Top AI Demos #30: Clinical Co-Pilots, Pull Agents, and Life Planners

Joel Horwitz from Synter demoed how their agentic system builds and manages paid media campaigns across 10+ ad platforms, with creatives generated and bidding and budget allocation handled without human clicking. The demo showed an MCP server that let tools like Claude and Cursor control ad accounts directly, backed by a unified abstraction layer over wildly different platform APIs, auth flows, and rate limits. Most importantly, they mapped the autonomy versus human-approval boundary in a way people found practical. We liked it because it turns a messy integration problem into reusable patterns developers can adopt.

Tom Kinyanjui from Neural Labs Africa presented Building NeuralSight, showing how NeuralSight turns existing ultrasound, X-ray, CT, and MRI workflows into a 24/7 tele-radiology diagnostic center without new hospital hardware. The project handles secure imaging data storage and compression, then routes remote reports and sets up AI-powered triage and report drafting, with Python deep learning (Fastai/PyTorch) built around real clinician delivery. What stood out was the interoperability-first design and practical low-resource focus, which people seemed to enjoy. We liked it because it connects model work to the messy deployment gaps builders hit every day.

Ryan Waliany presented AI-native workspace, a collaboration demo that organizes 15+ SaaS tools around human and AI joint work. The workspace treats agents as teammates with shared context, so decisions, drafts, and handoffs stay in one place rather than bouncing between chat tabs, using lightweight agent orchestration patterns. It felt especially timely for teams exploring multi-step agent workflows, and (people noticeably leaned in for the communication mechanics). We liked it because it made “agent collaboration” concrete, with potential to become a real product layer for day-to-day operations.

Levi Turk presented 3D UI for Openclaw + Hermes, where agents build shared real-time worlds and let users “cast” interactions via voice, portals, and 3D buttons that trigger API calls or effects. The demo uses a Three.js-style rendering workflow plus a web-first UI layer, and it contrasts a Mindcraft harness with classical roles like Curator, Orchestrator, Coder, Reviewer, and Tester. We liked how it made agent orchestration visible, not abstract, and people seemed especially into the practical “shipping lessons” angle. It felt directly relevant to today’s push for smoother agent UX and faster iteration loops.

Ricardo Kerr presented VOXMAP, a dialogue panel that analyzes customer conversations in real time and turns them into structured summaries, sentiment signals, recommended solutions, and next steps. It runs via a web NLP dashboard that includes automated spell correction plus visual analytics like word clouds and word co-occurrence graphs. Given his background in conversational AI personas and RAG, the demo felt especially practical for teams. People seemed to like how quickly the insights showed up, and it fit the broader shift toward agentic tooling that helps reduce operational guesswork.

Abdallah Abadah, a senior full-stack engineer from Amman, demoed a lightweight LLM-enhanced ecommerce storefront that focused on squeezing strong UX out of small models. The app presented practical ways to optimize user inputs and data structures before the model runs, keeping latency and cost in check, while letting features plug in with minimal friction. It stood out because it came with a clear runbook for AI empowerment, and people seemed to like how approachable the integration felt. We liked it as a blueprint for real-world agentic storefront upgrades.

Thiago Nobre Mascarenhas presented Cortical Eval, a demo that judges whether an AI output makes sense using a Wernicke-style E0 to E3 cascaded pipeline with early-exit. It computes an N400-inspired Hallucination Index, triggers automatic repair inspired by P600, and verifies claims against enterprise knowledge bases. Thiago’s background in data governance and compliant AI made the “no nonsense” framing feel especially grounded for builders. We liked it for how it operationalized quality checks in a way people subtly flagged as immediately useful, and it fit the broader push toward safer agentic systems.

Jonathan Malkin demoed Jules, a Claude Code collaborator plus a “live studio” voice and video loop that turns structured skills into a real conversation. Jules uses routed workflow skills and hard gates like eliciting, challenging, adversarial review, and only then recommending, to keep the model from just agreeing or anchoring. It also ties into Whisper for speech-to-text and ElevenLabs for spoken output, with avatar visuals and OBS multitrack recording. Builders liked that it stayed lightweight, ran reliably, and even worked for Portuguese dialogue without extra prompt magic, showing how honest agentic behavior can be product-ready.

Chocks E. presented agentctl, a local Go security control layer for coding agents like Claude Code and Codex that gates risky action categories and records every decision to a JSONL trace. The tool attaches via Claude Code PreToolUse hooks or Codex MCP, then replays prior sessions against a different policy.yaml to show what would have been blocked, escalated, or approved. We liked how it prioritized a small, governable surface over heavy sandboxing, and the community response was telling (people seemed to love the counterfactual audit loop). If productized, this could become a practical safety baseline for shipping agentic tooling to teams.

Chris Forrester of Hypernym presented “Traction + Grip with Claws,” a CLI coding assistant that learns from your repo context and OS habits to keep long-running agents from going amnesiac. It runs locally with open-source solvers like Minimax 2.5 and Gemma 4, using semantic compression to manage continuous context as projects grow. We liked that it directly tackles the “agentic coders need better system qualification” pain point, and (people seemed to really click with it) because it felt usable for both quick one-offs and multi month builds.