Glassray (@glassrayai) header
Glassray (@glassrayai) profile picture

The evaluation layer that makes your AI agents self-improve.

London, United Kingdom
Joined June 2026
Photo by Glassray (@glassrayai)Photo by Glassray (@glassrayai)Photo by Glassray (@glassrayai)Photo by Glassray (@glassrayai)Photo by Glassray (@glassrayai)Photo by Glassray (@glassrayai)
This week we're sending 10 AI founders a harness to learn how they build theirs. We flew in to SF and leave on Saturday. X do your magic and help us get a coffee with them ❤️ • Andrew from Numeric • @woodenstick from Salient • @mkjung_ from @heyivoai • @sgondala2 from @furtheraicom • @cszy from @fieldguide • @mlmanapat from @Rowspace_AI • @sambhav_anand from Fulcrum • @thepericulum from @gcai • Timmy from @FazeshiftHQ • @AugustFR from @QurrentAI
4
4
8
443
Super44's agent sends cafes and restaurants a daily briefing built from up to 20 tool calls. One wrong number and the merchant stops trusting it. 6 weeks with Glassray: reliability up 2.2x, weekly traffic almost 3x, same headcount. More here: glassray.ai/customers/super4…
1
63
gpt-4.1 is 1.5x cheaper than the trace judge we run today and caught more real failures. We benchmarked four models on 24 real traces, three runs each, scored against a pre-written golden set. Full method and numbers: glassray.ai/blog/cheap-model…
1
1
103
Glassray retweeted
Same 50 traces. Same digest. Same prompt. Claude Sonnet 4.6: $3.45/1k gpt-4o-mini: $0.11/1k We're moving our highest-frequency LLM call to gpt-4o-mini, same search results. But the cheapest, Gemini 2.5 Flash-Lite, broke. Full writeup: glassray.ai/blog/cheap-model…
1
4
3
251
We moved our code explorer to a cheaper model: 32% cheaper, no accuracy lost. The lazy version, just swapping in a cheaper model, failed. All 10 candidates missed our bar. The win came from the harness around it: incremental submission, strict citation checks, a repair pass. Then the cheap one shipped at $0.78 a run. Full method and numbers: glassray.ai/blog/cheap-model…
1
1
59
1/ Glassray reads your code and watches hundreds of thousands of production traces to catch the mistakes your AI agent makes in production. Think one of its findings is wrong? Correct us - we'll learn from your feedback.
1
77
2/ Dismiss the example, via MCP, Slack or Ask Glassray. Glassray takes it into account when looking for new examples, and every related graph and number updates to match. See what else we shipped this week → glassray.ai/docs/changelog
20
1/ Your AI app is live on Vercel. Do you actually know what it's doing in production? Now you can: add a Trace Drain, copy three values, no code changes, and every production run streams into Glassray.
1
1
53
2/ Vercel streams your AI SDK telemetry straight in, and a multi-step run shows up as one clean trace. Then Glassray gets to work: it maps your runs into flows, flags where an agent deviated from what you intended, and suggests fixes it checks against the traces that come after
1
1
32
3/ Already on Better Stack? Connect it as a source and pull your traces in directly. For teams shipping AI on Vercel who want to know what their agents actually do in prod. Have a look at our latest features! glassray.ai/changelog
13
Wonder how Brex, Deel or Elevenlabs hire? We all think we hire for talent and then panic 3 months later. Metaview has analysed more than 3 million interviews to work out what actually predicts a great hire. I'm sitting down with Siadhal Magos on Thursday to get all the insights! Spots are filling fast! Link in the comments! #london #ai #hiring
2
6
202
What a great turnout, over 60 AI builders in London joined us next to Westminster Abbey alongside @vercel! London AI is totally popping!
2
34
1/ You merged a PR to fix a hallucination your agent kept making. Did it actually work? Glassray checks the traces that come in after your commit ships and tells you. 🎬
1
1
1
29
2/ When a commit looks like it fixes a deviation, we don't assume. We re-check reality: scan the traces since it landed. One recurrence and the fix is refuted; zero and it's closed, with the evidence attached.
1
13
3/ So instead of hoping the hallucination is gone, ask, "Is it still happening since my fix?" and get a real answer! glassray.ai/docs/changelog
8
We put Claude on the night shift at @glassrayai. It picks up tickets overnight, opens PRs, and responds to @greptile feedback until the PR is cleaner for human review. The goal is not autonomous engineering. Just better focus for the hard problems. linkedin.com/pulse/we-put-cl…
1
1
5
205