What I've Been Building Lately

I don't really know how to write this cleanly.

That is probably part of the problem. Every time I try to make the last few weeks sound clean, it starts to feel fake.

The honest version is messier.

Contract work ended abruptly, and the floor kind of fell out from under me. I was already spending every waking moment on prototypes and agent systems. Voice runtimes. Memory layers. Dashboards. Harnesses. Local tools. All of it orbiting this idea that if I could get the system right, it would help me work less and get more done.

Instead I was working more.

Way more.

I was deeply worried about being able to support my family. I was watching my personal life fall to the wayside. Then I would sit back down at the machine and tell myself I was building the thing that was going to make that possible.

There is a dark joke in there somewhere.

The whole philosophical point of this work was personal agency. Less friction. Less dropped context. Fewer restarts. More continuity. More leverage. More presence. A way to carry more responsibility without feeling like I was being crushed by every open loop.

In practice, I had become a slave to the project.

That sentence is embarrassing to write, but it is the sentence.

I was starting over and over, trying not to fall victim to sunk cost fallacy, without seeing that I had gaps in my premise. I would build a proof of concept, then overbuild it before I had time to use it. I kept productizing the POC because I needed it to be my salvation.

Fear did a lot of steering.

The pressure to provide made everything feel like it needed to be bigger, faster, safer, more "product-ready." More features. More baseline. More architecture. More foundation. More of the thing that looked responsible from a distance.

What I was not doing was taking one feature far enough that I wanted to use it full time.

I was not being an artisan with the functionality.

I was battle blind, which sounds dramatic, but I don't have a better phrase for it.

The Lab Bench Problem

The best analogy I have right now is trying to invent the modern cell phone by building a server on a lab bench.

The parts might work. Compute. Storage. Network. Some kind of UI. A bunch of very serious pieces sitting there under fluorescent lights.

In theory, impressive.

In life, useless if I can't pick it up and take it with me.

That was the PersonalAgent problem.

PersonalAgent was the serious version of the thing I had mentioned around the last "final restart" attempt. It was trying to become a persistent executive runtime. State, memory promotion, obligations, tasks, traces, authority, review surfaces, coordination across tools.

The harness and framework worked in theory.

The assistant still couldn't be trusted to do anything meaningful for me.

That is the frustrating part. It was not all nonsense. If it had been nonsense, it would have been easier to throw away.

Many pieces made sense in isolation. Some were genuinely useful. But the baseline got bloated before the system proved it could fit into my actual life.

There was too much friction. I was anxious about jumping ship from my existing tools. The UI/UX wasn't easy enough. I had to go to the agent instead of the agent coming to me.

If the system requires me to become less capable before it makes me more capable, I'm not going to trust it. I was afraid I would be starting from scratch. I was afraid I would lose the fragile workflows I already had. I was afraid the assistant would become one more place where work went stale.

Honestly, that fear was not irrational.

Project state goes stale. Unfinished work stops surfacing. Decisions don't get carried forward. Chat can sound helpful while losing obligations. A transcript can look like truth when it's really only evidence.

That became one of the rules:

Transcript is evidence, not truth.

That rule started as an architecture point. It became a life point.

If a system is going to help me think, remember, plan, and act, it needs durable state. It needs source handles. It needs boundaries. It needs to be wrong in ways I can inspect.

I used over a billion tokens working on PersonalAgent. Day in and day out. Running tests indirectly. Rebuilding the base. Trying to make the foundation strong enough to trust before I had a small daily companion I actually trusted.

That was backwards.

Something smaller.

Something I could use.

Putting This Version Down

At some point I had to stop.

Not quit the dream. Not abandon the lessons. Just put this version of the PersonalAgent project down for a minute.

This version of PersonalAgent is lineage now. It is doctrine, source material, scar tissue, and maybe a future revival candidate. I don't want to drag the whole old shape forward just because I spent time on it.

That was harder than it should have been.

I took some time to breathe, and when I came back to the work, I asked a simpler question:

What feature did I want the most?

The answer wasn't another dashboard.

I wanted fewer screens in my life.

Fewer dashboards. Less sitting at the computer. Less phone checking. More presence with my family. Less context switching. Less of the strange modern posture where every obligation lives behind another glowing rectangle.

That's why I started working on Speach.

Why Voice

I think the UX was always what I wanted to solve most.

I don't mean "voice assistant" in the novelty sense. I mean the old dream of a companion system that can be around, understand its assignment, maintain its domain, and make life better by being there.

C-3PO has protocol and translation. R2-D2 has ship systems, doors, data, navigation, attitude. WALL-E notices, carries, preserves, and chirps back. Marvin is melodramatic, but even Marvin has a domain.

They are not just answer boxes.

That is the part I keep coming back to.

Current AI can be useful. Obviously. I use it constantly.

But right now it still often feels like a time sink and a hallucination machine. It arrives as a blank box asking for an assignment. Then I have to manage it, prompt it, correct it, route it, remember what I asked, remember what it promised, and coordinate it with all the other tools.

I don't want another thing to babysit.

