![Top AI Demos #29: Agent Memory, Image Relighting & Software Testing [AI Tinkerers - Post-Training]](https://sloppy-joe-app.imgix.net/ai_tinkerers/logos/post-training_laptop-sticker_190_de2487c31e50813a3471e79627268ffb.png?w=300&fit=max&auto=format%2Ccompress&q=75){kind=small}
Top AI Demos #29: Agent Memory, Image Relighting & Software Testing
Issue #29 · Week of June 1
Anther week of impressive builds drawn from the most significant and highly rated AI Tinkerers presentations around the world!
Aurelian Jibleanu’s ArgosBrain: Persistent Code Memory tackles agent efficiency by improving code understanding, while Vesh Thakker’s 50 Ants: Agentic Software Testing demonstrates automated defect remediation.
We also explored how AI can enhance continuity and interaction. Ady Ngom’s HQIQ Maestro: Voice-First Generative UI offers seamless, context-aware voice interactions, and Sean Cofoid’s consulting-os: Git-Versioned AI Memory shows verifiable self-improvement through structured learning.
Builders are also creating more robust operational systems. Samir Akarioh’s Rasa: Conversational DevOps Agent makes infrastructure management interactive, and Devashish Meena’s Clawrium: Orchestrating Local Openclaw Agents provides tools for managing multiple local agents.
Top 5 Picks (June 1)
1
TOP PICK
ArgosBrain: Persistent Code Memory
Aurelian Jibleanu from ArgosBrain presented a live demo of persistent code memory in Rust, showing how ArgosBrain indexes a codebase into a unified graph using SCIP, LSP, and tree-sitter, then serves sub-50ms symbol lookups over MCP. It runs local-first on Kubernetes and in tooling like VS Code and OpenClaw, driving per-task agent costs down while eliminating hallucinated symbols for structural queries. After instrumenting tool calls, he reported 73% of tokens were redundant re-reads, and people seemed to like the pragmatic, determinism-first shift. We liked it because it turns agentic token waste into something measurable and actionable for builders.
|
2
RUNNER UP
AI Post-Capture Image Relighting
Jack Shapiro demonstrated an app that relights photos after they are taken, turning off-the-shelf photography ideas and prompts into a repeatable end-to-end workflow. He’s the kind of designer-photographer who treats lighting as meaning, and he approached the build like a UX system, pairing careful prompt structure with practical ML image tooling. It felt especially approachable for builders, and people (quietly) seemed to love how it made “photo magic” learnable. This kind of pipeline hints at real product value for photo workflows, from creators to enterprises.
TECH STACK
|
instagram.com
3
COMMUNITY FAVORITE
50 Ants: Agentic Software Testing
Vesh Thakker, the CEO behind Aspired Future, demoed 50 Ants - Automated Agentic Testing and Verification, showing an agentic orchestration engine for software validation, stress-testing, and defect auto-remediation. 50 Ants coordinates stateful agents that run tests both inside the codebase and externally, then hands failures to an AI coder workflow for targeted fixes. The recorded capability and co-remediation clips made the approach feel practical, and people seemed to like how it tackles the real code-quality bottleneck at AI-accelerated speed. It felt like a product-shaped blueprint for keeping systems trustworthy as teams scale.
|
4
STANDOUT
Rasa: Conversational DevOps Agent
Samir Akarioh presented a conversational DevOps agent built with Rasa that lets you manage real infrastructure through natural language. The demo connected an NLU layer to the Scalingo API so a single sentence could create apps, scale services, check logs, or update environment variables, while keeping the workflow grounded in concrete DevOps actions. We liked how the interface made everyday admin tasks feel like chatops, and audience response nudged it even further (people enjoyed the practicality). It’s a nice blueprint for turning agentic “understanding” into production-ready control.
|
5
NOTABLE
Daily: Gradient Bang AI Game
Kwindla Kramer of Daily presented Gradient Bang, an LLM-centered massively multiplayer game that grew from sample code into a community-driven world. The project treats generative logic as gameplay and uses procedural CSS to produce endlessly varied visual gradients while the experience stays interactive and responsive. With Pipecat and Daily’s real-time voice and systems background behind it, it felt like a practical blueprint for building agentic mechanics people can actually play. It also lined up with what folks quietly kept pointing to, how fun ideas become durable contributions, not just prototypes.
