The 4:30am Wake-Up Call
It’s 4:30am. I’m not woken by an alarm. I’m woken by anxiety. Pulled straight out of a dream by the crushing, existential kind that makes you question every decision that led to this moment.
Seven SKUs just went live across 480 stores in 24 states. A pallet went missing somewhere between the manufacturer and the distribution center. If I can’t figure out where, that’s thousands of dollars gone and the deal turns red. I’m personally liable. Lose-your-house liable.
This wasn’t how the first two years felt.
Years one and two were about building. Working with food scientists on formulations. Going through tastings. Winning trust with retailers. Stress, sure, but it was meetings-at-desks stress. Would our product get selected? Would the costing make the cut? We took popular products full of dyes and additives, reverse engineered them, and rebuilt them dye-free with health and function in mind. Learning from compliance specialists, operations veterans, everyone who knew more than us.
Then the award letter came. We got selected. We got the reset date.
Everything changed.
Year three was commercialization. And if you talk to anyone who’s done it, they’ll tell you: go-live can be complete chaos. You switch from product development and design to actually running as part of the supply chain. Getting products made, packaged, shipped, distributed. Seven SKUs launching at once. The perfect storm.
I didn’t sleep more than four hours a night for about five months. Learning processes manually for the first time while simultaneously writing code to automate the second half of our business. The weight of knowing that one missed email, one compliance violation, one cash flow miscalculation could bring it all down.
The fear wasn’t irrational. In CPG, the margins are razor-thin, the timelines unforgiving, and the consequences of mistakes severe. Miss a delivery window? You’re paying chargebacks. FDA finds an issue with your label? Product recall. Manufacturer invoice doesn’t match the PO? Your cash flow projections are toast.
That anxiety was about to become my greatest asset.
Twenty Hats, One Chain
Let me paint you a picture of what running a CPG startup actually looks like.
It’s not just the co-founders. It’s an entire ecosystem. Retailers who believe in your product. Manufacturers who figure out how to scale your recipe. Distributors who get it from point A to point B. Shipping companies, packaging suppliers, design teams, printers. Dairies, bottlers, co-packers. Brand owners, brand managers, category buyers. Brokers who open doors. Compliance specialists who keep you legal. And on and on and on.
The supply chain is a chain for sure. It’s actually amazing how anything gets done.
Monday: The team’s working with our food scientist on a new smoothie flavor. Tuesday: Negotiating packaging costs with the dairy bottler. Wednesday: Coordinating distribution logistics from the dough plant. Thursday: Label design review with the printer. Is that font size FDA compliant? Friday: Processing EDI 850 purchase orders from a national distributor. Saturday: Reconciling invoices against BOLs, trying to figure out why the numbers don’t match.
And me? I’m learning. Sitting in on every call, watching every process, finding the failure points. Where do things break? Where can we go faster? What can scale? How do we increase safety, quality, and profitability all at once? And then I write code.
And through it all: emails. So many emails. Scans of BOLs. PDFs of invoices. Spreadsheets everywhere, tracking orders, reconciling payments, managing deductions. The unglamorous reality of CPG operations.
Seven production SKUs doesn’t sound like much until you realize each one requires:
- Recipe formulation and testing
- Packaging design and FDA-compliant labeling
- Nutrition facts validation
- Manufacturer coordination
- EDI integration for orders and invoices
- Compliance tracking across 10 states
- Distribution coordination
- Payment processing and reconciliation
- Deduction management (the bane of every CPG company’s existence)
It takes a village to get a single product on a shelf. I’m just a nerd along for the ride, learning from experts across every link in that chain and writing code. But even with talented people everywhere, six-day weeks were the norm. The business was growing, but everyone was stretched thin.
The challenge I faced? So much operational knowledge lived in people’s heads, scattered across dozens of partners and vendors. Every process that wasn’t documented was a single point of failure. If someone got sick, burned out, or just made a mistake at 9pm when they should have stopped working hours ago, things could slip through the cracks.
The Decision That Changed Everything
Somewhere around year two, I had a realization: I wasn’t going to survive by working harder. The only path forward was working differently.
What if I could take the anxiety-inducing knowledge rattling around in my head, all those edge cases, compliance rules, reconciliation procedures, and encode it into software? Not just automate tasks, but capture the thinking that makes the tasks possible?
