FullStack Developer, passionate about technology, startups, and building products
Medellín, Colombia
Joined March 2019
- Tweets117
- Following259
- Followers52
- Likes135
I've been using a Linux PC at home as my remote dev machine for the past few months, with my MacBook Air as a lightweight client. I've really enjoyed being able to delegate work and close my laptop while it keeps running.
But some workflows still brought work back to the Air, especially iOS simulator testing and Codex computer use on macOS. So I added a used MacBook Pro as another remote node.
I wrote about the setup and included a demo of the agent testing my app on web and iOS, then sending me a recording:
dev.to/juandastic/my-coding-…
A recent video from @theo about his Linux dev fleet resonated with an experiment I’d already started: moving my coding agents and dev environment off my MacBook and onto Linux.
I turned an old gaming PC into a Linux dev box. Then came the real wow moment, from a restaurant, with my MacBook off, I requested an app change through my Android phone.
The agent worked on Ubuntu at home, and I tested the result on the same phone.
This video shows that loop: Tailscale + wireless ADB + Metro + Codex.
Full story:
dev.to/juandastic/i-updated-…
I’ve been experimenting with personal agents like Hermes, and it made me wonder how many AI products are really skills wrapped inside standalone apps.
I built an AI nutrition tracker to understand meals from text or photos, estimate calories and macros, and save everything to a Google Sheet I control.
To make it reusable, I originally built a Python backend, web app, Telegram bot, OAuth flow, database, and user system.
Eight months later, I recreated its core workflow as a plain-text skill in around 15 minutes and two prompts.
This connected with two ideas I had recently encountered:
@theo arguing that capable models and tools can reduce some products to structured instructions:
youtube.com/watch?v=xUnRQ9vL…
And Julien Avezou writing about building for agents, not just humans:
dev.to/javz/start-building-f…
That experiment made me question how many other AI products could be compressed into simple agent instruction.
Today, I would start with the skill and only build the standalone product when the workflow proves it needs one.
dev.to/juandastic/i-built-an…
Juan David Gómez retweeted
React Native on iOS doesn't support ReadableStream. When users switch apps mid-response, the AI keeps generating to nothing. This dev solved it by making the server keep going after disconnect and persist the full response to the DB.
{ author: @juandastic }
dev.to/juandastic/i-built-a-…
12 years building web apps. Zero mobile experience. Last month I needed an iOS app for my side project so I just built one. Took 3 days.
The real unlock was not React Native or Expo. It was having AI in the workflow. It did not write the app for me but it removed the friction of learning a new platform. I moved at the same speed I move on the web.
Then iOS broke everything. Turns out when your user switches to WhatsApp mid-response, iOS kills your network request. The AI keeps generating server-side, burning tokens, but the response never arrives.
Had to rethink the whole streaming architecture to handle disconnects gracefully.
Full writeup: dev.to/juandastic/i-built-a-…
Building AI side projects is fun until you have to pay for it
I built Synapse, an AI companion for my wife with a memory system that makes Gemini know her life, her patterns, her emotional triggers. She uses it daily.
Two weeks ago, I connected PostHog to track costs. $24. One session alone: 28 messages, $2.42.
Every message sends ~30K tokens of context. 80-90% of those tokens are the exact same compiled knowledge, repeated every turn.
Most AI providers offer automatic caching. Send the same prefix enough times, and they might cache it for you. But you have no control. One small change in the prompt (like a datetime updating every turn) breaks prefix matching. You never know if it is working.
𝗚𝗲𝗺𝗶𝗻𝗶'𝘀 𝗲𝘅𝗽𝗹𝗶𝗰𝗶𝘁 𝗰𝗮𝗰𝗵𝗶𝗻𝗴 𝗔𝗣𝗜 𝗶𝘀 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁.
You create a cache resource, get a name back, and reference it on every request. Guaranteed hit. 75% cheaper on cached tokens. You decide what gets cached and you can verify it is working.
I separated the stable knowledge compilation (~25K tokens) from the volatile parts of the prompt, cached the compilation after hydration, and referenced it by name on every message.
The client sends both the cache_name and the full compilation on every request. If the cache is hot: 75% savings. If it expired, the server inlines the compilation at full price. The user never notices.
𝗧𝗵𝗲 𝗻𝘂𝗺𝗯𝗲𝗿𝘀
Before: $0.017-0.039 per generation After: $0.0088 per generation
That $2.42 session? Would cost ~$0.25 now.
Same knowledge graph. Same memory quality. Just a lot cheaper to remember.
Full breakdown: dev.to/juandastic/my-ai-send…
My wife is a psychologist. She started using an AI between therapy sessions but got frustrated that it forgot everything.
So I built her one that doesn't.
297 messages in 15 days. Here's what I learned 🧵
Synapse uses a knowledge graph instead of a flat memory list.
With the right therapeutic frameworks (ACT, DBT, Polyvagal Theory) and all her context, she found her ideal AI personal companion
synapse-chat.juandago.dev
In the last 15 days of real production data:
→ 297 messages → 15.2M tokens processed → 100% daily active usage → Peak: 69 messages in one day
It's fully open source. The full product story:
dev.to/juandastic/my-wife-se…
You can't ask "find the most connected entity and give me all its facts" if the facts live in a different store than the structure.
I'm calling this "context blindness" and for a personal companion it matters more than retrieval scores.
Full benchmark with architecture diagrams, code, and results (open source):
dev.to/juandastic/i-benchmar…
I built a bidirectional sync between a Neo4j Knowledge Graph and Notion for the @ThePracticalDev Challenge with @NotionHQ
AI memory is usually a black box. Now, it's a fully editable, dynamically generated Notion workspace.
Here is the architecture
dev.to/juandastic/full-circl…
Juan David Gómez retweeted
Context windows are expensive resources.
This dev demonstrates how to tame a 120k token prompt by implementing a deterministic GraphRAG approach for more efficient memory scaling.
{ author: @juandastic }
dev.to/juandastic/scaling-ai…
1 Million token context windows are a luxury, not an architecture.
My AI side-project hit a wall when the system prompt bloated to 120,000 tokens per message.
Here is how I fixed it using a Knowledge Graph, a "Waterfill" budget, and Agentless GraphRAG. 🧵👇
To get those details back, I built Deterministic GraphRAG.
I hate Agent tool-calling (adds 3-5s latency). Instead, a straight-line Hybrid Search runs in <1s. It checks what's already in the prompt and only injects missing facts. Zero redundancy.
The Result: Token usage collapsed back to my 40k limit. Latency stayed under 1s. The AI still feels like it knows everything.
I wrote a full deep dive on @ThePracticalDev with the architecture
dev.to/juandastic/scaling-ai…