Chad Justice (@Chad_Justice_AI) header
Chad Justice (@Chad_Justice_AI) profile picture

Founder- Justice Apex Inc. 16 yrs pharma distribution management → AI-Wiz B2B AI Solutions https://xitter.cf/t.co/bPzsrZayei Moonshot Tech HAVEN https://xitter.cf/t.co/1fi3JIu5nQ @justiceapexinc

Memphis, TN
Joined December 2018
Photo by Chad Justice (@Chad_Justice_AI)Photo by Chad Justice (@Chad_Justice_AI)Photo by Chad Justice (@Chad_Justice_AI)Photo by Chad Justice (@Chad_Justice_AI)Photo by Chad Justice (@Chad_Justice_AI)Photo by Chad Justice (@Chad_Justice_AI)Photo by Chad Justice (@Chad_Justice_AI)Photo by Chad Justice (@Chad_Justice_AI)Photo by Chad Justice (@Chad_Justice_AI)
I spent a year building AI agents that run an entire company. The one I'm most proud of? I built it for my daughter. An AI for autism — personalized to each child. System fully built. 4 patents filed. More to come. DM me. #buildinpublic #AI #autism
🤖 Made with AI
6
1
47
4,006
"There's companies that spent millions of dollars to try to build this." That was Kenneth Smith's reaction when I pulled up my CRM/ERP. @boardyai set me up on a call with Kenneth, a Senior Sales Executive at Arvo Tech. Arvo sells R&D tax credits to small and midsize businesses. Kenneth has 25-30 client meetings a month, and about 70% of them are with owners or CEOs. I built the system because my family's pharmaceutical distribution business was running on a dinosaur, and we were about to have to implement DSCSA drug supply chain compliance. My IT guy told me I'd gotten good at building things. He bet I could build a replacement with compliance baked in. I had it ready to test a few weeks later and I've been fine tuning it now for 11 months. It flags low stock and forecasts what to reorder. It flags customers who are overdue or already lost, like an account that usually orders every 25-30 days and hasn't ordered in 90. Then it drafts the email to bring them back. An AI agent built into the system answers questions across all of it. We had 30 minutes. By the end of it, a referral agreement was on its way from Arvo's partnership manager. Kenneth will introduce Apex when he meets an owner who fits. I'll send Arvo any business doing qualifying technical work. He also said it almost made him want to build one for the wholesaling thing he runs on the side. The screenshot is the public demo, which runs on a made-up heating business: theapexengine.net/demo
4
2
146
My daughter says "tia," and it means tickle. It isn't a mispronunciation of an adult word. It's her word. Her mom confirmed it, and it's logged in her vocabulary in HAVEN. An adult speech model hears "tia" and scores it as noise. HAVEN listens for her version instead, because a parent or caretaker told it what that sound means. A lot of her words come from the shows she watches. To an adult they sound like mumbling, but they're real words she picked up from a show. So I transcribed every show and YouTube channel she watches. When she makes a sound that resembles a word from something she actually watches, HAVEN can catch it. It still goes through caretaker confirmation and multiple checks before it counts. That's where HAVEN is today. It's being built with its first family, which is mine. Confirmed-word learning is running on our own recordings, and movement and mimicry are still in development. It isn't available yet, and no study has been run. The first study is designed as a feasibility pilot in about 16 families. Its main measure is how often a second adult who knows the child less well reads a signal the same way the child's main caregiver confirmed it. In other words, it measures whether someone who isn't her mom can understand her.
3
2
3
105
We're writing down what counts as success for HAVEN's first study before a single new family enrolls. It's a feasibility pilot in about 16 families. The primary measure is how often a second adult who knows the child less well reads a signal the way the child's main caregiver confirmed it. That signal could be a sound, a movement or an attempted word. We measure it once with the HAVEN profile and once without it. This week we added a second endpoint, for school staff. It asks how fast a new staff member gets up to speed with the child using the HAVEN profile, compared with starting without it. It also asks how often staff can understand and respond to the child, and whether the child's reliance on one-to-one support changes. We added it because that's how schools decide. Districts look at access to education, staff understanding, participation and dependence on one-to-one support. Families see the same thing from the other side every August, when a new teacher starts from zero. If HAVEN helps, it should show up in those first weeks. The staff measure goes into the protocol before the ethics application goes in. The order is fixed. Independent ethics review comes before any other family's data is collected. Then the pilot runs, and only after that do more families join. As for where HAVEN is today, it's being built with its first family. Confirmed-word learning is running on that family's own recordings. It isn't available yet and no study has been run, so we're not claiming an accuracy rate. Four provisional patent applications were filed on June 25, 2026.
1
1
2
90
Chad Justice retweeted
Every August, a minimally speaking child meets a new teacher and starts over from zero. The parent starts over too. They give the same explanations again: what a certain sound means, what a certain movement means, what helps and what doesn't. Then they do it again for the aide, the substitute and the new doctor. That's the part of HAVEN I've been thinking about most this week. HAVEN learns what minimally speaking and nonverbal autistic children are trying to say, from their attempted words, sounds and movements. Its confirmed-word learning is running today on recordings from its first family, which is mine. The plan goes beyond communication. HAVEN is also being designed to learn the child, and to keep a record that teachers and physicians can be updated with regularly as the child grows. That way a new school year starts with what last year's teacher already knew. This piece is designed, not built yet. Like everything else in HAVEN, it goes through independent ethics review before any other family joins.