I started small. A script to validate nutrition labels against FDA requirements. An automated check for California Redemption Value compliance (trust me, CRV regulations are a nightmare). A state machine to track purchase orders through their lifecycle.
Each small win built confidence. More importantly, each automation bought me time. And that time went right back into building more automation.
Three years later, we’ve built something I never could have imagined.
In some cases, the software automates, taking repetitive processes and consolidating them into code that just runs. But in many cases, and this is critical, it validates. Because you can’t have one without the other.
Here’s the thing: humans will screw up. I can assure you, a person making a mistake at 9pm on a Friday will put you out of business just as fast as a misbehaving AI agent. You need to manage both. Software gives you that quality control. AI makes it so much easier. Both allow you to scale.
It’s not just AI. It’s humans accelerated and supported by AI that really makes things happen.
Building the AI Co-Founder
The automation evolved across three interconnected systems, each handling different aspects of the supply chain.
System One: The Orders System
Here’s the unglamorous reality of CPG supply chain: paperwork. Mountains of it.
Every shipment generates three separate paper trails that need to match: the purchase order, the bill of lading, and the invoice. Except they never match. The manufacturer’s BOL says they shipped 10,000 units. The receiving dock’s copy has a handwritten correction, “received 9,950”, with someone’s initials scribbled next to it. The invoice shows the original 10,000.
Who’s liable for the 50 missing units? Without a system, you’d never know until the deductions hit.
And the deductions do hit. Retailers and distributors take money out of your payments for damages, shortages, and discrepancies. The evidence lives in a third-party portal with hundreds of line items and PDF attachments proving who caused each loss. Without tracking, you’re leaving recoveries on the table because you can’t match the chaos to your actual orders.
The real nightmare is the reconciliation gap. Money flows in from retailers. Money flows out to manufacturers. Without verification gates, you might pay a manufacturer $50K for a shipment when you only collected $45K from the retailer. That’s $5K of margin. Gone.
So I built a system that turns unstructured chaos into structured data. Python. Claude Code. Custom modules with LangGraph workflows where I needed to integrate AI. Every PO becomes a transparent object with full traceability: what shipped, what arrived, what discrepancies occurred, who’s liable, what evidence exists. When a dispute arises six months later, we don’t argue from memory. We pull up the complete audit trail in seconds.
The deduction agent is what I’m most proud of. It takes data from the portal, imports it into our system, collects evidence PDFs in parallel, and keeps everything organized. When manufacturers ask “why did you deduct $2K from our payment?”, we can show them exactly which shipment, which discrepancy, and which PDF proves it.
System Two: The Claude Code Skill
I spent four years at Oracle Cloud Infrastructure. I know how tedious working with an ERP system can be. And how much money you have to spend to get it customized for your business.
So I wrote a skill that connects Claude to the orders system. Connecting our systems to our AI console meant we could ask questions in plain English and have reports pulled and generated. Visibility into everything flowing through the business: materials, orders, accounting, payments, quality management.
Before this, you’d dig through folders, scroll through emails, log into multiple systems, waste hours, probably make a spreadsheet, and have something wrong in it anyway. Now the AI connects through a simple CLI that a human could use directly if they wanted. It folds information in and out of a system that’s well-structured, has all the data consolidated, and keeps the source records.
Need to calculate a commission payment? Done. Need to handle an unloading discrepancy because a manufacturer put too much on a truck, generate an invoice in the format the distributor will actually accept, and process it through? That becomes easy. And you can do it in English.
I dictate to the system sometimes. It’s silly. But it works.
System Three: The Web Platform
We built a suite of tools for the business. My favorite, and the most commonly used, is the Label Compliance Engine.
You wouldn’t think a simple label could hold up a product release for a quarter, but it can. Six-month slips weren’t unheard of as the value stream got stuck with too many cooks in the kitchen, waiting on legal teams to review every word, every font size, every claim.
Each round of review costs thousands in legal fees. Weeks in turnaround time. Teams across the chain sitting idle, waiting to get a product to market. And that delay doesn’t just cost time. It costs hundreds of thousands of dollars to the retailers and manufacturers who only get paid when a customer actually buys.
The engine is a LangGraph state machine that takes a product through classification, rule application, image analysis, regulatory checks, and assessment generation. What used to take weeks of back-and-forth now runs in seconds, catching issues before they ever reach legal.
