SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Uttar drafts grounded customer replies for small sellers using a local open-weight model

0
·3 views

Uttar is an app that helps small Instagram and WhatsApp sellers answer routine customer questions about price, cash-on-delivery, and delivery times. It matches each message against the seller's own product catalog and shop rules, then uses a locally run model to produce a few draft replies in the seller's language and past voice. Sellers review and send the drafts themselves, and the app flags questions it cannot answer from the seller's data instead of guessing. The developer chose a self-hosted model because per-token API pricing would exceed the planned subscription cost, and it keeps customer conversations on the device. Early testing also showed the model inventing policies, which the developer addressed with explicit rules against unlisted answers.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Flower-counting pipeline scores perfectly only because its simulator and detector are tuned to each other

The author argues that the Bloom-Counter project's reported 99.7% detection recall says little about real-world performance. The simulator paints flowers in colors chosen to fit the detector's HSV ranges, and its scoring compares only totals rather than matching individual objects. To test this, the unmodified detector was run on simulated scenes under altered lighting and with leaf occlusion and clustered blooms added. Stricter per-object scoring still gave a perfect score on the original scenes, which the author attributes to the co-tuned design. The stress tests were then used to measure how the detector fails and whether two simple remedies help.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AppTutorial: Open-Source Windows Tool Gives On-Screen, Step-by-Step Software Guidance

A developer has built AppTutorial, a Windows tool that offers interactive, on-screen guidance for complicated software such as Excel, IDEs, and web apps. Users press a hotkey over an app, and the tool captures the window, identifies the software, and proposes tutorials or accepts a custom goal. It runs entirely on a local open-weight vision model, Qwen2.5-VL served through Ollama, so no cloud AI or API key is required. Each step is highlighted on screen and verified against the live display, with Excel currently the most thoroughly tested. The project was submitted to the Hacktoberfest Open-Source AI Challenge, and friends and acquaintances who tested it gave generally positive feedback.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Cheap Proxy Pricing Hides the Real Cost of Scraping, Says New CLI Tool

A proxy reseller's advertised rate of $0.49 per 1,000 requests looks inexpensive, but scraping teams are billed for attempts rather than usable pages. Blocks, soft challenge pages, timeouts, and silently incomplete responses all count toward the invoice without producing data. A developer built check-proxy-tax, a roughly 200-line standard-library Python CLI, to measure the gap between sent requests and usable responses. The tool counts a response as usable only if it returns an exact 200 status and, optionally, contains a required marker string. Its output reports usable percentage, effective cost per 1,000 usable pages, and a cost multiplier that can be used to evaluate proxy plans.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

MyZubster Details Session-Bound Actions and Identity Model for Its Metaverse

The MyZubster team has described its metaverse architecture, including a public interface showing 25 characters and 7 verified explorers. Its backend ties world actions to authenticated sessions and checks lease validity before any state change. Authentication, authorization, and lease status are treated as separate checks. Contributor identities drawn from GitHub activity are kept distinct from accounts, professional profiles, and verified external identities. Characters are added only after explicit contributor opt-in, so no one is represented without consent.

0 comments Read more at DEV Community