Metriqual (@metriqual) header
Metriqual (@metriqual) profile picture

The Rust-based AI Gateway. less than 4ms Overhead. Zero GC Pauses.

India
Joined November 2025
Photo by Metriqual (@metriqual)Photo by Metriqual (@metriqual)Photo by Metriqual (@metriqual)Photo by Metriqual (@metriqual)Photo by Metriqual (@metriqual)Photo by Metriqual (@metriqual)Photo by Metriqual (@metriqual)Photo by Metriqual (@metriqual)Photo by Metriqual (@metriqual)Photo by Metriqual (@metriqual)
Your LLM gateway may give you an HTTP 200 response for every request but might still be muting conversations and manipulating state and API calls. We have identified exactly where problems have occurred, across five architectural levels. Two of our actual findings from stress tests are: > 28% drop in conversation continuity > 25% rate of permanent request failures. None of this triggered alerts. This is FailureAtlas. Complete research and open catalog can be found in the link below.
🤖 Made with AI
2
11
1,699
Something’s cooked 👨‍🍳 @zerithstudio x @metriqual
1
3
65
2,099
I built a fake bank chatbot, loaded it with 30 made-up customers' Aadhaar and PAN numbers, and attacked it 55 ways. Some of them in Hinglish. 54 leaked!!! Then I put Metriqual in front of it. 0 leaked. 0 of 20 normal questions blocked. 1.44 ms added. Why we are building this: It strips Aadhaar (checksum-validated), PAN, account numbers, IFSC and UPI IDs before the prompt leaves, blocks injection in English, Hinglish and Hindi, scans the reply on the way back, and refuses the request outright if any check itself fails. Every bank, NBFC, insurer and health platform shipping AI is sending customer text to someone else's model. DPDP safeguards apply from 13 May 2027, penalties go up to ₹250 crore, and these bots are already talking to customers today. For more information DM me or request access at metriqual.com.
3
5
11
690
coolest LLM gateway site to ever exist metriqual.com
1
3
122
Metriqual retweeted
Alright, going hard so far. I may integrate an edge coding agent with @metriqual
Man, this is so freakishly addictive. @kaggle
1
1
8
392
Most brand colours get picked because someone likes them. I picked one by ruling out red. When red means "provider down", the brand colour's job is to stay quiet so errors pop at a glance. How I landed on a lavender blue: medium.com/@rj45010/i-picked…
2
3
57
Metriqual retweeted
same word, different context. what changes inside the model? this metriqual prototype tracks "bank" through transformer layers to visualise and measure those changes
6
5
1
47
1,107
switching AI models shouldn't mean rewriting your code with metriqual, it's just a new route
2
146
YOUR AI-CONVERSATION SHOULD NOT START OVER JUST BECAUSE THE PROVIDER WENT DOWN!!! Behind the platform our founders researched into this and explored deeper questions: When a model fails, does the conversation survive? Our research studies how forwarding conversation history to a fallback provider preserves context, so users don’t have to repeat their details, preferences, and instructions. Across 750 tested failover events, the stateful proxy achieved a 99.2% Continuity Preservation Rate. What happens inside a model as it processes your prompt? In “Trajectory Geometry of Transformer Representations Across Layers,” we trace how internal representations evolve, exploring how related meanings converge, complex tasks follow more curved paths, and ambiguous words become distinguishable through context. With Metriqual’s stateful architecture, model routing, failover, and observability come together to help keep conversations moving WITH their context intact wrapped in an all-new, creative UI and a smoother UX that makes the platform easy and fun to use. Curious about the experiments, findings, and ideas behind our work? Read them here: 📖 ContinuityBench: arxiv.org/abs/2607.15899 📖 Latent Trajectories: arxiv.org/abs/2606.09287 Explore the new Metriqual. Same mission, a whole new experience.
