gllmdev Profile Banner
Guillem Profile
Guillem

@gllmdev

Followers
91
Following
570
Media
15
Statuses
368

Software Architect @kirha_ai | Building the future of AI data access

Barcelona
Joined March 2020
Don't wanna be here? Send us removal request.
@kirha_ai
Kirha
4 days
What people pay for is structure, freshness, and accountability. We benchmarked Kirha vs Web Search on domain-specific queries. Results: 87/100 vs 61/100 95% fewer tokens https://t.co/eCkemnfd0P
124
128
231
@kirha_ai
Kirha
1 month
This is not advice. It’s an obligation: stop giving your models a bad data trip 🤖 With traceable context from premium providers, you reduce your LLM’s hallucinations while tracking when it goes wrong. Use verified, real-world context. Build responsibly:
0
9
25
@kirha_ai
Kirha
1 month
Our Monthly Recap Blog is live! X402 dynamic pricing, Neo4j migration, new CyberSec/OSINT vertical, deterministic tool planning, enterprise-grade privacy, full API tracing and more. Kirha is leveling up fast. Read the full breakdown 👇 https://t.co/OgxTX7706R
Tweet card summary image
kirha.com
November 2025 updates: X402 integration on Coinbase Bazaar, Neo4j migration for MCP metadata, new CyberSec/OSINT vertical, deterministic tool planning, enterprise privacy features, and B2B traction.
8
9
26
@rektorship
Pierre Hay
1 month
one mcp a day keeps the doctor away. welcome @AbuseIPDB to the CyberSec vertical. 1.5M addresses added in the last 24 hours. Build your cybersec AI agents with @kirha_ai
3
1
5
@gllmdev
Guillem
2 months
https://t.co/T8wHmgaLQX come and play, we deployed a beautiful graph representing our data based on our recent neo4j migration @kirha_ai
1
6
13
@kirha_ai
Kirha
2 months
AI agents need to pay for data dynamically. But today, most payment systems assume flat pricing or subscriptions. We just published how Kirha “hacked” X402 @base (Coinbase’s micropayment protocol) to support dynamic pricing for AI agents. Here’s the recap: 🔗
3
7
40
@kirha_ai
Kirha
2 months
Kirha is now on Telegram! Join our private community of AI builders to get the latest development updates and share insights. Connect, build, and grow with Kirha! https://t.co/hQ0QyLuN9T
1
6
18
@gllmdev
Guillem
2 months
beautiful to see this in a graph. mcps, verticals, tools, relationships all connected. makes tool discovery intuitive
@rektorship
Pierre Hay
2 months
@gllmdev is exploring Neo4j to store relationships between our ever-growing catalog of MCP tools. goal is to pre-filter the most relevant ones to save context for our LLM planner. > models can plug @kirha_ai with minor impact on the TTFT. better doesn't always mean slower
0
0
2
@gllmdev
Guillem
2 months
finally we got compact in claude chat
0
0
0
@gllmdev
Guillem
2 months
my current life can be defined as: spending 80% of my time on infra that is invisible to the user but would immediately cause a company ending panic if it went down. the other 20% is drinking coffee.
1
0
4
@gllmdev
Guillem
2 months
we're building the future of data for AI
@kirha_ai
Kirha
2 months
Our new teaser is live. One minute to reveal what Kirha really is: the private data layer powering the next generation of LLMs and agents. Fresh, verified, real world data on demand. Deterministic. Reliable. Built for scale. This is only the beginning. Our roadmap is packed
0
0
1
@gllmdev
Guillem
2 months
must read
0
0
1
@gllmdev
Guillem
2 months
anthropic added native http to mcp since then. but the pattern works for any stdio program. legacy servers. custom middleware. anything you cant modify. maked it public today at @kirha_ai https://t.co/qkNyQBA2SV
Tweet card summary image
github.com
A Go application that wraps stdio-based programs and exposes them via HTTP. Originally built to enable remote access to MCP (Model Context Protocol) servers when only stdio transport was available....
0
0
1
@gllmdev
Guillem
2 months
docker makes it practical. wrap any stdio program: FROM iowrapper RUN npm install your-mcp-server ENV WRAPPER_COMMAND=node ENV WRAPPER_ARGS=/path/to/server EXPOSE 8080 instant remote deployment
1
0
0
@gllmdev
Guillem
2 months
result: true parallel request handling over serial transport. each http request waits only for its response. no queuing. no polling. just channels doing their thing
1
0
0
@gllmdev
Guillem
2 months
separate goroutine reads stdout continuously. parses response json. finds our uuid. looks up tracker. swaps uuid back to original id. sends to waiting channel. that specific request completes
1
0
0
@gllmdev
Guillem
2 months
solution: request id swapping. client sends id "123". we replace with uuid. write to stdin. store tracker that maps uuid to original id and result channels. https://t.co/0Izp6jgCWz handles concurrent access
1
0
0
@gllmdev
Guillem
2 months
concurrency: http handles 10 concurrent requests. you write them all to stdin. responses come back in random order. matching response to request when everything goes through one pipe required some thought
1
0
0
@gllmdev
Guillem
2 months
format: stdio needs single-line json. http json has newlines everywhere. we normalize every request to one line before writing to stdin.
1
0
0