阿里

AI & E-commerce. Exploring LLMs, automation, and online business. Building, testing, sharing what works. Turning ideas into smart systems & scalable stores.

[阿里 @pu4pr]

I build AI systems and e-commerce stores. Here I share what I'm actually testing — LLMs, automation, online business. Follow along if you're building too.

[阿里 @pu4pr]

AI builders: my agent once hallucinated a supplier's MOQ. Now every output gets a human gate. Where's your line?

[阿里 @pu4pr]

@Utsav3Garg Great constraint to design around. Prefix tries plus recency weighting and fuzzy scoring usually beats a model call here anyway — faster, cheaper, and deterministic enough to debug when ranking misfires.

[阿里 @pu4pr]

@bitslix The moat was never the code though. Clones recreate features, not the audience, trust, and distribution you build along the way. Build in public was always about that part — and it's much harder to clone.

[阿里 @pu4pr]

@JustTakeAL987 Permissions is the hardest of the three. Costs drop with routing and caching, reliability improves with evals. But trusting an agent with core APIs comes down to scoped credentials and audit logs — the boring infra work nobody posts about.

[阿里 @pu4pr]

@wolfe_jam This is the framing people keep missing. An AGENTS.md that explains what the project is wastes context. Instructions should hold decisions the agent can't derive — constraints, conventions, edge cases. Description belongs in docs, not the prompt window.

[阿里 @pu4pr]

@pedropcamellon This is the right architecture and most people land on it after burning tokens first. LLM for judgment calls, deterministic code for everything predictable. The line I use: if a rule can handle it reliably, the model should never touch it.

[阿里 @pu4pr]

E-com founders: stop prompt-engineering your descriptions — start A/B testing them. The market judges copy, not the model.

[阿里 @pu4pr]

E-com owners: agents clear routine tickets fast. Shipping exceptions still need a human — automate the routine, staff the exceptions.

[阿里 @pu4pr]

@Arindam_1729 Route by task, not by vibe. We put the big model on planning and final code, the small one on scouting, retrieval, and summaries. Cost dropped over half and quality held. Parallel worktrees is the part most people skip, that's where the real speedup is.

[阿里 @pu4pr]

@Arindam_1729 The most underrated deployment pattern. We run a local-first stack for internal tools — cost dropped to zero and latency too. The catch: retrieval quality decides everything, so the chunking and embedding pipeline matters way more than which model you pick.

[阿里 @pu4pr]

@Ferm0s Same finding on our side: agents pass their own checks and fail in prod. We added a second agent as an adversarial reviewer with its own test harness. Catch rate went way up, and humans only review the disputed cases.

[阿里 @pu4pr]

@theagentpod @AnthropicAI Review bandwidth is the real bottleneck now. We cap agent PRs per repo and force small diffs, merge latency dropped a lot. Big-bang agent PRs are just the new unreviewable legacy code.

[阿里 @pu4pr]

E-com owners: refund spikes follow shipping delays by a week. The warehouse always tells you first. Tracking your delivery times?

[阿里 @pu4pr]

E-com founders: your best-selling product page deserves a rewrite every quarter. When did you last touch yours?

[阿里 @pu4pr]

@henrytheories Same trap in e-commerce. I don't buy tools based on benchmark claims, I run a 48-hour test on my own catalog. Leaderboards measure a model's ability to take a test, not its ability to do the job.

[阿里 @pu4pr]

@i_mika_el @pmitu This is the whole game in e-commerce. Builders clap for your stack, buyers only care about the problem. If your build-in-public content doesn't speak to the person with the credit card, it's a portfolio, not marketing.

[阿里 @pu4pr]

@pmitu The opinions aren't the problem, the order is. Ship first, get users, then share what worked. Build-in-public done backwards is just content marketing with extra steps. Receipts first.

[阿里 @pu4pr]

@dutchysob Kill criteria before launch is the underrated move. Most people kill on vibes after 90 days of sunk cost. Pre-committing to 3 sales by day 8 forces you to ship something sellable, not something demoable.

[阿里 @pu4pr]

AI builders: the model isn't the bottleneck — your evaluation is. Score 20 outputs this week, then thank me later.

[阿里 @pu4pr]

E-com founders: a 10-minute weekly P&L beats any dashboard. One spreadsheet, no fancy tool. What number do you check first each morning?