AI without design is a slot machine.
Sometimes you win. Every pull still stacks tech debt and burns tokens, money, and yes, literal water. The gold rush forgot the oldest truth in software: boring, predictable code is cheap, faster than AI, and runs the same way every time.
Tokenmaxxing fails for the same reason throwing money at any problem fails. Spend without design is a more expensive way to be wrong.
Deterministic core. Agentic edges.
The most robust production AI systems share one shape. Plain code and rules handle the 90% of cases that are predictable, debuggable, and nearly free. AI handles the 10% where judgment actually earns its tokens. Better-placed AI is better.
The payoff is a system you can own. Reproducible. Debuggable. Predictable cost. On the job-search pipeline I built for myself, that split delivered 80% time savings with full visibility into every step.
Departures
A fail-closed AI agent toolkit for live customer dataIf your AI writes to the wrong customer's account, would anything stop it?A green light is not proof
In productionConnecting a language model to a live, multi-tenant production API is genuinely dangerous. An agent can write to the wrong customer's data, assume an operation exists when the API silently does nothing, or get a "200 OK" back while nothing actually changed.
I designed a credential model binding every agent profile to one tenant at mint time (mismatch, and every operation blocks), a 47-entry capability ledger routing each operation to its safest path, and write paths that default to dry-run, demand an explicit expected-tenant argument, and re-read and diff after every write. High-risk config changes halt for human review. Shipped as 10 dated releases over roughly three weeks; now live in production with non-engineering staff, rolling out to enterprise and AI-forward clients.
The scariest AI failure is not a wrong answer. It is a success message. An agent connected to live customer data can write to the wrong tenant, assume an operation exists when the API silently does nothing, or collect a "200 OK" while nothing actually changed. Any of those reaching a customer costs more than the AI ever saved.
So I made failure loud. A credential model binds every agent profile to one tenant at mint time: mismatch, and every operation blocks. Every write defaults to dry-run, demands an explicit expected-tenant argument, and re-reads and diffs after the write, because a success-shaped response is not proof of a change. A 47-entry capability ledger routes each operation to its safest path, and high-risk config changes halt for human review. Shipped as 10 dated releases over roughly three weeks; now in production with non-engineering staff at a B2B PIM/DAM SaaS platform.
Most AI tools show you a green light. An agent on a live, multi-tenant production API can get a "200 OK" while nothing actually changed, do the right thing to the wrong customer, or assume a capability that silently does not exist. It looks healthy from the outside the entire time. Verifying that the answer is actually correct is the job.
I built this toolkit because I do not trust my own systems either: every agent profile bound to one tenant at mint time (mismatch blocks everything), every write defaulting to dry-run and re-reading and diffing afterward, a 47-entry capability ledger where every entry carries a last-verified date and the probe evidence behind it. High-risk changes stop and wait for a human. Shipped as 10 dated releases over roughly three weeks; in production with non-engineering staff. One number I will not claim: the business impact was never instrumented, so there is no savings figure here. When I have one, it will be measured.
Owning the product surface at a PIM/DAM platformCan you tell which AI feature bets are real before the devs burn a sprint?The paper trail, including the part that is not mine
Verified via JiraTwo years as a de facto technical product manager: designing application functions with clients and stakeholders, mapping the platform's MCP and API surface, and feeding a steady stream of specs and bug reports into the dev team's pipeline. I also led the MCP server pilot with enterprise customers and took it live.
I specified, championed, and verified these features. The developers wrote the code. That distinction matters, and I keep it honest.
Most product roadmaps run on the loudest voice in the room. Over two years at a B2B SaaS company, my job was to replace that with a paper trail: sit with clients and stakeholders, design the application functions they actually needed, map the platform's MCP and API surface, and hand the dev team specs they could build without guessing. I also led the MCP server pilot with enterprise customers and took it live.
The record, verified live against Jira: 250+ product tickets over two years, 72 feature proposals with 35 shipped to production, 165 bugs reported and 116 fixed. I specified, championed, and verified these features. The developers wrote the code. That distinction matters, and I keep it honest.
Anyone can claim product instincts. Here is a record you can audit instead, verified live against Jira: 250+ product tickets over two years at a B2B SaaS company, 72 feature proposals with 35 shipped to production, 165 bugs reported and 116 fixed, and an MCP server pilot led with enterprise customers and taken live.
The honest part: I specified, championed, and verified these features. The developers wrote the code. Most portfolios blur that line. Blurred lines are also how AI projects fail, which is rather the point of this site.
Shopify to enterprise PIM, zero data errorsMoving 33,202 products is easy. Not corrupting any of them is the job.Zero data errors, and why that number is the whole point
LandedAn e-commerce catalog needed to move off a Cloudflare-protected Shopify storefront and into an enterprise PIM, cleanly. I scraped 33,202 products through the storefront with Playwright after the standard requests approach got blocked, handling pagination, resume, and cookie expiration. Then I normalized and renamed 11,963 SKUs with parent grouping, duplicate detection, and handle normalization.
Deterministic validation caught the errors before they reached production, and the AI only touched the edge cases a rule couldn't cover.
