Projects in Marcos PazProjects in Marcos Paz
Projects
People
Cover image for I built Brandsmith, a 3D
I built Brandsmith, a 3D logo tool for people who looked at their flat SVGs and thought: “What if this had more chrome, drama, and questionable lighting?” Built in @Figma Make in about 3 hours with one white Monster Energy and the kind of confidence usually only seen before a bad git commit. Upload an SVG, tweak the extrusion, bevels, edge softness, metalness, roughness, lighting, backdrop, and environment, then export a PNG or loop video. Basically, it turns a polite little vector logo into something that looks like it’s about to headline a cyberpunk product launch. Check the 3d tool here: https://deal-coach-21183333.figma.site Made for the @Contra x @Figma Makeathon #ConfigMakeathon (https://www.linkedin.com/search/results/all/?keywords=%23configmakeathon&origin=HASH_TAG_FROM_FEED) #FigmaMake (https://www.linkedin.com/search/results/all/?keywords=%23figmamake&origin=HASH_TAG_FROM_FEED) #BrandDesign (https://www.linkedin.com/search/results/all/?keywords=%23branddesign&origin=HASH_TAG_FROM_FEED) #3DLogo (https://www.linkedin.com/search/results/all/?keywords=%233dlogo&origin=HASH_TAG_FROM_FEED) #CreativeTools (https://www.linkedin.com/search/results/all/?keywords=%23creativetools&origin=HASH_TAG_FROM_FEED)
1
84
Cover image for BE PAWLY — A COMPLETE
BE PAWLY — A COMPLETE LAUNCH KIT, ORCHESTRATED BY AN AGENT IN ONE AFTERNOON The premise I'm an indie dev about to launch Be Pawly, an AI-powered app that matches photos of lost and found pets in real time, helping reunite them with their owners. The product is ready. The launch kit wasn't. Hiring an agency for identity, mockups, ads, and a launch video wasn't on the table for a solo founder. So I gave Melius a single brief and let Mel orchestrate the entire kit. The result 18 connected nodes. One coherent brand. About 4,000 credits consumed out of 50,000. → Hand-drawn bilingual logo with the tagline "Conectamos corazones, reunimos familias" → 3 mobile app UI mockups (Upload Lost Pet, AI Match, Reunion Confirmed) → 4 emotionally segmented Meta ads (Community, Gratitude, Hope, Urgency) → 1 redesigned modern "missing pet" street flyer with QR → 1 cinematic 15-second vertical Instagram video, anchored to a keyframe The process, in three phases I gated execution by phase to manage credits and review the agent's output at every dependency boundary. Phase 1 — Foundation (72 credits). Just the logo. The logo anchors every downstream asset, so it was the only node I needed to lock before generating anything else. Mel routed GPT Image 2 (Medium) and delivered a hand-drawn terracotta mark, and proposed a bilingual tagline I never asked for, but that fit the Latin American neighborhood-first tone perfectly. Phase 2 — Derivatives (around 1,000 credits, 8 parallel nodes). Three UI mockups, four ad variants, one flyer. Mel propagated the Spanish tagline across every asset without prompting, invented credible character names (Max, Toby, María & Simba), and used context-appropriate Latin American Spanish ("publicador" instead of "poster"). The brand feels like a real product, not stock. Phase 3 — Motion (around 3,000 credits). A cinematic keyframe of the reunion, then a 15-second vertical Instagram spot anchored to it. When I asked Mel which model it would use, it didn't just pick — it offered me Kling 3.0 Omni (2,520 credits) versus Grok Imagine (1,260 credits), explained the tradeoff ("Kling handles the emotional human-first nuance a bit better"), and let me decide. What separates an agent from a generator Three behaviors Mel did on its own that no generator I've used does: Model routing with reasoning. GPT Image 2 across the 8 image nodes for stylistic coherence; Kling 3.0 Omni for the video, chosen for emotional nuance. Transparent about cost tradeoffs. Style propagation through wired connections. The tagline, color palette, and character voice flowed from one node into every downstream node automatically. No copy-paste, no re-prompting. Initiative within taste. The bilingual tagline. The character names. The neighborhood-aware Spanish. These were unprompted but always on-brief. Feedback for the Melius team What worked beautifully: → The canvas-as-conversation model. Being able to point at a specific node and refine it without re-running the whole graph is a step-change versus linear chat-based generation. → Mel's proactive credit transparency. Asking "which model and how many credits" before running was huge for budget management. → Auto-grouping into Brand Identity, Meta Ad Campaign, and Social Video was an unprompted decision that made the canvas instantly readable. Suggestions: → Surface estimated credit cost on hover for every node before approval, not just when prompted. → A "fork node" action to create alternative variants in parallel without losing the original would be useful for A/B-ing ads. → An export-as-launch-kit shortcut (a zip with all assets in presentation order) would close the loop from creative to delivery. Built entirely on Melius. Models routed: GPT Image 2 (Medium) and Kling 3.0 Omni. Total credits consumed: about 4,000 of 50,000. The 15-second reunion video is attached. Project canvas: https://app.melius.com/projects/04e916b8-35de-4957-8db6-53ac1af180ad/canvas/4d8d194a-d0d5-4e4c-a59b-d7a4561923bb
1
70