Developer Builds Local Voice AI Assistant That Responds in Under 3 Seconds

·4 views

A developer has built a fully local voice AI assistant called Mana that runs entirely on a personal machine with an 8GB VRAM GPU, eliminating cloud dependencies and API calls. Inspired by fictional AIs like SAO's Alice and Iron Man's JARVIS, the project aimed to achieve conversational response times that feel natural rather than robotic. The key engineering decision was replacing a chain of multiple specialized models — which collectively produced 7.5 seconds of latency — with a single 4B parameter model handling transcription, reasoning, code generation, and explanation in one pass. This architecture reduced end-to-end response time to roughly 2–3 seconds, with optional 7B or 8B quality modes available when speed is less critical. The assistant uses local tools including Whisper for speech recognition and open-source TTS engines, with only the Live2D avatar component relying on an external service.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How to Extract PDF Form Field Data in Bulk Using C# and Free Spire.PDF

Enterprise workflows frequently rely on PDF forms for data collection, requiring automated extraction of field values for storage and processing. A tutorial on DEV Community demonstrates how to read interactive PDF form fields — including text boxes, checkboxes, radio buttons, list boxes, and combo boxes — using C#. The approach uses the free Spire.PDF for .NET library, which provides the PdfFormWidget class to access and iterate over all form fields in a document. Developers can install the library via NuGet and retrieve field values by casting each field to its specific widget type. The guide also notes limitations such as the free version's 10-page cap and the need for separate handling of XFA-based PDF forms.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Differential Privacy and Federated Learning Can Secure Health Data

A technical guide published on DEV Community outlines how developers can build a decentralized health data platform that protects user privacy using differential privacy and federated learning. The approach relies on two open-source libraries, PySyft and Opacus, to process sensitive biometric data — such as heart rate and step counts from Google Health Connect — directly on users' devices without exposing raw records. Mathematical noise is injected into local model gradients before any data leaves the device, ensuring that no individual's personally identifiable information is transmitted to a central server. The system uses gradient clipping alongside Laplacian noise to satisfy the formal epsilon-delta differential privacy guarantee, making it mathematically impossible for an attacker to confirm whether a specific user's data influenced the model. The guide targets Python developers with a background in basic statistics and demonstrates the workflow through a simulated health dataset and logistic regression model built with PyTorch.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Cursor Launches Google Workspace Plugins for Gmail, Drive, and Calendar

Cursor released Google Workspace plugins on August 3, 2026, enabling coding agents to interact with Gmail, Google Drive, and Google Calendar directly within the IDE. The plugins connect to Google's remote MCP endpoints, which remain in Developer Preview, meaning available tools may differ from what marketing materials describe. Security guidance recommends installing one plugin at a time using a low-impact test account, rather than connecting accounts holding sensitive customer, financial, or production data. Developers are advised to verify the live authenticated tool inventory via tools/list, since observed capabilities — such as Gmail offering draft creation but no send tool — may not match product descriptions. As Google's MCP servers are still in preview, tool inventories should be rechecked after every plugin or server update.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developers Add Native GCP Cloud Spanner Support to KEDA Autoscaler

A team running Kubernetes workloads that process jobs stored in Google Cloud Spanner tables built and contributed a new native Spanner scaler to KEDA, the open-source Kubernetes Event-Driven Autoscaling framework. The scaler allows Kubernetes to automatically adjust worker replica counts based on actual database queue depth, using any custom SQL query that returns a single integer value. Prior to this contribution, KEDA supported other GCP services such as Pub/Sub and Cloud Tasks, but lacked a Spanner integration. The implementation supports all three GCP authentication methods and is accompanied by 18 unit tests, 6 integration tests against the Cloud Spanner emulator, and an end-to-end test against a real Spanner instance. The pull request has been submitted to the official KEDA repository and documentation is available on keda.sh.

0 comments Read more at DEV Community