Top AI Demos #32: Cloud Waste Agents, MVP Validation, & AI Skill Orchestration

Top AI Demos #32: Cloud Waste Agents, MVP Validation, & AI Skill Orchestration

Issue #32 · Week of June 22

Joe Heitzeberg • Founder at AI Tinkerers • ⏱️ 1 min read

Creating space for leading builders to share ideas, grow, and make an impact.

This week’s AI Tinkerers submissions dive deep into practical agent orchestration and building robust developer tooling. We saw builders like Hugo Rodriguez in Medellín demonstrating how an AI agent can find and quantify cloud waste on AWS and Databricks with his Cloud Zombie Hunter. Louis Marcondes in Curitiba showcased a sophisticated multi-agent system for WhatsApp using n8n, integrating Claude and Supabase RAG in his Santé Nutrir project. Ryan Waliany from Seattle presented SPEAR, a framework for autonomous agents designed to prevent process failures and rework, in his talk on the SPEAR: Autonomous Agent Framework.

Several demos focused on building specialized AI systems, highlighting distinct technical choices. Luiz Goncalves in Curitiba shared how to validate product ideas with low-cost pre-MVPs using landing pages and analytics in Validando MVPs com Landing Pages. Abid Waqar in Islamabad Rawalpindi detailed his Gryter: AI Skill File Orchestration, where specialized agents achieved team-level output. Danish Munib also from Islamabad Rawalpindi tackled the complex problem of REACH: Spatial Disaster Alerts, addressing the challenges of vague location names.

We also saw exciting advancements in developer tooling and code systems. Ady Ngom in Dubai and Doha explored MCP for separation of concerns in AI, composing independent packages into generative UIs with HQIQ Maestro: Voice-First Generative UI and Maestro: MCP Separation of Concerns. Shimin Zhang in Seattle introduced “Inhabited-design,” a Claude Code skill for generating AI-powered UI designs with personality in Inhabited-design: Claude Code UI Skills.

Cloud Zombie Hunter: AWS/Databricks agent Hugo Rodriguez Senior Architect at Source Meridian 📍 AI Tinkerers - Medellín • Jun 09 Hugo Rodriguez from Source Meridian presented Cloud Zombie Hunter, a terminal-based Claude Code agent that hunts AWS and Databricks cloud waste while requiring explicit go-ahead before it would act. It inspects abandoned resources like EBS, Elastic IPs, stale snapshots, and lingering CloudWatch alarms, plus Databricks clusters, SQL warehouses, and jobs, and it grounds savings by pulling live pricing and billing data instead of guessing. We liked how it handled the “idle is not waste” trap with tag-based intent checks and visible math, and people seemed to enjoy that real-world guardrails beat confident hallucinations. If productized, it could become a safer cost-savings co-pilot for teams. Read more → TECH STACK Claude CodeAWSAWS CLIAWS MCP ServerDatabricks Unity Catalog PROJECT LINKS @harf2310

Validando MVPs com Landing Pages Luiz Goncalves Founder | System Integration Architect at Zanta Opsvia & MediWhats 📍 AI Tinkerers - Curitiba • Jun 10 Luiz Goncalves showed how he builds pre-MVPs fast by validating demand before writing backend code. The demo uses an AI-generated landing page, PostHog analytics, a Cloudflare Worker to capture leads, and Resend for confirmation emails, then ties decisions to a small ads budget. It’s a practical system integration pattern for founders in regulated, compliance-minded spaces, and it felt especially reassuring to attendees since (people loved it) the “measure first” approach. We liked it because it makes agentic-style workflows measurable and cheaper to iterate, not just clever. Read more → TECH STACK AI StudioCloudflare PagesCloudflare WorkerPostHogResendAI Studio (DeepMind)Cloudflare Pages & Cloudflare WorkerClaude skill PROJECT LINKS

Gryter: AI Skill File Orchestration Abid Waqar Founder/CEO at Gryter 📍 AI Tinkerers - Islamabad / Rawalpindi • Jun 12 Abid Waqar demoed Gryter, his solo-built AI fitness coach, showing onboarding and generated workouts in the first 90 seconds, then digging into the orchestration harness behind the scenes. He kept the system maintainable with 12 skill files, subagents with scoped context, and a CLAUDE.md so Claude reloaded the codebase correctly, plus MCP connectors that drove his real Firebase-backed Flutter app and CI tests on-device. What we liked was the clear symptom-to-fix story from code drift, which felt relatable to builder pain. It also seemed (people loved it) because the harness made a solo workflow output what a team used to do. Read more → TECH STACK Claude Opus 4Claude CodeMCP connectorsFlutterFirebase PROJECT LINKS gryter.com @abidwaqar98

