Olu Akele / Portfolio 2026Now building: ShelfReadyPrefer the technical details? Flip this.

Built in publicOne post per episodeRSS

Build log

What I decided, what I built, what went wrong, and the result. One write-up per step.The product call, what I built, what broke, and the number. Every episode, written up.

  1. Portfolio
    Episode 06

    One site, two readers: adding a Simple / Technical switch

    My wife said my website was too technical. So I gave it a switch: plain English by default, and the technical details one tap away. No number changed.My wife read the portfolio and said it was too technical. Instead of choosing between her and an engineer, I gave the site a switch: Simple by default, Technical one tap away, with every number unchanged.

    PM takeaway
    When two audiences need different things, don't average the copy. Pick a default for the reader you're losing, and give the other one a fast way in.
    Build takeaway
    Render both versions and let CSS pick, set before the first paint. No flash, no hydration mismatch, and the page still reads with JavaScript off.
  2. ShelfReady
    Episode 07

    A public AI shopping demo that can't run up a bill or show an error

    I put a live AI shopping assistant on a public page, then planned for the two ways a public demo goes wrong: it costs too much, or a visitor sees an error. I solved both.I put a live AI shopping agent on a public page, then designed for the two ways a public demo fails: it costs too much, or a visitor hits an error. Both ended up with a hard answer.

    PM takeaway
    A public demo is a product with a budget. Decide the worst-case spend and what a visitor sees when you hit it before launch, not after the first bill.
    Build takeaway
    Test protection in the deployed bundle, not in a script. A variable named `global` turned the rate limiter off, and only once the code was bundled.
  3. ShelfReady
    Episode 06

    360 agent shopping sessions, 0 wrong products, and a finding I didn't expect

    I had an AI shopper do 60 shopping tasks, 3 times each, on the messy store and the fixed one. It never bought the wrong item, and the biggest improvement came from fixing its search, not the product data.I had an AI agent run 60 shopping tasks, 3 times each, on the messy catalog and the fixed one. It never bought a wrong product, and the biggest gain came from fixing its search tool, not the data.

    PM takeaway
    Measure the whole path to the customer, not just the part you built. The catalog was my product, but the agent's search tool turned out to be the bigger lever.
    Build takeaway
    Run evals more than once, and audit the harness as hard as the system. At 40 tasks one task is 2.5 points, and a harness bug quietly made "before" look worse.
  4. Portfolio
    Episode 05

    Building my portfolio in public, in five episodes

    I built this website the way I build products: in small steps, with a plan before each one and notes along the way. It went from a "coming soon" page to a full site with a build log, a case study and proof of results.I built oluakele.com the same way I build products: small slices, a plan before each one, and a written log, from a coming-soon page to a site with a build log, a case study and proof that links to its sources.

    PM takeaway
    Ship the smallest public version first, then let real feedback steer. The first color refresh read as "LinkedIn" to me in the preview screenshots, so it changed before it ever shipped.
    Build takeaway
    Measure where the measurement is reliable. Local Lighthouse on a busy laptop swung from 57 to 87 on the same page, so performance gets checked on the deployed site.
  5. ShelfReady
    Episode 05

    Claude shopped my store and came back with a checkout link

    I connected an AI assistant to the store. It found the closest match to a shopper's request and came back with a working checkout link, and the test uncovered a data problem my checks had missed.I gave the store four MCP tools and connected Claude. It found the closest match to a shopper's request and returned a working checkout link, and the test surfaced a data bug the audit couldn't see.

    PM takeaway
    Put the product in front of a real agent early. A live session found a spec error that every score had missed.
    Build takeaway
    Grounding checks must define what counts as evidence. A URL handle isn't product text, and treating it that way let a bad spec look proven.
  6. ShelfReady
    Episode 04

    101 fixes, one human reviewer, and a store score from 94.1 to 98.3

    I built a tool that suggests fixes to product details, with a person approving each one. Then I approved the fixes myself and checked the store again.I built a fixer that proposes catalog fixes, a queue where a human approves them, and write-back to Shopify. Then I reviewed the whole queue myself and re-audited.

    PM takeaway
    A confident wrong spec costs more than a missing one. So the fixer drops anything it can't prove and hands the gap to the merchant.
    Build takeaway
    Shopify has no transaction across mutations, so all-or-nothing writes are a fiction. Record each change's own result.
  7. ShelfReady
    Episode 03

    My audit said 89% of the store was ready. That was the bug.

    My first check said almost the whole store was ready for AI shoppers. That was wrong, so I fixed how the check grades products before trusting it.The first full audit ranked products correctly but called almost the whole store agent-ready, so I changed how the bands work before trusting the baseline.

    PM takeaway
    A score can rank things correctly and still mislead. Check the headline number against the question it's supposed to answer.
    Build takeaway
    Make the model quote its evidence, and check the quote in code. It fixed a lenient judge and gave every finding a source.
  8. ShelfReady
    Episode 02

    Seeding 150 products, a third of them broken on purpose

    I made a practice store with 150 products, broke 50 of them on purpose, and then my laptop froze halfway through loading them.I generated a demo catalog with 50 deliberately messy products, and then my laptop froze halfway through seeding it into Shopify.

    PM takeaway
    Build the answer key before the test. Because I injected every defect myself, every audit and eval from here on is scored against known answers, not a model's opinion.
    Build takeaway
    Keep resume state where the work lands. Storing a hash on each Shopify product made the interrupted seed restart cleanly, with nothing lost or duplicated.
  9. ShelfReady
    Episode 01

    A green status page before a single feature

    Step one had a single goal: a live page showing the store, the database and the AI all connected. It took about three hours and one confusing first launch.Episode 01 had one finish line: a live URL showing Shopify, the database and the model key all connected. Getting there took about three hours and one confusing first deploy.

    PM takeaway
    Define done as something you can open in a browser. "Live URL, all green" ended every scope debate before it started.
    Build takeaway
    Make the health page name what's missing. A red check that says which setting is absent turned a blank deploy error into a checklist.
  10. ShelfReady
    Episode 00

    Before any code: six accounts, a $0 stack, and three gotchas

    Before writing any code, I set up every free service the project needed, and hit three surprises that would have derailed the first step.I set up the full ShelfReady stack on free tiers before writing a line of app code, and hit three surprises that would have cost me the first episode.

    PM takeaway
    Decide the cost ceiling before the architecture. A $0 budget made the model choice for me, and I wrote that down as a decision instead of an accident.
    Build takeaway
    Verify every credential with a real call the night before. Two of mine behaved differently than the docs said.