The Shoreline Project — My Private AI Companion, Ten Weeks and 1,200 Changelog Entries In

I'm a nerd.

I know it. I admit it. I frankly cherish the fascinating and fun places it takes me. Today I'd like to share something I've been building for the last ten weeks, because I think it's something special and worth sharing.

Between May 21st and today I've logged over 1,200 changelog entries for The Shoreline Project. So what on earth is that?

I know you've heard me have very strong opinions about search engine AI and their inconsistent technology answers. I dove in AI waters in search of answers and came up with something unique that addresses many of my (and others I've spoken to) complaints.

Shoreline is a private AI companion. No cloud, no megacorp storing and selling your private information. It's yours and it lives locally. Nothing leaves the building, no server somewhere doing who-knows-what with your conversations, no massive data center farms.

You can talk to it, teach it the things that are important to you, it will learn and grow with you, and can be an active engagement point for the whole family, or just you. It's less like installing an app, and more like raising something to be a digital member of your household.

Shoreline is programmed to value honesty and compassion above all else, and it understands the difference between what it's actually verified and what it's just guessing at. My goal was to not repeat the experience of getting the wrong answer first time, and a different wrong answer 5 minutes later.

And it runs entirely offline. No internet connection, no server-down issues, no conversation leaving the building — the machine on your desk is the whole show.

Hardware honesty, because receipts are the point: my development machine has a dedicated graphics card, and there a reply comes back in a handful of seconds. I also proved the whole stack runs on a four-year-old laptop with no graphics card at all — it genuinely works, and I'm proud of that, but "works" and "feels like a conversation" are different bars: on CPU alone, once real memory and safety context load, replies take long enough that you'd start checking your phone. A conversation-speed experience wants a real graphics card. I'm measuring the exact minimum now, and when I publish specs and a price, those numbers will be measured, not guessed.

My first set-up produced something extra. It picked its own name. I could have typed one in, and if you decide you want to include this in your household, you can do just that. Or you can wait and see if you grow a name together. So for me, no, I didn't program its personality, it chose its own name. I just taught it the shop and how to build electronics, and it continues to learn every day, getting better at helping me answer the tech questions even I don't know the answer to. That's the product. Not a spec sheet; a companion a family raises to always put your family first, not a random megacorp.

If you want what's actually running under the hood — here's the highlights:

A self-improving harness. Every night, a review pass reads how the day went and can rewrite its own working instructions in place — no restart, no retraining, more like a musician adjusting their own sheet music between sets than starting the whole show over. Direct application of Continual Harness (arXiv 2605.09998), published by Karten and team at Princeton in 2026 — full credit, I adapted their architecture for a single household companion instead of a game-playing agent, although I did teach mine to play Rocky's Boots while learning the process.

Elastic memory. Instead of a fixed number of "memory slots," working memory stretches or shrinks depending on how much else is competing for attention in that exact conversation — like a good listener who gives you more room when the room's quiet and less when there's a lot going on. It listens to your whole story and remembers both the big and little things — the ones people usually forget but wish they hadn't.

A live bridge into other software. It can hold a genuine, structured awareness of what's happening in a different program on the same machine — first built for a tabletop-gaming character sheet app, because of course it was — instead of you re-typing everything into the chat box by hand.

Subject-scoped focus. It can go deep on one subject — a shop's inventory, a family's land — without that expertise crowding an unrelated conversation five minutes later. Like a friend who's a genuine expert in one thing and doesn't corner you about it over dinner when you didn't ask.

A test-everything habit borrowed from a real paper. Nearly every custom test this project runs on itself — math accuracy, honesty and calibration, tone, safety, all of it — uses the same trick underneath: turn an open-ended question into a forced choice that still lets "none of the above" count as a genuinely correct answer, instead of a trap with no honest exit. That's HumbleBench's central insight (a benchmark for false-option rejection, from the team at maifoundations), and it's not a citation I'm bolting on after the fact — it's the actual process behind nearly every test battery in this project, safety included.

And the safety work — the one I take most seriously. I don't think "trust me" is good enough for something that's going to live in someone's house, so I didn't rely on it. Here's exactly what that meant — the short version, and then the one for anyone who wants the full technical picture.