Santé Nutrir: Multi-Agentes no n8n Louis Marcondes IA & Automação para PMEs | Nutricionista Clínico at Louis Marcondes 📍 AI Tinkerers - Curitiba • Jun 10 Louis Marcondes presented Santi, a self-hosted WhatsApp multi-agent that answers product questions for a Brazilian “clean label” supplements brand and escalates to humans for clinical uncertainty. The system runs on n8n as the orchestrator, routing tool calls into separate sub-workflows that each own RAG over Supabase pgvector, prompts, and guardrails, with Evolution API handling inbound messages. He also walked through real-world ops lessons like WhatsApp debounce and the hard parts of closing the Wix checkout loop. We liked it because it felt usable, the kind of pattern builders (quietly, the community) can copy for production. Read more → TECH STACK n8n (self-hosted): orquestrador central; State Machine de sessãodebounce de mensagensCodeJavaScriptOpenAI GPTEmbeddingsVector storeEvolution APIClauden8nSupabasepgvector PROJECT LINKS

SPEAR: Autonomous Agent Framework Ryan Waliany CEO at Ambiguous 📍 AI Tinkerers - Seattle • Jun 08 Ryan Waliany presented SPEAR, a framework for autonomous agents that turns “prompt to output” into a repeatable Scope, Plan, Execute, Assess, Resolve loop. The demo showed how the agent makes its plan visible, runs work against a MECE rubric, logs decisions, and iterates at the right cadence from quick tasks to full workstreams, not just one shot. His background across big-scale operations and agentic systems shaped the focus on process over raw model capability. People seemed to like that it made Plan and Assess feel lightweight but necessary, and it hints at agentic workflows becoming real infrastructure for teams. Read more → TECH STACK Claude Code PROJECT LINKS ryanwaliany.com @rwaliany

More Great Builds

Quick hits from the community — demos worth bookmarking:

Using Agents to Improve Myself

Mashhood Rastgar • AI Tinkerers - Islamabad / Rawalpindi • Jun 12

Mashhood Rastgar showed his personal agent that runs on a cloud VM and helps manage calendar, meeting recordings, interview notes, learning, talks, podcast workflow, and even finance. The demo walks through what the harness looks like, the shift from “Software 1.0” to “Software 3.0,” and how OpenClaw was set up with a security-focused mindset on a VM. We liked how it combined day-to-day utility with practical compound engineering ideas (people enjoyed the real-world scope). It felt especially relevant for this moment of more capable, agentic workflows and safer execution.

Loading tech tags...

Maestro: MCP Separation of Concerns

Ady Ngom • AI Tinkerers - Doha • Jun 15

Ady Ngom presented Inside Maestro, showing how separation of concerns can be enforced with MCP as an orchestration layer for Maestro’s voice-first UI. The demo runs in modular packages: Pandini compresses screenshots in-browser, the Avatar Generator exposes an avatar lifecycle via Fast MCP wrapped around REST, and Avatar Locales uses skill MD to add new languages without redeploying. It felt especially practical for builders, and people seemed to like the “swap-in tools, don’t rebuild agents” pattern. We liked it because it made agent reliability and packaging more approachable.

Loading tech tags...

hqiq-ai-dev.netlify.app

Doomsday Stick

Jason Gardner • AI Tinkerers - Upstate NY • Jun 10

Jason Gardner from contextFound presented Doomsday Stick, a Pip-Boy-style macOS app that runs fully offline while still surfacing practical “survival knowledge” via local LLM inference. It leverages Apple’s MLX on Silicon to keep responses available when the network is gone, turning the offline edge cases into the main feature. Builders appreciated the tight, distributable setup (people seemed to love it) since it made “token-pocalypse” constraints feel solvable. If this ever packaged as a product, it could power rugged, privacy-first assistants.

Loading tech tags...

doomsdaystick.comVideo

Strawberry: Local Gemma Voice Assistant

Parth Parikh • AI Tinkerers - Seattle • Jun 08

Parth Parikh from Microsoft shared Strawberry, an entirely local voice assistant that transcribes with Whisper, reasons with Gemma via LiteRT-LM, and speaks using Kokoro TTS. He built a low-latency 3 stage producer consumer pipeline where LiteRT-LM streams tokens, a synthesizer thread generates per sentence audio, and a main playback loop stays in sync with a parallel barge in monitor using websocket based AEC. He also crafted a wake word pipeline with Whisper Tiny, plus sandboxed code execution and persistent memory. We liked how practical and responsive it felt, and people seemed to enjoy the focus on real performance wins for local agent style apps.