TECH STACK
|
More Great Builds
Quick hits from the community — demos worth bookmarking:
Raphaël Semeteys • AI Tinkerers - Paris • May 21
Raphaël Semeteys presented “Build Your Book: Engineering Sci-Fi with an AI Software Mindset,” where he treated story production like spec-driven software engineering. He used AI as an execution layer for lore and scenarios, then connected text clues to an Alternate Reality Game that unlocks digital assets, supported by a custom publishing workflow. He even shared the stable-node setup and the spec-driven narrative architecture from his projects, including the slides that documented the approach. We liked it because builders can reuse the workflow immediately to ship faster with fewer inconsistencies, and people seemed especially to enjoy the bridge from prose to playable worlds.
Loading tech tags...
Hishaam Abdul Razik • AI Tinkerers - Dubai • May 23
Hishaam Abdul Razik from AmpiereLabs presented an end to end Transformer diffusion pipeline for molecular battery material generation, showing a terminal run that mapped discrete constraints like cobalt free composition and symmetry into conditioning vectors for a diffusion model. The demo then denoised 3D atomic coordinates, ran “zero strain” and fed raw structures into CHGNet for fast relaxation, followed by ALIGNN to validate energy above hull in under three minutes. We liked how it routed topology with the Transformer and let physics models correct spatial hallucinations, and (people loved the hands on speed). It felt like the exact local frontier progress open weights builders want, with clear product potential for automating materials R&D loops.
Loading tech tags...
Ady Ngom • AI Tinkerers - Dubai • May 23
Ady Ngom presented Maestro, a voice first generative UI agent that builds the interface live as you speak, with scenes, listings, and comparison views composing themselves in real time. The architecture stayed package driven, using small tools like enable_translator_mode and disable_translator_mode to switch personas by updating chat context, while signaling UI state through data packets. It also ran provider agnostic on Gemini Live with an OpenAI Realtime fallback, keeping the “agent is the frontend” idea truly practical, and (people seemed to love the translator flow). We liked it because it showed how modular orchestration can make multilingual, accessible product UX feel natural, not bolted on.
Loading tech tags...
Diego Alejandro Giraldo Ascencio • AI Tinkerers - Pereira • May 21
Diego Alejandro Giraldo Ascencio presented Calificame, a system that grades handwritten Spanish exams from phone photos in seconds using vision agents, RAG, and generative feedback per question. He built it with a hybrid pipeline: an AI-guided image-preprocessing pass for corners and CLAHE or gamma settings, plus a deterministic OpenCV-only fallback (Canny, contours, perspective warp, LUT-based enhancement) if validation fails, so hallucinations do not break large batches. People seemed especially to like the reliability tradeoff, and it felt grounded for builders chasing practical agentic workflows and dependable document automation. It’s a nice template for turning flaky vision into something you could ship.
Loading tech tags...
Sean Cofoid • AI Tinkerers - Nashville • May 21
Sean Cofoid presented consulting-os, a git-versioned personal OS for AI agents that keeps typed memory in plain files and runs role-separated skills with a learning loop that persists across sessions. The demo walked through an artifact chain: a plan that committed to expected outcomes, then a debug report to a separate triage role, a regression that corrected a wrong-shaped fix, and a meta-heuristic that generalized the lesson. We liked it because it made self-improvement falsifiable and repeatable, and (people really valued that clarity). It feels timely with today’s shift toward more agentic reliability and efficient long-running workflows, and it could easily become a product for teams who need auditable, self-correcting agent behavior.
Loading tech tags...
David Ricardo Rivera Arbeláez • AI Tinkerers - Pereira • May 21
David Ricardo Rivera Arbeláez from Universidad Católica de Pereira presented Edge Power, an IoT energy monitoring and optimization platform that runs a Next.js dashboard on top of MQTT/WebSocket real-time data flows. The demo showed a modular edge-to-web pipeline, Dockerized services, and a clear separation between ingestion, processing, and visualization, with embedded devices feeding live updates while logs and architecture stayed easy to follow. Builders seemed to love how quickly it scaled from prototype to reproducible setup. It also fits the current shift toward more efficient agentic and edge workflows, and it could become a practical product for utility-style analytics.
Loading tech tags...
Cyrus Nouroozi • AI Tinkerers - San Francisco • May 18
Cyrus Nouroozi demoed YALC, an open-source GTM OS and CLI that programmatically manages LinkedIn outreach, lead qualification, and multi-channel campaigns. YALC is built for repeatable workflows, so you can wire together targeting, sequencing, and qualification steps from the command line instead of juggling spreadsheets and brittle scripts. As someone who’s shipped AI tooling in DSPy and previously led product at WOMBO, he kept the focus on practical engineering. We liked how clearly it maps to a real builder pain point, and it sounded like people wanted to reuse pieces of it.
