allocation. One member, one vote regardless of how much code, design, or capital they brought.
On license changes, governance decisions, leadership selection, and major technical direction. The vote isn't weighted by contribution size - every member's voice carries equally.
Income generated by FinnG (commercial licensing, hardware sales, paid services, grants) flows back to members through a defined revenue-sharing formula, governed by member vote.
The cooperative's bylaws make acquisition impossible without unanimous member consent. FinnG cannot be sold off to a corporate buyer. The commons stays a commons.
Sustained contribution over a defined period, then a vote of existing members. The pathway is documented; the criteria are public; new members are admitted regularly.
FinnG Retire is live, but could be improved. FinnG Transition is in design. Cooperative legal structure will be drafted with member counsel. First contributor cohort onboards. Software runs on existing NPU hardware users may be utilized soon.
Partner with Hailo or Kinara for bulk NPU modules. Pre-configured FinnG hardware kits ship to early users. FinnG AI prototype resumes - privilege-clean local AI for solo and small-firm attorneys, with healthcare and accounting verticals queued. Cooperative will be formally incorporated, first member meeting.
FinnG-designed external dongle ships under CERN OHL-S v2 - same NPU silicon as Phase 1, custom board and enclosure owned by the cooperative. FinnG AI expands into healthcare, accounting, and journalism use cases. Framework Laptop integration explored.
If traction supports it: train a small FinnG-native foundation model (1–3B parameters) with fully documented training data provenance, governed by the cooperative. The whole stack, model, software, hardware design, owned by the community that built it.
FinnG isn't a product looking for users yet. It's a project looking for the right people. If you've been waiting for an AI project structured like a commons rather than a startup, this is that.