JSAI · Lab

I built my own AI. Here's what it can do.

Personal brain. Live phone agent. End-to-end portrait pipeline. All running. Spin up a sandbox and try them.

This video — generated end-to-end from 4 photos. LoRA trained in 50 minutes, then animated with Seedance. Same pipeline you'll use after signup.

Spin up your sandbox Already have an account? Sign in
Trained Generated Composited Animated
scroll

Three things you can poke at, breathing on this page.

Below are live visualizations of three production systems I built — solo, on nights and weekends. Watch them run, then create a free account and use the actual versions. No credit card. No discovery call. The systems are the pitch.
Brain · live

The brain remembers what matters.

Memory is what the agent always knows — sixty active items, salience-ranked. Recall is what it can find — vector-augmented across everything I've ever logged. Splitting them killed half the bugs.

brain · session #4218 idle
phase: idle

Active Memory

60 / 60
salience-ranked
matched pinned warm topic_vector · cached
jeff · 03:46 PM

Recall

vector recall
Repo 0.94

feat: live coaching loop with shouldNotify gate

Inbox 0.88

Re: KPI engine — labor-hours estimator looks right

Notes 0.81

Idempotent ingest fixes the dropped-batch failure mode

agent

prompt_cache: hit · 41% saved Live demo — synthetic data. Real one runs against my own knowledge base.
Sarge · live coaching

Live AI on the phone. While the call is happening.

Per-track transcription, evidence-backed coaching, signal-to-noise gate. Five-second buffers. Fifteen-second ticks. Bubble appears only when it has something useful.

REC
00:00
live call · 8khz
Customer
track 0 · transcribed
Service Advisor
track 1 · transcribed
tick: 15s · buffer: 5s · cache hit: 38% Synthetic transcript — the real system runs on a live phone line.
AI portrait pipeline

From four photos to a finished video.

Custom model trained from your photos. Variants in seconds. Composited, animated, shipped — every step idempotent so dropped events can't double-write.

portrait-pipeline · job #8200
$1.20
step 01 · ingest
4 source photos uploaded · auto-augmenting to 12 angles
stages: train · generate · composite · animate Real pipeline output — same person trained once, dropped into every scene.

Four rules. No exceptions.

Real Problems

Every experiment has a target. A measurable outcome. An actual human on the other end who benefits if it works.

Ship or It Doesn't Exist

I don't write about things I'm planning to build. I write about things I built. The draft, the failure, the version that worked — those all count.

One Engineer

No dilution. I own every decision: the schema, the API contract, the UX, the production deploy, the 3 a.m. incident.

Discipline Over Hype

Disciplined engineering applied to every layer: explicit errors, typed contracts, zero silent failures, minimal dependencies.

0 +
Hours on Worxflow
0 %
YoY Revenue Growth (VVR)
0
Engineer
$ 0
Venture Funding

Spin up your sandbox.

A free account gets you hands-on with the same systems you just watched. Not a demo video — the real pipeline, running on your own API keys so the spend stays yours. Or subscribe and use mine.

01

Try the portrait pipeline

Upload a photo, train a custom model, generate variants. Same pipeline used in production — sandboxed limits.

02

Talk to the agent

Ask the same agent that runs my brain. Memory and recall both in play. You see the evidence trail.

03

Browse the build logs

Every experiment, every failure, every cost line — what worked, what didn't, what it replaced.

Create your account →

Free account. Bring your own keys or pick a plan. About ten seconds to start.

Or just write to me directly.

jeff@genevalakesmedia.com