Loading tech tags...
Dan Denney • AI Tinkerers - Nashville • May 21
Dan Denney demoed their personal OpenClaw setup on a Mac mini, using Discord plus Codex and Gemini web search to chain prompts into real outputs: reviews that get posted to a site, a read-later link library, and daily sports wager research notes. They mapped each workflow with simple tldraw diagrams and built a supporting Astro starter repo for tidy content publishing. We liked it because it made agentic tooling feel immediately buildable, and (people enjoyed the practical flows) more than the flash, with great product potential for everyday automation.
Loading tech tags...
Angus Robertson • AI Tinkerers - Los Angeles • May 19
Angus Robertson showed a competitive intelligence SaaS that benchmarks a company against up to 4 competitors across 9 dimensions, then delivers the full report in about 15 minutes for $45. The demo walked through the live report generation flow, a multi-LLM “AI visibility” testing pipeline, and the concurrency design that kept the 9-section output fast. He built the prototype quickly with Claude Code, then tightened QA and edge cases with beta feedback (people liked how it tied signals back to decision-making). We liked it because it turned a high-ticket consulting playbook into an agentic workflow you could adapt.
Loading tech tags...
Devashish Meena • AI Tinkerers - Seattle • May 21
Devashish Meena presented Clawrium, a CLI that manages a fleet of Openclaw assistants across your local network, including headless setup and a network-wide command center. The project uses a normalized configuration approach and orchestrates multi-host deployments via Ansible-style automation, with an easy one-command install and local model integration options. It felt especially useful to builders managing real infrastructure, and survey folks seemed to like how the “caveats-first” playbook made agent ops less mysterious. We liked it because it turns agent experimentation into repeatable, ops-ready tooling.
Loading tech tags...
Hasaan Khan • AI Tinkerers - Palo Alto • May 21
Hasaan Khan introduced Smart SMB Hub, where an ElevenLabs-powered Nova voice receptionist handles inbound calls and triggers a strict five-step Make.com automation sequence for Nest Driving School. The flow captures leads, confirms bookings, sends reminders, requests reviews, and re-engages unconverted parents, all using explicit state-machine style guardrails with fallback nodes instead of free-form “whatever the prompt says” replies. It stood out because it stayed reliable on off-script questions in a liability-sensitive niche, and people seemed to really enjoy that pragmatic lesson (people loved it). We liked it as a reminder that agentic UX still needs crisp orchestration.
Loading tech tags...
Travis McQueen • AI Tinkerers - Manchester New Hampshire • May 20
Travis McQueen, founder of Mimory, showed how Resonant Echoes structures multi-agent “conversations” from a 1-on-1 historical Echo into accelerated Lincoln-Douglass debates, then out to panel and seminar flows. The project uses portable prompt compilation so personas stay consistent across runtimes, plus an orchestration harness that manages shared transcripts and turn-taking without agents stepping on each other. People seemed to love how the protocol details were concrete, and it hints at product paths for privacy-aware, educational agent systems. For builders, it’s a practical blueprint for reliable multi-vocal dialogue.
Loading tech tags...
Latest Content
How to Ship Complex Features 10x Faster with AI Agents | Dex Horthy (HumanLayer)
Dex Horthy (HumanLayer) breaks down the “12 Factor Agents” approach to shipping multi-step agentic workflows faster: structured outputs, ...
Watch Now →
Homecrew: An Open-Source Package Manager For Agent Skills
Homecrew is an open-source package manager for agent skills: install a skill (a directory with `SKILL.md`) from a git tap and keep it syn...
Read More →
Top Job Matches
Matched based on your meetup activity and profile
Paxos Health
• New York & Toronto
• $110k - $175k (varies w/ location/level); generous equity
Stanford-founded Seed-stage healthcare AI startup with >$5M in VC funding and AI agents deployed in production with cu...
Apply Now →
Dex
• London (5 days on-site)
• £250,000
Frontier AI engineering role building the AI tooling layer for complex financial modelling.
Apply Now →
Jakib AI
• Columbus, OH
Jakib is a profitable, growing applied AI firm embedded with operator-led companies in logistics, manufacturing, and c...
Apply Now →
You are one of 95,000+ readers from a16z, Adobe, Amazon, Anthropic, Apple, Baseten, Black Forest Labs, Bloomberg, Boston Dynamics, ByteDance, Cohere, Cursor, and others — spanning frontier labs, big tech, startups, and top universities.
Top AI Demos #29: Agent Memory, Image Relighting & Software Testing