Messy product data is where most e-commerce AI projects quietly break. This catalog had to move off a Cloudflare-protected Shopify storefront into an enterprise PIM without dragging its mess along: 33,202 products behind a bot wall, 11,963 SKUs needing normalization, duplicates and broken handles hiding throughout.
Playwright got through where the standard requests approach was blocked, handling pagination, resume, and cookie expiration. Deterministic validation with parent grouping, duplicate detection, and handle normalization caught the errors before they reached production, and the AI only touched the edge cases a rule couldn't cover. The catalog landed with zero data errors.
"Zero data errors" is usually a claim nobody checked. Here it was the acceptance bar: 33,202 products scraped off a Cloudflare-protected Shopify storefront with Playwright (pagination, resume, and cookie expiration handled), then 11,963 SKUs normalized and renamed with parent grouping, duplicate detection, and handle normalization.
The method is the message: deterministic validation caught the errors before anything reached production, and the AI only touched the edge cases a rule couldn't cover. That split is why the error count is a number I can defend instead of a vibe.
Three fixes in 20 minutes. Free.
You talk about your business. I listen for the one workflow that eats the most of your week: invoicing, follow-ups, scheduling, reporting, whatever hurts. You leave with three fixes ranked by effort and impact: one you can do yourself this week, one worth automating, and one to skip entirely.
No pitch, no jargon, nothing to install, no access to grant. Nothing you tell me goes into any AI without your written ok. If a fix is worth building, the 48-hour micro-sprint below puts $0 on the table.
The evidence ledger
Every number below is measured, not marketed. 48 billion tokens of tuition on the road from vibe coding to agentic engineering.
I design agentic shells. Then I keep them healthy.
An agentic shell is an AI layer around your production app: modular, isolatable, testable workflows that scale with the task. Your core stays deterministic and cheap. The shell handles the judgment calls you can see, test, and afford.
Scout, Plan, Build: agents on parallel git branches, fully auditable and reversible. I test three approaches at once, keep the winner, and throw the rest away, so you don't pay for dead ends. Token spend is an engineering variable, not weather. Tuning it is part of the service.
APrompt pack + eval harness
Your prompts treated as versioned software with regression tests. Baseline versus v1 scorecard on 20 to 50 real examples.
BMini-RAG with citations
Grounded answers over your docs and policies, with quote spans and citation coverage so the answers come with receipts.
CAgentic flow stub
Plan, Act, Check loops with tool adapters and a JSONL audit trail. Fully reversible via git worktrees.
The 48-hour micro-sprint
First 8 hours free, if you send prerequisites within 24 hours of my reply. We agree on acceptance criteria upfront. I ship a demo plus artifacts. Miss the agreed deliverables and you pay $0. Hit them and we continue at an hourly or fixed scope.
"Alex was great to work with. The whole process felt easy and smooth. He asked smart questions, gave clear direction, and knew exactly what he needed from me. Because of that, he delivered a quality product I can use right away. I see a lot of future work with him because he knows his craft. I'm grateful for his work on my project and I highly recommend him."
I came up in sales and implementation before I came up in AI.Most AI projects do not die in the model. They die in the rollout.I have probably wasted more time on AI that did not work than anyone you will meet this year.
That's why the AI I build actually ships. 70+ SaaS implementations across two companies. 45+ enterprise HCM rollouts and $2.1M+ in implementation revenue I owned personally at a payroll and HR software company. Then a B2B SaaS sales seat, where I was promoted from SDR to AE in 11 months and hit 116% of quota at an 18.5% close rate.I know because rollouts were my job before AI existed as a job. 70+ SaaS implementations across two companies. 45+ enterprise HCM rollouts and $2.1M+ in implementation revenue I owned personally at a payroll and HR software company. Then a B2B SaaS sales seat, where I was promoted from SDR to AE in 11 months and hit 116% of quota at an 18.5% close rate. I learned what breaks, and why, years before I could automate any of it.I measured it: 48+ billion tokens across Claude and Codex, computed from my own session logs, on the road from vibe coding to agentic engineering. That tuition is why my advice is cheap, and why nothing I recommend pays me a commission. The other reason the AI I build ships: 70+ SaaS implementations across two companies, 45+ enterprise HCM rollouts, $2.1M+ in implementation revenue I owned personally, and a promotion from SDR to AE in 11 months at 116% of quota and an 18.5% close rate.
I'm not a traditional engineer, and after 70+ implementations I know exactly what breaks and why. The best AI practitioners aren't syntax wizards. They're coaches and communicators who can say precisely what they want from a team, whether that team is people or machines modeled on our own language. That's the skill that transfers. Clarity.
These days I run AI Ops at a B2B SaaS company and consult through Arkai Growth. On the side I coach tennis for kids on the autism spectrum through ACEing Autism, chase physics and math rabbit holes for fun, and fall asleep to AI videos. DePaul University, B.S. Health Sciences. Former pre-med who now diagnoses systems instead of patients.
Got an expensive problem? Let's talk.
Need someone who can ship working automation in 30 days? If your data is a mess, your AI hallucinates, or you need workflows that run themselves, let's build something you can measure in 48 hours.