HalfStack Vienna 2026 • Nov 19, 2026

The Feature That Didn’t Need to Be Smart

Carmen Huidobro

How do you build a genuinely useful LLM feature without relying on cloud APIs or sacrificing user privacy? See how a skeptical developer tamed local open-weight models using strict JSON validation and zero-temperature retries.

Starting at 0:42 · Watch from the beginning

Pause
Mute Enter Fullscreen
#1 about 4 min

Early career in Mac OS X and developer relations

Navigating a twenty-year software engineering career from early Objective-C and Subversion days to Ruby on Rails, Rust, and developer education.

#2 about 6 min

Navigating the uncanny valley of generative AI

Comparing modern generative AI hype and industry burnout to historical computing milestones like the 1961 IBM singing computer.

#3 about 3 min

Scoping an AI feature for a batch file renamer

Adding an LLM integration to an Electron batch file renaming app by focusing on discrete user assistance rather than reinventing the core workflow.

#4 about 6 min

Establishing ethical boundaries and JSON schemas for AI

Enforcing ethical software standards by using local open-weight models like Mistral, making AI opt-in, and translating natural language into strict JSON instead of YAML.

#5 about 7 min

Designing the multi-step prompt and Llama CLI workflow

Building a local AI pipeline that normalizes multilingual input, infers action intent, and executes mega-prompts via the Llama CLI.

#6 about 3 min

Implementing graceful failure, JSON validation, and corpus testing

Mitigating LLM hallucinations by validating output against a strict schema, retrying failed generations at zero temperature, and running an evaluation corpus.

#7 about 5 min

Evaluating local LLM tradeoffs and user communication

Balancing the privacy benefits of local LLMs against high CPU usage, large binary sizes, and prompt ambiguity while communicating transparently with users.

#8 about 6 min

Live demonstration of local AI file renaming

Testing the local AI assistant live by translating filenames into German and attempting to replace spaces with emojis.

Matching moments

3:24 min

Preserving artificial intelligence history with localized open-source models

Audrey Kim Audrey Kim · World Congress 2026 North America

1:44 min

Summarizing developer experience and artificial intelligence companions

Robert Hoffmann Robert Hoffmann +1 · LIVE

4:29 min

Designing AI applications defensively for inevitable failures

Krzysztof Cieślak Krzysztof Cieślak · World Congress 2026 Europe

5:43 min

Pushing back against predictable AI interfaces

1:12 min

Enhancing conversational intent through modern large language models

Nathaniel Okenwa Nathaniel Okenwa · World Congress 2025

4:46 min

How generative models disrupt traditional machine interaction

Mary Grygleski Mary Grygleski · LIVE

Upcoming sessions on this topic

Starts in 6 days

Conference India 2026

November 25–26, 2026

Bengaluru, India