Avi

@ryolu_ @cursor_ai There is no moat in just your code anymore. Anyone can create anything super fast now.

@arxivlens is three runtimes pretending to be one product: - Web + jobs: Core + Hangfire + Razor + Tailwind - Mobile: Flutter + Hive - ML pipeline: Django PostgreSQL is the only shared contract. Three teams could own three repos. One person can't.

Bare-metal k8s lesson on @arxivlens: Service type=LoadBalancer reports 'pending' forever without a cloud provider. So the deploy script falls back to printing the NodePort URL. Hangfire dashboard now lives at :31120. The fallback ships before the platform team does.

Features I shipped this month: - Bulk export to CSV - A 'cancel' button - Better error messages - Sane defaults Zero of them used AI. All of them moved the needle.

The hardest LLM bug is the one that happens 1% of the time. The most expensive LLM bug is the one users notice. The most embarrassing LLM bug is the one your investor finds.

@arxivlens ingests papers from arXiv, bioRxiv, medRxiv, PubMed, and HuggingFace. Each source has its own IDs, its own DOI coverage, its own update cadence. Some publish XML. Some ship JSON. One has a 2003-era RSS feed. The hard part isn't fetching. It's normalizing.

5 citation formats on @arxivlens mobile, all from one Paper model: - BibTeX via a custom builder - APA / MLA / Chicago / IEEE via per-style templates - Author disambiguator handles 'et al.' thresholds - Bottom sheet is DraggableScrollableSheet with copy + share intents

@arxivlens pod sizing was a lesson: Web (ASP.NET Core SSR, CPU-bound): 45Gi / 10 CPU Scraper (Hangfire + HttpClient, I/O-bound): 8Gi / 2 CPU 200 concurrent papers per scraper pod, idle CPU at 30%. async/await all the way down. Provision by latency profile, not by feel.

PM: 'Can the AI just figure it out?' Me: 'Figure what out?' PM: 'You know. The thing.' Six weeks later we shipped a dropdown. The hardest part of AI products is defining the problem.

Deploy script on @arxivlens compares image digests before restarting: - buildx imagetools inspect → new SHA - kubectl get deploy -o jsonpath → running SHA - rollout restart only when they differ In-flight Hangfire jobs survive. ~80 lines of PowerShell. No ArgoCD.

Six months ago I thought the bottleneck was the model. Now I think the bottleneck is everything else: data, UX, distribution, support. The model is a commodity. The product is the work.

Replaced my embedding model with one half the size. Search quality dropped 0.4%. Latency dropped 60%. Bill dropped 80%. Half of optimization is being willing to lose the 0.4%.

Switched my structured output from JSON-mode to XML tags. Parse failures dropped from 4% to 0.2%. Same model. Different format. The biggest LLM wins always look like nothing.

Watching Big Tech announce 'AI-first' the same way they announced 'mobile-first' in 2014 and 'cloud-first' in 2009. By the time they ship, indie hackers are already on the next thing. Being first to announce isn't being first to ship.

Two years building AI products, one lesson: Intelligence is cheap. Reliability is expensive. Trust is priceless. Most founders optimize the first. Users pay for the third.

Build the version with no AI first. Then ask whether AI actually makes it better. The answer is 'no' more often than indie Twitter wants to admit.

Your users don't care which model you use. Your users don't care how you prompt it. Your users don't care about your eval suite. Your users care if the button does what it says it does. The moat is the button.

Investor: 'Where's your moat?' Me: 'Distribution.' Investor: 'No, your AI moat.' Me: 'My AI is the same as everyone else's. Distribution is my moat.' Investor stopped emailing. Probably for the best.

Hot take: The best prompt engineering technique is having fewer features. Every feature is a new failure mode for the model. The AI products that win ship one thing that works, not ten things that sort of work.

@Railway Why do people even use goggle cloud at this point

Google Cloud has blocked our account, making some Railway services unavailable. We have escalated this directly with Google. The Railway Platform team has since confirmed access to Google Cloud and is working on restoring access to all workloads. We have access to some of our Google Cloud–hosted infrastructure and are working to restore the rest of the service. We apologize for the disruption.

Wrote a 1,200-token system prompt with 17 'always do this' rules. The model ignored 12 of them. Cut the prompt to 280 tokens. The model ignored 2. Less prompting, less ignoring. There's a lesson in there somewhere.