A.I. in Production: Learnings & Findings

TLDR
———-
We will have multiple presenters speaking on the A.I. implementations they've shipped to production, and are working on now.

More details about the event here

Topics include:

• what we’ve pushed to production, how it’s going
• what we’ve found (tools, patterns, vendors) along the way
• Raspberry Pi + A.I. demo

SCHEDULE
—————-

• 6:30 PM: Doors Open & Socialize
• 7:00 PM: Announcements, Presentations begin
• 8:00 PM: Presenters finish, Extended Socialization
• 8:30 PM: Doors close

PRESENTERS
———————

## Featured Presenters

## Scott Lichtenstein

## When AI Picks Up the Phone: Making Technology Work for the Non-Tech Crowd

Most small business owners aren’t sitting around training models or writing prompts, they’re answering phones, serving customers, and trying not to miss sales. This talk explores how AI can be designed to meet them where they are, through tools that feel natural, solve real problems, and quietly deliver the power of AI without any tech know-how.

## David Pickett

## Fast, Small, Accurate(-ish): A 4-Bit LLM Tour with Qwen3-32B

• A 10-minute, data-first comparison of 4-bit quantization formats for the Qwen3-32B model across three environments: Hybrid CPU/GPU, NVIDIA-only, and Apple Silicon.
• Benchmarks focus on accuracy (LiveBench; possibly Aider Polyglot) and performance: prompt processing, time-to-first-token, steady token rate, and multi-request concurrency/throughput.
• Formats covered (time permitting): GPU (mxfp4, nvfp4, AWQ-4bit, ExLlamaV3 4bpw), CPU/GPU via GGUF (Q4_0, Q4_K_M, UD‑Q4_K_XL, IQ4_NL), and Apple MLX (MLX 4‑bit, MLX 4‑bit DWQ).
• Goal: provide practical rules of thumb on speed/accuracy/memory trade-offs and a simple decision path to choose the right 4-bit format for specific hardware and workloads. No hype—just results.

## Ben Ward

# Building Real-Time Conversational AI for Mobile: Architecture Lessons from ConvoLive

• Deep dive into ConvoLive’s real-time conversational pipeline on React Native, orchestrating multiple AI services.
• Components: OpenAI GPT with scenario-specific prompt engineering; Google Cloud TTS with streaming; 3D avatar rendering and lip sync.
• Focus on engineering tactics: reducing end-to-end latency, optimizing token usage/costs, handling mobile constraints.
• Concrete examples: generating dynamic language-learning scenarios and prompts.
• Production lessons: rate limiting, offline/spotty connectivity strategies, multilingual TTS quirks (e.g., dialect/locale issues like “Spanish isn’t one Spanish”).

## What to expect

#### Presentations on AI projects shipped to production

#### Discover essential tools, patterns, and vendors

#### Live Raspberry Pi + AI Demo!

Crowd
* What have *you* learned over the last few months and years? We are especially interested in anything that is running in production!

Sponsors & Amenities
🌺 Meeting space generously provided by Hub Coworking Hawaii and Hawaii Center for AI 🌺

🤖 Food and drinks will be provided by Flowing Blue 🤖

Event cost: FREE

Want to help out? Send us a message!