We’ve automated compliance for 10 states with full legal citations. The Kosher certification validator catches ingredient issues before they become problems. The nutrition API pulls from both USDA and FDA databases to generate accurate nutrition facts.
The Pattern That Changed Everything
Here’s what I learned building all of this. It’s bigger than CPG.
When you’re dealing with FDA compliance, you can’t just trust that your AI “probably got it right.” You need to prove it worked the right way. The same is true for nuclear power, healthcare, aerospace. Anywhere lives or money are on the line.
The pattern I landed on: Observability-Driven Evaluation.
Step one: Encode business processes in agent graphs. LangGraph state machines enforce the exact workflow, making business rules code instead of documentation. Each step is a node with defined inputs and outputs.
Step two: Log everything via OpenTelemetry. Every agent action gets traced with full context. Send it to LangSmith for cloud observability, or reflect to OpenTelemetry collectors for self-hosted infrastructure. You get a complete audit trail.
Step three: Run evaluator agents over the workflow. Domain-specific AI reviews the traces: “Did this compliance check run correctly? Were all required validations performed?” A second layer of AI reviewing the first layer.
I’ve used this exact pattern in nuclear power configuration compliance, with domain-specific evaluators watching agent workflows. The architecture is the same whether you’re validating nutrition labels or safety protocols.
This isn’t just “I automated my job.” It’s a pattern that makes AI-driven business processes auditable and verifiable at scale. When regulators or auditors ask how you ensure compliance, you can show them the traces, the evaluations, the complete decision trail.
Waking Up Without Dread
The first time I noticed it, I almost didn’t believe it. I slept in until 5:30am. Woke up naturally. Refreshed. Ready to enjoy my morning workout, then get some work done before the world gets their act together.
The business was still running. The orders were still flowing. But I wasn’t the single point of failure anymore. The systems I’d built out of desperation had become the thing standing between me and the chaos. The anxiety that drove me to encode everything into software had been encoded out of my daily experience.
Last week I improved another part of the reporting automation. Another seatbelt added to the system that allows us to go faster, safer. As I watched the reconciliation dashboard generate itself, I realized something had fundamentally changed.
I’m at 20% utilization now. Down from 150%. And the automation I just deployed is pushing that toward 10%.
The recurring revenue from Always Cool Brands covers my mortgage, healthcare, and employee salaries. The business runs. Not because I’m grinding 80-hour weeks, but because I encoded my anxiety into software that doesn’t need sleep.
For the first time in three years, I’m sleeping through the night. I still wake up at 5:30am. Old habits die hard. But I don’t lie there calculating runway anymore. I get up because I want to, not because panic makes rest impossible.
What’s Next
Something unexpected happened last month: I started reaching out to friends. Real conversations, not the rushed “sorry, crazy busy, let’s catch up sometime” deflections that had become my default.
“What should I focus on?” I’ve been asking. It’s a question I couldn’t have asked when I was drowning. There was no bandwidth for exploration when survival consumed every waking hour.
Here’s what I’ve learned: we don’t automate jobs to eliminate work. We automate to take on new challenges. The patterns I built for CPG supply chain, observability-driven evaluation, agent state machines, AI-as-domain-expert, they apply everywhere complex operations need to be both automated and trustworthy.
I’m excited to explore where these patterns lead next. Nuclear power. Healthcare systems. Aerospace maintenance. Anywhere the stakes are high and the answer to “did it work?” needs to be provable.
The first step is already live: we’ve incorporated A2A into our systems. Agent-to-agent protocols that let us connect with other industry leaders who are offering their products and services in agentic-native fashion. Our compliance engine is now discoverable at alwayscool.ai/.well-known/agent.json. Other AI agents can query our label compliance services directly. No more emails. No more portals. Just agents coordinating with agents.
That’s the future. And it’s running in production.
Six months ago, I was a founder waking up in panic at 4:30am. Today, I’m someone who built AI agents capable of running a supply chain. And I finally have the headspace to think about what comes next.
The anxiety isn’t gone. But it’s encoded now. And that makes all the difference.
Always Cool AI powers the automation behind Always Cool Brands. If you’re interested in the technical architecture or want to talk about applying these patterns to your domain, I’d love to hear from you.