Loading tech tags...

github.com

Inhabited-design: Claude Code UI Skills

Shimin Zhang • AI Tinkerers - Seattle • Jun 08

Shimin Zhang shared Inhabited-design, an adversarial Claude Code skill that takes a “build me an X for Y” prompt and iterates toward a presentation-ready page design with a clear point of view for a specific named user. The demo walked through the markdown artifacts (claude_designer.md, claude_icp.md), an inspiration-bank protocol, and an iteration loop that continues past typical one-shot outputs until convergence. We liked it because the feedback loop and adversarial critique kept the design from drifting into generic slop, and it matches how agentic workflows are trending toward tighter self-checking.

Loading tech tags...

wolfpeachlabs.comshimin.ioVideo

Browser Agent Data Entry

Yazan Risheh • AI Tinkerers - Dubai • Jun 13

Yazan Risheh showed an agent that reads a weekly timesheet from Excel and then logs every hour in Salesforce with no clicks, no copy paste, and no manual entry. In the demo, Claude drives the workflow by reading the screen, matching dates, projects, and tasks, and entering hours in real time. It was a nice example of browser agents and RPA-style automation stitched into an LLM feedback loop. We liked it because it tackles a real ops pain point when systems do not integrate, and it felt like the kind of thing people kept talking about.

Loading tech tags...

HQIQ Maestro: Voice-First Generative UI

Ady Ngom • AI Tinkerers - Dubai • Jun 13

Ady Ngom presented Inside Maestro, showing how MCP can act as a separation of concerns layer for voice-first generative UI. Maestro composes independent packages via MCP: Pandini compresses screenshots without installs or APIs, the HQIQ Avatar Generator builds a full avatar lifecycle, and HQIQ Avatar Locales adds new languages through Skill-MD and on-demand tool calls. Fast MCP wraps REST so tool surfaces materialize quickly, and adding Arabic live required no redeploy. We liked how the boundaries added real reliability tradeoffs, with builders seeming to enjoy the clear, reusable patterns (people loved the composability). It also maps nicely to the current shift toward cheaper, more modular agentic systems and could become a practical product toolkit for regulated concierge workflows.

Loading tech tags...

hqiq-ai-dev.netlify.app

KML Earthworks: AI-built engineering app

Caio Zanetti • AI Tinkerers - Curitiba • Jun 10

Caio Zanetti showed KML Earthworks, a Streamlit app that turns a Google Earth road alignment into deterministic earthworks results, with prismatoid cut fill volumes, mass balance, and Excel export. He built a conversation-driven pipeline (parse to stationing every 20m, elevation enrichment with API fallback, grid optimization) and ran the 58 tests while walking repo structure, plus a CLAUDE.md “constitution” that guided the agent across commits. The standout idea was brutal data validation, because bad elevations would mean bad contracts. It felt especially relevant to builders chasing reliable, cost effective agentic workflows and, from attendee reactions, it landed as the kind of practical demo people wanted to reuse. If productized, it could become a fast planning utility for bid ready quantity takeoffs.

Loading tech tags...

kml-earthworks.streamlit.appgithub.comVideo

REACH: Spatial Disaster Alerts

Danish Munib • AI Tinkerers - Islamabad / Rawalpindi • Jun 12

Danish Munib showed REACH, a disaster-alert system that treats Pakistan’s PDF bulletins as a spatial data problem, extracting structured events and mapping them to real polygons. The pipeline scrapes bulletin boards, uses multimodal extraction, and then resolves messy direction-style locations like “Upper Sindh” by intersecting grids with admin boundaries, with an assistant that queries the data via SQL. People in the audience seemed to really enjoy the “hard part” being handled thoughtfully, and it felt especially relevant to today’s agentic, tooling-rich workflow era. This is the kind of build that could become real utility for disaster teams and neighborhoods.

Loading tech tags...

reach-prototype.netlify.appgithub.com

You are one of 95,000+ readers from OpenAI, Anthropic, Google DeepMind, Mistral AI, Cohere, ElevenLabs, Perplexity, Midjourney, Scale AI, Together AI, Groq, Weaviate, and others — spanning frontier labs, big tech, startups, and top universities.