Manuel Del Verme

Transparent, interoperable FOSS AI is the great equalizer. I invest//advocate//build for developers accelerating progress toward expanded human potential 🚀🌟

[Manuel Del Verme @ManuelDelVerme]

the AI diffusion bottleneck is reliability. not capability.

most teams don't have the resources to measure agents.

the right way to transition to agents safely is open evals infrastructure. that's what @silverstreamAI @ServiceNowRSRCH @nvidia @IBM @thealliance_ai are doing

[Manuel Del Verme @ManuelDelVerme]

It's funny how the main feedback we got from people was that claude code's scheduled tasks are opaque😂 @AnthropicAI i have a great pitch for you

[Silverstream AI @silverstreamAI]

Bench for Claude Code was #1 Product of the Day on Product Hunt 🏆

Also featured in the @ProductHunt newsletter

Thanks to everyone who supported us 🙏

Built to store, review, and share your Claude Code sessions

More coming soon

[Silverstream AI @silverstreamAI]

Bench for Claude Code was #1 Product of the Day on Product Hunt 🏆

Also featured in the @ProductHunt newsletter

Thanks to everyone who supported us 🙏

Built to store, review, and share your Claude Code sessions

More coming soon

[Manuel Del Verme @ManuelDelVerme]

Agents touch real systems: databases, APIs, permissions, configs.

We've been using Bench to give customers curated traces and get full observability into what our agents did and why. Share traces directly in PRs.

Now we're releasing it. Live on Product Hunt today.
https://t.co/DLAlvH69jH

[Manuel Del Verme @ManuelDelVerme]

At GTC, the same question kept coming up:
Is there a way to track what Claude Code does and share it?

Tomorrow, we’re launching the answer on Product Hunt.

[Manuel Del Verme @ManuelDelVerme]

The @silverstreamAI and @ServiceNowRsch teams built the infrastructure and observability, we host a managed visualization layer compatible with CUBE: https://t.co/hebBPv3zgm

if you're running agents in production right now, what has stopped you from creating broader evals?

[Manuel Del Verme @ManuelDelVerme]

the AI diffusion bottleneck is reliability. not capability.

most teams don't have the resources to measure agents.

the right way to transition to agents safely is open evals infrastructure. that's what @silverstreamAI @ServiceNowRSRCH @nvidia @IBM @thealliance_ai are doing

[Manuel Del Verme @ManuelDelVerme]

right now every team builds eval infra from scratch. no way to compare results across models. no standard way to measure failure. every team starts from a vibecoded shell.

wrap a benchmark once, run it everywhere. no custom integration. built on MCP, Gym and @opentelemetry . not another benchmark. infrastructure for all of them.

[Manuel Del Verme @ManuelDelVerme]

@jacobmbuckman Happy to catch up! Are you coming for GTC?

[Manuel Del Verme @ManuelDelVerme]

Day 2 and it still didn't figure out calling a class AckermannModel and implementing a bicycle dynamics doesn't make for a very useful model

[Manuel Del Verme @ManuelDelVerme]

Hey agent reverse engineer the car protocol and map the room

[Manuel Del Verme @ManuelDelVerme]

we entered the 90s, the age of optical flow

Lucas-Kanade works!
dense opt flow didn't :(

ego motion: floor features move 80px downward, distant windows move 40px upward

[Manuel Del Verme @ManuelDelVerme]

@grok @grok do inverse kinematics for this

[Manuel Del Verme @ManuelDelVerme]

@LiTianleli We've been evaluating Grok on enterprise tools: ServiceNow, Oracle, Odoo, obscure high revenue SaaS, full computer use with DOM replay, Grok has lower FP for task refusals (good TP) but pixel control is lagging.
happy to share finetuning data if useful.