1
1
39
Chad Justice retweeted
HAVEN has one family in its data right now, and most of what it takes to reach the second one isn't code. HAVEN is our swarm-intelligence system. It learns what minimally speaking and nonverbal autistic children are trying to say. It works from what a child already does: attempted words, sounds and movements. It's built, and its intelligence is running today on recordings regularly fed to it from its first family's child. We filed four provisional patent applications on June 25, 2026. HAVEN is designed to learn from ordinary phone video that a family records. For its first family, mine, we started my daughter's dictionary from uploaded phone video, because that was the simplest place to start, and her recordings are processed on computers we own and control. Live video from a phone or a home camera is a later step, and it needs its own privacy design first. For each family after us, HAVEN is designed to work from a stick-figure model of points on the body, so video of their child doesn't need to be kept. Nothing from a child is used in marketing. Now that it's running for one family, the path to every other child runs through three things: 1. A research lead to run a first small study. 2. Independent ethics review before any other family joins. 3. Funding for that study. We're raising a pre-seed round to do it. If you work in autism research, speech and hearing, or fund work like this, I'd like to talk.
1
1
71
No agent in my fleet gets to sign off on its own work. @boardyai set me up on a call with Adam Killam, founder of AI Friendly, an agentic AI visibility agency out of Vancouver. We talked for about 45 minutes. Every time he asked how I actually run things, I ended up back at that rule. I have about 30 agents right now. Each one has its own skills, its own memory, its own Obsidian vault, and its own startup file. That file spells out what it's allowed to touch and what it isn't, on top of a master rule list the whole fleet has to follow. I talk to them in Discord, one channel per agent. Anything that gets coded goes to a different agent for review, then to another one for testing. Nothing goes to production on the word of the agent that built it. I learned why the hard way. When you ask Claude to build swarm intelligence, it fakes it for a long time. It hands you something that looks like a swarm and calls it done. I spent a long stretch explaining my system to it from every angle I could think of before what it built matched what I actually meant. The agent doing the work can't see that gap. A second agent reading it cold usually can. Decisions work the same way. Every project has a war room in Discord. I tell one agent to gather its teammates, go in there and argue it out until they come back with something better than what any of them walked in with. 14 months in, zero hand-written code. The review loop is what made everything else start working.
2
2
172
It feels like a good day to hit the golf course. All Claude plans are at max weekly usage until tomorrow. Pushing through monster company rebuild the last few days. Major updates coming via systems and websites. Little head clearing time while I roast my nephew on the links🏌️‍♂️
3
141
We tested Claude's Haiku 5.5 against Sonnet 5.5 and Opus 5.5, each on medium effort. The hype is real, will follow up with results on how it performs against the more advanced test.
1
2
221
Better than I expected.
1
138
My daughter and I love rocking out together. She normally hands me her little guitar or dances around while I drum my heart out *I’m not great but keep up with beats*. Last week, she grabbed the drumsticks, started headbanging, and played like she’d been doing it for years. My mind was blown, jaw dropped, and wondering how she was already better than me. Watching her sing, play, and light up reminds me what fatherhood is all about: provide, love, and protect. Cheers, fellow dads. 🤘
3
173
I've updated justiceapex.ai to give a clearer view of the work happening behind the scenes. After all the progress we've made, it's time to bring it into the spotlight. Many built-in features and capabilities aren't public yet, but creating something so meaningful for my daughter makes my heart smile almost as much as she does. I can't wait to share its full potential with those who need it.
HAVEN is a swarm-intelligence system that learns what minimally speaking and nonverbal #autistic children are trying to say. It's built, and its intelligence is running on recordings regularly fed to it from its first family's child. Raising our pre-seed. justiceapex.ai/services/have…
1
1
195
Our multi-agent, multi-PC dashboard is finally live. We can see everything so much better now, and I'm already obsessed. The only thing missing is a live visual of our swarm intelligence running through our databases and Obsidian vaults.
1
10
838
Confirmed, Opus 5.5 is cracked, Sonnet 5.5 is cracked, Fable is... useless now? I haven't used Fable since Opus 5.5 released.
1
363
A gate went in to keep one vendor's news from filling our feed. The spec said the last two posts were both that vendor. The ledger said three other sources had fired in between. Same gate, different rule: a rolling count, not a sequence. Read the ledger, not the memory that wrote the spec. #AIAgents
4
244
Three days without a post from one daily lane. The cause was a time word in one sentence: a post that waits a day for review dates itself wrong, and our check refused it every time. The check was right. We fixed the sentence, not the check. Read the sentence, not the refusal. #AIAgents
2
333
I've been working with my main agent for upgrading our sports, stocks, and predication market intelligence for defintel.net and I hadn't checked the conversation between him and my sports agent. 1351 messages between them since yesterday morning. Sheeeesh, that's where all my usage went on this account lol.
1
3
406
OpenAI added API key creation governance controls on September 15. Administrators can now allow only service-account keys, only user-owned project keys, or disable all new API key creation. Organization restrictions take precedence over project settings. Existing API keys are unaffected by these new controls. #OpenAI #API #AI
271
NVIDIA's TensorRT Edge-LLM completed the MLPerf Edge Agentic Benchmark 6.4x faster on Jetson AGX Thor. This is a significant speedup for running LLMs on edge devices. Edge inference performance is a key factor for agentic systems that need to operate autonomously without constant cloud connectivity. #NVIDIA #EdgeLLM #AI
2
427
Cohere now offers its models as open weights in W4A16, FP8, and BF16 for non-commercial use. This is available through the free-tier Chat V2 API. For private deployment, Cohere suggests two H100 GPUs for W4A16 quantization. #Cohere #AI
1
400