Community Spotlights • Issue #15

Julian Ghadially, founder of The Product LLM, presented DSPy and Prompt Optimization Experiments. DSPy is a prompt optimization package for self-improving AI programs, and he demonstrated iterative experiments on a fact checker. The project has 30k stars on GitHub and audience feedback was positive. It leverages iterative web search and a feedback loop to improve verification performance. This approach helps teams tune prompts quickly and move toward scalable AI workflows. A practical, repeatable path from prompt to product.

John Zila from Watt Data presented Canopy Agent Orchestrator: Everything Everywhere All at Once. Canopy runs many agents in parallel in isolated sandboxes, with automatic merging and conflict resolution, using Beads for in-repo task tracking and copy-on-write isolation. The live demo defines dependencies, launches parallel work, and shows a dashboard tracking progress; it is open source. Audience feedback is positive, and the demo hints at scalable, production-ready multi-agent orchestration for practical AI workflows.

Victor Wong from BEA Union Investment showcased a local-hosted version of an AI deep research engine for finance, a fully self-contained Agentic research engine with a chat interface running on a Mac Studio. It is a completely local stack that replaces all API dependencies and is optimized for financial workflows. The audience appreciated the non API architecture and the potential for cost-effective deployment in trading environments. For builders, it shows how portable, on-device AI agents can scale in finance.

Sree Pradhip from Pulsar Ventures presented LinkLibrary, an AI-powered link manager born from quick experiments into an AI-native architecture. LinkLibrary defines a core data model for links, collections, and social signals and uses semantic chunking, embeddings, and retrieval. It shows AI agents for tagging, summarization, and discovery, and compares FastAPI, Supabase (pgvector), and React. Survey feedback was positive (people loved it) and it highlights practical RAG patterns in a production setting.

Gerard Fernandez from Vivid Gallery presented Zensei, an AI-powered market regime engine that translates macro data into structured, interpretable explanations. Zensei combines a deterministic rules-based core with carefully designed prompts to deliver real-time, human-readable reasoning about regime changes, constraints, and transitions. It shows a path toward trustworthy, production-ready decision engines through regime logic and prompt design in high-stakes domains. (survey feedback hints at strong appreciation for interpretability) For builders, it signals scalable decision-support potential in finance and beyond.

Zeke John from MediScan AI presented Codecall, a demo that lets agents write and execute code in Deno sandboxes to orchestrate multiple tools like an API. Codecall provides a TypeScript library that implements programmatic tool calling and uses the MCP standard for tool use, enabling agent tasks without tool-call bloat. It demonstrates secure orchestration for AI workflows. It matters to builders because orchestration can stay lean while handling complex tasks.

Pablo Diego Rosell from Adri.Care presented Engineering a Consistent Image Generation Pipeline for Augmentative and Alternative Communication, showcasing an AI-powered workflow for matching voice dictation to pictograms. The system maps user intent to pictograms using an offline prompt-to-image matcher, supported by an online intent recognition model to handle edge cases. This approach maintains low latency while ensuring predictability for a production-ready user experience. When new pictogram generation is needed, a constrained image generation pipeline with AI-based prompt augmentation enforces a consistent visual style. A compact AI scene builder then assembles full compositions from modular data structures, enabling reliable and scalable pictogram-based communication.

Travis McQueen from Mimory presented Synthesizing World Aware Agentic Assistants, a just-in-time assembly of agentic helpers for dynamic environments. The core offers a visualizer and the system prompt construction, mapping objects and assembling an agent from their relations into a condensed physical-digital graph. It highlights a bottom-up approach to fragmented systems, with strong potential as a production-ready IoT platform. Survey feedback hints it resonates with builders. Builders take away how modular orchestration can scale in real settings.

Keanan Koppenhaver from FloorboardAI presented How I Built a Fake TechCrunch with Claude Code and the Claude Agent SDK. The project takes a public GitHub URL and generates a satirical TechCrunch article using a Claude setup that analyzes the repo and writes copy. Built with Claude Code and featuring Nano Banana visuals, it hints at automated documentation workflows that could scale into productized content pipelines. Audiences reacted warmly, appreciating the practical agent architecture and the real-world usefulness.

David Botos, Co-Founder of Connect211, showed a CLI demo that converts unstructured community services schedules into RFC 5545 RRULE compliant JSON using a compact 0.6B model. The run highlights quantization-aware training and a lightweight pipeline that can tackle real-world data without big hardware. Subtly, the audience seems impressed by the 0.6B approach and QAT potential for production LLM data tasks. If it becomes a product, it could power low-cost, scalable data pipelines for community services across states.