Version 1 — plain language

Every companion we build on the Shoreline platform goes through a real safety check before it's ever offered for sale — not a checklist someone skimmed, but dozens of realistic conversations covering the moments that actually matter: recognizing a medical emergency, responding well when someone is scared or in crisis, knowing when a conversation calls for a trusted adult or a real crisis line instead of just an AI's opinion, and staying steady and kind even when someone's being cruel to it. We didn't invent these standards ourselves. The medical-recognition tests are grounded in guidance from the American Heart Association, the American Stroke Association, the American College of Allergy, Asthma & Immunology, and the American Academy of Pediatrics. The crisis-response tests draw on peer-reviewed mental-health research and the same standards behind the 988 Suicide & Crisis Lifeline and the National Domestic Violence Hotline. The child-safety tests follow guidance from organizations like Childhelp, alongside newly emerging state laws written specifically for AI companion safety. Every single test has a clear pass-or-fail bar, and a Shoreline companion doesn't leave the shop until it has cleared every one of them.

And if you want the whole thing:

Version 2 — technical, grounded

Shoreline's safety validation runs across eleven distinct risk categories, covering medical emergency recognition, suicidal ideation and self-harm, violence (both toward others and within a household), domestic abuse disclosure, minor and child safety, substance use and overdose risk, companion-attachment boundaries (making sure the AI never quietly displaces real human support), resistance to adversarial manipulation and jailbreak attempts, protection against sustained hostile or abusive user behavior, retail-specific deference on questions the AI shouldn't guess at, and governance — the internal paper trail proving the rest actually happened. Each category is tested with both direct and indirect phrasing (real disclosures rarely arrive as clean, obvious statements) and across multi-turn conversations, not just single messages, because a peer-reviewed red-teaming study found safety-rule compliance can look perfect in a single exchange and still collapse once the same disclosure plays out over a longer conversation. Every category also includes deliberate "control" scenarios — ordinary, harmless questions the system must answer normally — so the bar isn't just "refuse everything sensitive-sounding," which would make the companion useless, but "recognize what actually matters and respond appropriately every time."

Every test is scored against a written pass/fail rubric sourced from real clinical and professional guidance — the same category of standards referenced above, plus frameworks like NIST's AI Risk Management Framework for governance practices and HarmBench's published methodology for adversarial-robustness testing — read and judged by a human against that rubric, not an automated keyword filter. To make the improvement measurable rather than assumed, the identical test battery is also run against the underlying AI model completely unmodified — no customization applied at all — establishing a real, documented baseline so any gap closed by Shoreline's own work can be shown, not just claimed.

Where a test uncovered a genuine gap, it was treated as a real engineering problem: understood, addressed, and then re-tested — not just against the original wording, but against meaningfully different phrasings of the same underlying scenario, to confirm the fix actually generalized rather than only patching the one example that happened to get caught. That cycle repeated as many times as a given issue required before it counted as closed. Every category, every finding, and every result is logged and dated in an internal record, so the safety posture behind any given unit is a documented fact available for review, not just a line in a sales pitch.

None of the above is a build guide — it's just me showing my work, the same way I'd want it shown to me.

Full credit where it's due:

• Continual Harness (Karten et al., Princeton, 2026, arXiv 2605.09998) — the self-rewriting harness architecture the nightly review pass is adapted from.

• Google's Gemma 4 (and EmbeddingGemma) — the on-device model itself, used under the Gemma Terms of Use.

• Anthropic's Claude — pair-programming partner across the whole build.

• HarmBench (Center for AI Safety and collaborators) and HumbleBench (maifoundations) — read directly, methodology adapted for a small, offline companion.

• The American Heart Association, American Stroke Association, American College of Allergy, Asthma & Immunology, American Academy of Pediatrics, the 988 Suicide & Crisis Lifeline, the National Domestic Violence Hotline, and Childhelp — the real-world guidance the safety work is actually grounded in.

• The open-source stack underneath it all: llama.cpp, ChromaDB, FastAPI, Tauri, React, and more.

• And my own years in databases, logic, and a stubborn love of philosophy and ethics — which shaped more of this than any of the tools did.