Open-source models are getting better every week. A month ago we were on different ones; today it's DeepSeek V4; next month it'll be something else again. I install on-device AI in your firm, train your team, and keep your stack on the current best model — securely. The hardware sits inside your office. Client data never leaves the building.
A real NDA. Llama 3.3 70B. Wi-Fi physically off. lsof running live on screen to prove no network traffic. The model reads the document and flags unusual clauses — all on the laptop, with the receipts visible.
Our code makes zero network connections. Audited every dependency. Verified live with lsof.
I run benchmarks on every new open-source release that drops. When something meaningfully beats what's on your box, you get an update. Secure pipeline — signed USB w/ checksums or in-person install.
I'm your local AI guy. Install, train, maintain. No per-seat pricing. No SaaS lock-in. The hardware and the stack are yours — I'm here to keep them current.
Four steps. Then I stick around.
I set up a local AI server on Apple Silicon hardware inside your office (M-series, 64 GB+ recommended). Double-click launcher. No command line for staff.
I train two of your staff on the NDA + contract review workflow. 90-minute session, recorded for replay. You learn enough to be dangerous on your own.
Partners drop documents in, get structured clause analysis back. Wi-Fi can be off. Client data never touches a third-party server.
When a better open-source model ships, I bring it to you — securely. Signed USB drop with verified checksums, or I fly in and load it on-site. Your compliance posture, your call. You stay on the best stack without thinking about it.
Same speed. Radically different trust model.
| Scenario | Cloud (ChatGPT, Claude Cloud) | AirGap AI |
|---|---|---|
| Reviewing a sealed NDA | ⚠️ Risky — data leaves the firm | ✓ Stays on-device |
| Client asks about data retention | ⚠️ Depends on provider ToS | ✓ Nothing retained — no network call |
| Working on a plane / offline | ✗ Doesn't work | ✓ Works |
| Monthly cost for 4 partners | $80–400+ recurring | $0 after install |
| State Bar compliance burden | ⚠️ Depends on jurisdiction | ✓ Client data stays on-premise |
Open-source AI is moving fast. Here's what that actually means for your firm.
Every week there's a better open-source model. A month ago we were running a different setup; today it's DeepSeek V4; next month it'll be something else. If you bought a "frozen" local AI box, you'd hate it inside 90 days.
That's why this isn't a product purchase. It's a relationship. I install the hardware once, teach your team enough to be dangerous, then keep you on the current best model with a secure upgrade process — signed USB drops with SHA-256 checksums your IT verifies before load, or I fly to your office and load it on-site with full chain of custody. Your compliance posture, your call. When the next breakthrough drops, you get it. Without the cloud.
I'm taking one firm at a time right now. White-glove from install through the first 30 days, then optional ongoing model upkeep as your local AI guy.
After handoff: optional retainer for ongoing model upkeep — I track new open-source releases, run the evals, and bring you secure updates as the landscape evolves. Without the retainer, your hardware still works indefinitely; you just do the model selection yourself.
Request the intro callOr email matt@ineedhemp.com directly
Narrow on purpose. The first engagement is built around law firms — NDAs and contract review — and the playbook expands from there.
Good fit:
If you're in healthcare, accounting/audit, or therapy and want this — email me anyway. I'm taking one firm at a time, but I'm tracking which verticals come up next.
Nice Dreamz LLC designs and deploys local-first AI systems for firms that can't put client data on someone else's server. AirGap AI is our commercial offering for regulated industries — built on top of our open-source stack that now powers over 1,400 installations worldwide.
We publish our work in the open. Clients can audit every line of the inference layer, verify the on-device claims with their own tooling, and keep the stack running indefinitely whether we continue to support it or not. No lock-in, no black boxes, no license audits.
For bounded professional tasks — NDA clause extraction, contract diffing, precedent comparison, document summarization — performance is comparable to cloud frontier models on the workflows we deploy. For open-ended reasoning or drafting complex briefs from scratch, cloud frontier models retain an advantage. Part of the engagement is mapping which tasks belong on-premise and which can safely route to cloud with appropriate redaction.
Nothing outside standard application deployment. The stack runs as a local process listening on loopback only — no inbound firewall rules, no open ports, no VPN configuration changes. Installation is equivalent in complexity to deploying a standard Mac business application.
Apple Silicon Macs (M2 / M3 / M4 / M5) with 64 GB unified memory or higher are the recommended configuration. For firms running 32 GB workstations, a smaller-parameter model provides good results for most document workflows. Hardware sizing is scoped during the intro call based on firm size and document volume.
AI outputs are always treated as first-pass material subject to attorney review — the deployment is structured around that assumption, not around replacing attorney judgment. Every engagement includes a written compliance memo documenting the tool's failure modes and the review workflow designed to catch them.
Every week I evaluate new open-source releases against the document-review benchmarks I deploy. When a model meaningfully outperforms what's on your box, you get an update. Two delivery options based on your firm's compliance posture:
Signed USB drop: I overnight a tamper-evident drive with the model and an attached SHA-256 manifest. Your IT verifies the checksums before load. Full audit trail.
On-site install: I fly to your office and load the update in front of you. You watch the whole chain of custody. The Wi-Fi stays off the entire time.
You decide which method fits your firm. We set up the pipeline once, during the Foundation engagement.
Your firm owns the deployed stack outright. Optional ongoing retainer covers model upkeep (the secure update process above), new document types, additional workstations, and workflow tuning as the landscape evolves. Without the retainer the system continues to operate indefinitely — no remote kill switch, no license check-in, no subscription dependency. You just do the model selection yourself.
The underlying open-source stack is public at github.com/nicedreamzapp/claude-code-local — 2,600+ stars, MIT licensed, audited by people I've never met. You can run it indefinitely without me. The retainer pays for the curation, not the right to use the software: I track what's new, run the benchmarks, and bring you the right updates so you don't have to follow the AI news cycle yourself.
Larger consultancies typically quote six-figure engagements on multi-month timelines because their cost structure requires it. Nice Dreamz delivers the same outcome on a fixed-scope engagement because we built the underlying open-source stack and deploy it directly. Clients get the principal engineer for the full engagement, public source code they can audit without an NDA, and a clean exit path at the end of the pilot.
A fifteen-minute conversation to understand your firm's document workflows, answer compliance questions, and determine whether the Foundation engagement is an appropriate next step.
Technical team prefers to evaluate the code first? The builders' community is on Discord: discord.gg/ZdSqgAxUW