Adham Mahmoud

👨🏻‍💻Software engineer & AI architect
🤖Helping business owners scale with AI
🚀Helping tech & non-tech

[Adham Mahmoud @bulidwithadham]

Hi, I'm Adham. I'm obsessed with building and fixing things — 12 years of writing code.

I'll help you with tech, AI, and business, building and growth.

Drop Hi,
And let’s connect

[Adham Mahmoud @bulidwithadham]

@Dan_Jeffries1 Formal verification won't fix software built around misunderstood business rules. I've learned to make the math deterministic first, then add rules, then AI. Verifying the wrong specification just makes the mistake reliable.

[Adham Mahmoud @bulidwithadham]

@haider1 I care more about cost per completed coding task than cost per token. A smaller model that needs more retries or human fixes can end up being the expensive option.

[Adham Mahmoud @bulidwithadham]

@kimmonismus @AnthropicAI This is why I don't want critical agent workflows tied to one model subscription. Context, routing and fallback paths matter when capacity becomes unpredictable.

[Adham Mahmoud @bulidwithadham]

@theo After 12 years writing code, I think knowing what not to build is harder than building it. AI just makes it easier to avoid that decision.

[Adham Mahmoud @bulidwithadham]

@chalaska The model improving is only half the story. A design system needs to be usable as context, not just a pile of Figma frames. The way you expose components and constraints to the agent matters.

[Adham Mahmoud @bulidwithadham]

@paraschopra This only works when the hard part is implementation. If the difficulty is understanding the business, waiting won't help much. I've found that company context and rules take more work than the agent itself.

[Adham Mahmoud @bulidwithadham]

@nikks_techie Fine-tuning won't keep your company's changing facts current. I use retrieval for business context and live tools for store data. Teaching a model how to respond is different from giving it today's numbers.

[Adham Mahmoud @bulidwithadham]

@karrisaarinen I see this with agents at work. They'll keep optimizing the requested action even when the action itself stopped making sense. I started treating "should we do this at all?" as a separate decision before execution.

[Adham Mahmoud @bulidwithadham]

@theo The interesting failure mode is three reviewers sharing the same blind spot. I'd keep deterministic checks as the hard gate and use the council for risks the tests didn't anticipate.

[Adham Mahmoud @bulidwithadham]

@filippkowalski I use a similar boundary when building agents: automate detection, keep consequential fixes behind approval. Knowing something broke and knowing the safe way to fix it are very different problems.

[Adham Mahmoud @bulidwithadham]

@kentcdodds I disagree with the idea that all SaaS is in trouble. Apps that only wrap data in a dashboard are exposed. Apps that own reliable workflows, permissions and business logic become even more important.

[Adham Mahmoud @bulidwithadham]

@emollick I'd separate finding problems from judging whether they matter. AI can catch inconsistencies at scale, but scientific review also needs judgment about assumptions and what a result actually proves.

[Adham Mahmoud @bulidwithadham]

@levie Agents make weak access controls much more expensive. My rule when building them is simple: read access first, explicit approval for writes. A useful agent doesn't need unrestricted access.

[Adham Mahmoud @bulidwithadham]

@mardehaym I'd add one more thing: what happens when the agent is wrong? For agents I build, tool permissions and approval rules matter as much as ownership. A hosted agent can still cause real damage.

[Adham Mahmoud @bulidwithadham]

@BenLesh @DavidKPiano State machines are underrated in agent systems too. If a workflow has a fixed set of valid transitions, I don't want an LLM deciding what state comes next.

[Adham Mahmoud @bulidwithadham]

@testingcatalog I connected Claude to my store, ads and shipping through MCP. Credential handling is only one layer. The harder part is deciding what an agent can read, what it can change and what needs approval.

[Adham Mahmoud @bulidwithadham]

@GergelyOrosz I'm not convinced AI-native engineers will automatically cost more. The expensive skill is knowing where AI shouldn't decide. I've learned to keep calculations deterministic and use agents around them.

[Adham Mahmoud @bulidwithadham]

@rauchg The frontend getting cheaper to create makes the backend more important, not less. Data contracts, permissions and state still need to survive long after the agent that built the UI is gone.

[Adham Mahmoud @bulidwithadham]

@claudeai Risk checks shouldn't be the only boundary. An agent can make a perfectly valid API call that's still a bad business decision. I prefer hard rules around spending and actions before the model gets involved.

[Adham Mahmoud @bulidwithadham]

@TheGeorgePu Comparing token spend to engineer salaries misses the point. I'd compare cost per completed, verified task. An agent making 100 bad calls isn't replacing an engineer.