Alex Blom

Here is a quick tutorial using @aisle_sh to create an AI RAG on the fly. It works with complex documents like doc, ppt, xls and more. From there can chat and search the knowledge base with different models (GPT, Anthropic, Gemini, Grok, etc) or build more complex automations Prompts and Workflows.

Compare different prompts, models and settings with the @aisle_sh Playgrounds feature. Quick overview video attached.

Spent the last few weeks talking to mid-market and enterprise teams about why their AI rollouts are stuck. It's never the model. Every team built something that works in a demo. And then it hits the approval stage. The IC who built the prototype can't ship a change without a compliance review, and the iteration speed that made it good in the first place dies in a Jira queue. The two teams I talked to who actually shipped stopped treating governance as a separate step. They put it inside the tool the work happens in. One place, one audit trail, one shipping cycle.

Picking #3 first gets you a keynote slot. Doing #1 and #2 first is what compounds. aisle.sh/blog/the-order…

@tessi3r @businessbarista I come from an era where real product people used moleskins

@businessbarista The most important skill for anyone who wants to work in product now is taste.

AI slop should be called Human slop. It is almost always a human skill issue, resulting from bad direction, bad context, and bad taste.

Another day another Opus outage. If you ever wanted a case study on why we build reliable, retryable flows vs single shot agents.... @aisle_sh

Somebody needs to make Farming Simulator 2027 - but all the farms are just AI data centres

@brexton been telling a lot of our customers this was a bigger risk than they were imagining. need to be wary of locking in to any platform right now / there are too many ways to get screwed.

*sigh* *add OAI to list of neoclouds*

@inc1nder Am a little biased - but the model/vendor chaos is the biggest adoption barrier I see in orgs atm. You want people to experiment but not lose their history during switches; and the last thing you want to do is lock in a model/platform while it is so early.

we are here right now

Feel myself falling down the slippery X/Twitter slope again....looking back on when I was last active and man life has changed.

"Agents are deterministic or useful, never both" is one of the more confidently wrong takes about AI agents. Agents become deterministic when the space is configured: the tools they can reach, the prompts they run, the memory they read from, the system message that bounds them. That's what a Project is: a configured surface, not a chat.

A year ago I was arguing with customers that Claude deserved a real shot next to GPT. They couldn't see past "GPT Everything". The same team this week: except its Codex vs Claude. Models are going to keep leapfrogging for a while. What surprises me is how many teams have been through this cycle two or three times now and still lock into a static AI worldview at the first opportunity.

@jpwallhorn Have been working on this stuff for a while as well and fully agree. Sometimes I describe to customers that instead of agents they want a 'deterministic ai flow'.

Been building production AI systems at Suppli, and it's clear: prompt engineering is just the tip. Most of the work is in the harness - the systems that make models reliable, cost-effective, and ready for real users. Here's what I've learned matters most when you're past prototypes: - Orchestration and state: Models need solid plumbing for tools, error recovery, and loops. Without it, you're debugging chaos. - Caching strategies: Prompt caching for repeats, semantic for similar intents. Pick wrong, and costs spiral. - KV cache handling: Essential for inference at scale. Techniques like PagedAttention keep latency and bills in check. - Decoding and quantization: Speed up with speculative guesses or model compression. Know your workload to choose. - Output validation: Schemas fail; build retries, fallbacks, even human checks. - Evals mix: LLM judges for quick checks, humans for truth. - Cost tracking: Break it down by feature to spot the leaks. - Guardrails: Limits on loops and actions to prevent runaways. - Observability: Trace every call like it's code - metrics, logs, the works. - Routing: Switch models dynamically, with fallbacks for uptime. - Fine-tuning decisions: In-context for speed, tuning for efficiency - but only after basics are solid. AI engineering is systems work with probabilistic twists. Nail these, and your agents ship value without the headaches.

Your AI starts from zero every morning. every chat resets. every prompt rebuilds context from scratch. Projects on Aisle is the fix.

Projects bundles together instructions, tools, and knowledge into a workspace for you, your team, and your agents. Not exaggerating, but this has been the single biggest shift in the way I work that I can think of. aisle.sh/features-proje…

🧵Projects bundles prompts, workflows, scripts, MCP connectors, and memory folders as tools in Aisle. Configure them once, and every conversation in the project starts with the full context in place. aisle.sh/features-proje…

Been talking AI with legacy customers all week. The shift is wild: 18 months ago "AI automation" was novel and now it's table stakes. It strikes me that few are thinking about maintenance or durability of their automations. The new models make this easier than ever, but AI debt will bite your ass just as hard as tech debt if you're not careful.