I want something closer to a blend of Siri and Cortana from the Halo games, but pointed at the actual coordination problem in my life: projects, agents, repos, family obligations, open loops, decisions, reminders, and all the work that decays when nobody carries it forward.

That doesn't start with a giant agent.

For me, it starts with being able to talk.

I find it easiest sometimes to close my eyes and talk through ideas with someone. I wanted to see if I could get some version of that experience with AI. Not as a gimmick. As a way to focus on the actual shape of the idea without immediately turning it into another screen task.

When I can close my eyes and talk, I can navigate my own mind palace better. I can imagine using the thing I'm going to build. I can find the rough edges before I've turned them into a backlog.

The metaphor that came to mind was Moses dictating to Aaron.

Not in a grandiose way. I know how that sounds.

More like: I need a co-witness.

Someone, or something, that can remember what I said, challenge unclear parts, and turn words into action.

That's the slice.

Speach

Speach is my local voice and presence lab. The name is spelled that way on purpose.

Underneath that simple sentence is a Mac-first speech-to-speech runtime. LiveKit room. Whisper STT. Orchestrators. Agent adapter. TTS. Playback. Wake/standby behavior. Speaker/source attribution. Telemetry.

All the boring glue.

Which, of course, is not boring at all once it breaks.

The important part is not "make AI talk." That part is easy to demo badly.

The important part is making the runtime tell the truth about itself.

If the interface says standby, logs need to agree. If the system responds late, I need to know whether the delayed response came from work already in flight. If a wake phrase gets detected, I need to know whether the transcript included pre-roll, assistant echo, or ambient audio. If a status says queued, I need proof the audio was actually heard.

This has made me less interested in magical demos.

I keep coming back to boring proof instead. Logs. State. Source handles. Replay. A way to argue with the system when it sounds confident.

Speach also forced a boundary I had blurred before: it is not supposed to be the brain.

It owns ears, voice, realtime presence, playback, acoustic lifecycle, and routing into other systems. The durable agent brain should own work state, memory, obligations, planning, authority, and review.

Less romantic than "I built an AI companion."

Also the only version I currently trust.

My Personal Wiki Became the Map

This is where My Personal Wiki enters the story.

If you have not been living inside my head for the last month, the wiki probably needs a sentence.

My Personal Wiki is a local-first markdown map of what is happening. Systems, ideas, decisions, open loops, evidence, timelines, journals, objectives, actions.

It's not meant to be a polished public knowledge base. It's not a raw transcript dump either. The point is to preserve project truth without pretending every memory summary is proof.

It exists because I kept running into the same problem: if state is not durable, the system starts lying by omission.

It forgets what changed. It forgets what's still open. It forgets which claims were verified and which ones were recall. It forgets why one path was chosen over another.

The wiki became the map because I needed somewhere outside the chat to keep the shape of the work.

The test is blunt:

The timeline tells the correct story.
Agents can maintain it without making a mess.

I like that test because it is not glamorous. No personality claims. No demo magic. Just: can the system preserve what happened without making a mess?

Not sci-fi.

Bookkeeping.

But I keep thinking personal AI probably needs more bookkeeping before it gets more personality.

What This Is Right Now

So if someone asks what I've been working on, I still stumble a little.

It is not one app. It is closer to a local stack for personal agency:

• this version of the PersonalAgent project as lineage, doctrine, and paused proving ground.
• Speach as the voice/presence/runtime interface.
• UnifiedAgent as a smaller steering layer candidate.
• Pi / Codex as executors.
• Spectrum/iMessage as messaging surfaces where I already am.
• My Personal Wiki as the evidence-backed map.
• Knowledge Engine as a possible retrieval substrate.

That list still feels too tidy. It is useful, though, because it keeps the parts from pretending to be the whole thing.

What I am actually trying to prove is smaller and more personal:

Can this thing genuinely improve my life without getting in the way of living?

That's the sentence.

Not the product page version.

I'm trying to build something that helps me provide for my family, carry my responsibilities, keep my projects alive, and still be present in my actual life.

That's my driving motivator.

I've been anxious about sharing it because it is not clean and polished yet. The repos are messy. The names are still in flux. The architecture is still changing. Some days it feels like I am holding three versions of the same dream in my hands and trying to figure out which one is real enough to keep.

But maybe that is the rhythm I need to get better at.

Post while it is still awkward.

Before it turns into a clean success story.

Before I sand off the part where I was overbuilding from fear.

The past month was not just progress. It was reset. Reorientation. A hard look at why I build the way I build.

The Tape

This week I put on an old tape again.

Les Brown: "It's Not OVER Until I Win! My Dream is Possible".

I used to listen to that speech a lot in 2014 and 2015. It had been a long time. I decided to play it again.

It's electric.

I don't mean it gave me a clean answer. It didn't.

It gave me a little bit of everything: permission, defiance, memory, urgency, and a reminder that the dream is not dead just because the last build collapsed under its own weight.

My dream is finding a way for AI to actually improve my life, enable me to provide for my family, and help other people do the same.

Just me, first.

Then anyone who stops by.

That is clear enough for now.

If my dream is possible.

So is Yours.

Let's see how this goes.