1
6
10
257
Metriqual retweeted
how do you test whether a failover remembers your conversation? > turn 0: plant a fact (a name, a number, a date, an invented word, a preference) > a few turns of unrelated chat > last turn: ask for the fact, and kill the provider right then > if the fallback answers correctly, the history made it across 150 conversations, run 5 times 750 injected failures for each system we compared
🤖 Made with AI
5
6
51
1,134
metriqual in one line : your AI stays online, even during an outage
2
87
EVERY AI ANSWER HAS A JOURNRY!!! Tracking how transformer representations evolve across layers, revealing patterns in meaning, reasoning and ambiguity. Studying AI and building reliability around it is what matters in the long run. Source: arxiv.org/abs/2606.09287 @theunblunt @its_vayishu
Have you ever wondered what happens on the way when your AI answers something? There is a transformer and inside that you have different internal representations, connecting those you get a TRAJECTORY! Our founders’ paper, “Trajectory Geometry of Transformer Representations Across Layers,” studies the shape of that journey and answers technical questions like How far does it travel? How sharply does it turn? Do related meanings move closer together? Does a small wording change alter the path? Why does this matter? NOW, for AI used in real workflows, understanding that behaviour matters when investigating errors, ambiguity and sensitivity to wording. With access to internal model states, researchers can use trajectory geometry to compare prompts and locate layers for further experiments, without training an extra classifier. The findings are observational, drawn from 150 controlled prompts across three small models. At @metriqual our founders are contributing open-source tools to study model behaviour cuz understanding how AI behaves. Keeping AI applications working. Both deserve attention. Read the research: arxiv.org/abs/2606.09287
1
1
3
639
a transformer processes text through a stack of layers. most research papers asks what representations encode at particular layers this paper asks a different question: what does the path through representation space looks like? so basically measuring the trajectory. its length, curvature, and semantic convergence GEOMETRY, not just probes
reading a paper on how transformer representations "converge" across layers and learned that > you can treat the forward pass as a path and measure its "shape" > category structure is already present in layer-0 embeddings > GPT-2, tiinyllama and qwen behave differently at depth even on the same prompts
2
6
152
a prompt has a shape "bank" can mean river or savings at layer 0, the two are almost the same point by the last layer, they're far apart we measured the paths across 3 models and multi-step reasoning bends the path ~2.7× more than rewording does turns out, MEANING MOVES the split starts around 22% of the way through the network we measured this in GPT-2 Small, TinyLlama, and Qwen2.5 observational finding, not a claim that geometry causes behaviour checkout our paper [Trajectory Geometry of Transformer Representations Across Layers] arxiv.org/abs/2606.09287
2
4
107
The AI features that impact the most negatively do not raise an error but return an HTTP 200 and silently corrupt the conversation. Your health-check is green and the response is already wrong. We classified 5 of these failure modes and published on paper on it. It's called "failure atlas" arxiv.org/abs/2607.17525
2
5
148
traditional failover asks : did another provider return a response? stateful failover asks: did another provider return the response *with the state required to continue the task?* these might sound similar but architecturally, they're completely different systems that's the problem we're solving at metriqual.com : preserving the state required for an application to continue when it switches providers
4
8
149
When your AI crash it switches the model WITHOUT preserving the previous conversation. Metriqual preserves context across failovers, achieving 99.2% context preservation across 750 failovers. See how it works in 30 seconds. 👇 For more details and in-depth understading refer to the research paper published by our founders below. arxiv.org/abs/2607.15899
2
5
1
8
1,455
What happens to the conversation when your LLM fails over? Turns out that switching to a backup model isn't really enough. This paper made me think about LLM gateways differently as in we have spent a lot on AI-infrastructures and making them available but what about the failovers, the quality of the infra, is the context preserved? Is continuity preserved? What systems we have to keep a check on these things which initially seem little but are very much significant in the long run? AI agents being more stateful is the need of the hour. Paper: ContinuityBench — A Benchmark and Systems Study of Stateful Failover in Multi-Provider LLM Routing. arxiv.org/abs/2607.15899 @metriqual @metriqual.com
5
4
1
22
3,012
Most LLM benchmarks ask questions like: How good is the model? How fast is it? How much does it cost? But what happens when the model fails mid-conversation? That’s a different problem. A useful benchmark should measure whether a system can: → detect the failure → switch providers → preserve the full conversation history → continue from the exact point of failure This is why a continuity-preserving architecture is more representative of how AI systems actually behave in production. Because uptime isn't just getting a response. It’s getting the RIGHT RESPONSE WITH THE CONTEXT that came before it. That’s what our founders set out to measure with the Continuity Benchmark. @theunblunt @its_vayishu
4
8
359