Top AI Demos #26: PostHog MCP - Full Product Visibility, Self-Evolving Agents and Pooled Compute, and Parthas - Multi-agent Claude Code

Top AI Demos #26: PostHog MCP - Full Product Visibility, Self-Evolving Agents and Pooled Compute, and Parthas - Multi-agent Claude Code

Issue #26 · Week of May 11

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

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

This week, we saw builders shipping concrete systems, from local LLM assistants to sophisticated agent orchestrators. Fernando Correa Gomes in São Paulo showed how to get full product visibility with PostHog MCP, while Marvin Bitterlich in Dublin demoed Parthas, a multi-agent system for Claude Code.

Several projects focused on giving AI systems better memory and context. Adrian Knapp in São Paulo shared myfitplan, a personalized nutrition planner using RAG, and Gangadhar Payyavula in Seattle presented Wings Studio for actionable agent memory.

We also saw creative approaches to agent coordination and optimization. Steve Cosman in Toronto explored reflective optimization with GEPA, and Luiz Henrique Simoes in São Paulo discussed Self-Evolving Agents and Pooled Compute.

PostHog MCP: Full Product Visibility Fernando Correa Gomes Product Engineer at PostHog 📍 AI Tinkerers - São Paulo • Apr 30 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS hog.pet github.com

Self-Evolving Agents and Pooled Compute Luiz Henrique Simoes Data Scientist, Designer, PM at N/A 📍 AI Tinkerers - São Paulo • Apr 30 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS github.com Video @almostanhour

Parthas: Multi-agent Claude Code Marvin Bitterlich Senior Product Engineer at Intercom 📍 AI Tinkerers - Dublin • Apr 28 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS @warlockofcode

SousVide: GenAI Art Direction Bridge Yifei Shao Co-founder at MediumRare Imagery 📍 AI Tinkerers - Dublin • Apr 28 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS mediumrareimagery.com Video @shao__yifei

OpenClaw: Local LLM Home Assistant Amanda Lam Ambassador at Women Techmakers Hong Kong 📍 AI Tinkerers - Hong Kong • Apr 29 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS @amanda_lam

More Great Builds

Quick hits from the community — demos worth bookmarking:

Optimizing AI Agent Workflows

Jean Schnneider Arias Suárez • AI Tinkerers - Pereira • Apr 30

Jean Schnneider Arias Suárez presented an architectural guide and demo repo on optimizing developers’ workflows with AI agents, focusing on plugins, MCP servers, and modular custom skills for terminal-based autonomy. The project treats agent instructions as dependencies, uses Cognitive Load on Demand, and highlights practical security concerns like prompt injection and .env exfiltration risks. As a student builder, Jean kept it educational and hands-on for fellow learners. We liked it because it nudged people away from vibe coding toward safer, maintainable agent systems, the exact direction builders are moving in as agentic frameworks mature.

Loading tech tags...

github.com

Agents building Agents: Reflective Optimization

Steve Cosman • AI Tinkerers - Toronto • Apr 29

Steve Cosman, ex-Apple and CEO of Kiln AI, showed how Agents building Agents work through reflective optimization loops. The system (Kiln) runs multi-level GEPA reflective prompt optimization, then creates synthetic evals from real bugs, so agent development iterates on architecture, model choice, tool use, and subagents instead of thrashing. It backed up the approach with fleet-style improvements over traces, and (people loved it) for turning “hard to debug” agent behavior into repeatable evaluation. It felt timely for today’s shift toward real-time, agentic autonomy and for builders who want safer, faster iteration.

Loading tech tags...

github.com

Wings Studio: Actionable Agent Memory

Gangadhar Payyavula • AI Tinkerers - Seattle • May 01

Gangadhar Payyavula from Cognet AI presented Wings Studio, an AI workspace that builds memory agents can act on. The demo shows files, notes, tasks, decisions, meetings, and agent runs turning into typed objects inside an inspectable graph, so an LLM can extract entities and relationships from messy inputs, then trigger actions over the same evolving state. It leaned on agent gateway style tool and message execution plus graph-based schema linked “cards.” We liked how builders could make memory visible and correctable, with folks seemingly appreciating the practical architecture for long-horizon agents.

Loading tech tags...

cognet.ai

ParcoursLab: AI Course Planning

Aaryan Shroff • AI Tinkerers - Toronto • Apr 29

Aaryan Shroff, a UWaterloo Data Science Club project lead, demoed ParcoursLab, a human-centered course planning copilot that builds a degree-respecting schedule while accounting for goals, skills, and extracurricular interests. The app uses Flask and a TypeScript front end backed by Supabase, with LLM help constrained by an ESCO skills database and deterministic prerequisite parsing using a regex and mapping “accessibility graph.” We liked how it avoided slow, expensive parsing failures and still gave the model judgment where it mattered, with people seemingly enjoying the transparency and advisor-friendly summaries.

Loading tech tags...

github.comVideo

PairSpace

Mikele Gajda • AI Tinkerers - Lausanne • Apr 30

Mikele Gajda presented PairSpace, a modular dual-layer dashboard that keeps a personal space and a shared space synchronized for relationship management workflows. PairSpace runs the “always in sync” logic as a clear app-layer data model, with the UI reflecting shared state changes immediately. What stood out was the practical focus on collaboration hygiene, not just chat, and it felt especially relevant to builders juggling high-friction shared context. We liked how / people responded positively to the straightforward, inspectable approach at https://pairspaceapp.com/ and how it could scale into team-friendly relationship ops as a product.

Loading tech tags...

pairspaceapp.com

myfitplan: Personalized RAG Memory

Adrian Knapp • AI Tinkerers - São Paulo • Apr 30

Adrian Knapp shared how myfitplan uses RAG to make an AI diet planner remember each user across sessions, not just improvise from the latest prompt. The demo walked through a Supabase-backed retrieval setup and Langfuse for prompt versioning plus full pipeline tracing, showing exactly what chunks were fetched and why the meal plan looked the way it did. It stood out because personalization was treated as a retrieval design problem, with traces surfacing stale or wrong context that prompt tweaks could not fix. People seemed to really click with the practical debugging lesson, and it fits the current shift toward real-time, multimodal and memory-driven assistants in everyday products.

Loading tech tags...

myfitplan.aiVideo

shelf: AI Catalog Optimization

Anirudh Dabas • AI Tinkerers - Toronto • Apr 29

Anirudh Dabas, a University of Waterloo CS student building agentic tooling for ecommerce automation, demoed shelf, an open-source CLI that applies the autoresearch keep/revert loop to Shopify product catalogs. shelf generates shopper queries, routes them through real AI shopping agents via APIs, scores product visibility, then proposes an atomic catalog metadata change through the Shopify Admin GraphQL API and re-runs the evaluation until you stop it. We liked it because the demo made stochastic LLM judgment feel measurable, using MAD-based confidence scoring that kept noise out. The community feedback was the kind people quietly brag about because it’s adaptable to any optimization target with a clear metric, which fits where AI builders are heading.

Loading tech tags...

github.com

You are one of 95,000+ readers from Apple, Amazon, Microsoft, Google, Nvidia, OpenAI, Anthropic, Cohere, ElevenLabs, Scale AI, Groq, Mistral AI, and others — spanning frontier labs, big tech, startups, and top universities.