The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Bell Tuning surfaced bimodal retrieval at r=0.999 correlation to ground
truth. Noise chunks flagged, prioritized fix list delivered in 48 hr.
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
Guitar knobs turning, jittered depth of field, rendered in Eevee with raytracing on. Trying to see how close I can get to that Cycles-style DOF without the Cycles wait times.
This test landed way closer than expected; good sign for faster turnarounds on close-up shots like this.
Getting Cycles-style depth of field out of Eevee with raytracing is exactly the kind of workflow hack more people should be talking about. What's your typical render time on a shot like this compared to what Cycles would've taken?
Hey Contra 👋
Two months ago I posted Promptly here as a work-in-progress and said I'd launch in about three weeks.
Then a health issue knocked me flat, and the three-week plan was gone.
I'm good now — and as of today, the build is done. I just published the full case study as well.
Quick recap: after shipping 20+ projects as a tech lead, I kept watching the same thing break. A founder says one thing, the developer hears another, and a month in, the whole build is based on a misunderstanding nobody caught early.
Promptly closes that gap. You describe your idea in plain English; the AI interviews you, pushes back on the vague parts, and writes a developer-ready spec while you chat.
But the chat is the easy part. The real engineering is everything that stops the AI from being trusted too much:
→ It never touches the database. 16 agents can only propose changes through typed tool calls; backend code checks each one, and invalid calls get rejected and logged.
→ Nothing ever gets lost. It snapshots before anything risky, so you can roll back a single module or the whole project.
→ Spend is capped, always. Every AI call passes four checks before it fires, and each one is logged with its exact cost.
One line sums up the architecture:
The AI proposes. The code decides.
The model gets to be creative. It never gets to be in charge.
Full case study is on my profile, and the product is live if you want to poke at it.
One question:
Shipped anything with LLMs in production? What guardrail saved you — or what missing one burned you?
Every comment on the last post shaped what got built. Thank you for that 🙏
building solo, in public
Trending
Claude
Claude has entered the design space. How are you using Claude Design?
Contra University
Learn from expert creatives how to earn more using next-gen AI tools.
creativeaiflow
Creative AI workflows are evolving. What tools do you use, and what are their strengths and weaknesses?
freelancerlife
Freelancer life is wins, pivots, and everything in between. What’s yours right now?