Developer builds AI law firm template where buyers own their own API keys

·3 views

A developer building CounselForm AI 2.0, a Framer template for law firms, designed the AI layer around a Bring Your Own Key (BYOK) model so buyers retain full control of their AI infrastructure. The template features a tool called Ask Lexora, which lets website visitors explore firm expertise through either structured guided questions or a natural-language AI mode. To keep API credentials secure, a Cloudflare-based server-side layer sits between the Framer frontend and the AI provider, preventing secret keys from being exposed in client-side code. The AI is grounded in a curated Knowledge Sync workflow using approved firm content, rather than allowing the model to generate unsourced or fabricated firm-specific answers. This architecture ensures each buyer bears their own inference costs and is not dependent on the template seller's infrastructure or subscription.

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)

Purdue Study: Removing Teacher Model Boosts Small AI Reasoning Performance

Researchers Yi Ding and Ruqi Zhang from Purdue University found that teacher supervision in on-policy distillation introduces noise rather than useful guidance, with larger teacher models producing even more signal degradation. Their analysis revealed that training gains in standard distillation come almost entirely from penalizing low-probability tokens, not from transferring teacher knowledge. Building on this insight, the team developed On-Policy Self-Adaptation (OPSA), which replaces the teacher with entropy-adaptive penalties calculated directly from the student model's own distribution. Tested on Qwen3-1.7B, OPSA improved AIME24 Avg@32 performance by 35.41 points, outperforming standard teacher-guided distillation by 16.77 points. The approach also eliminates the computational overhead of loading and running a separate teacher model during training.

0 comments Read more at DEV Community

ProgrammingHacker News ·

Quill, YC W20 Startup, Seeks a Fullstack Software Engineer

Quill, a startup that participated in Y Combinator's Winter 2020 batch, is currently hiring a fullstack software engineer. The job listing was posted on Hacker News, a popular tech and startup community platform. The posting received no points or comments at the time of reporting. No further details about the role's requirements or compensation were publicly available in the listing.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Claude Code's Workflow Tool Offers Real Power but Suffers from Poor Adoption and Flawed Defaults

Claude Code features a dynamic workflow system built on a JavaScript-like scripting dialect that lets users orchestrate multi-agent tasks through primitives such as agent, pipeline, parallel, and nested workflow calls. Unlike traditional agent-team setups, Claude Code's harness manages orchestration directly, reducing token consumption and avoiding dependence on a lead model to babysit subagents. Despite these structural advantages, the feature has seen minimal real-world adoption — a search of the top 20 GitHub results for 'ultracode' found zero repositories shipping actual Claude Code workflows, with most instead porting the concept to rival tools like Codex. A key flaw undermining practical use is that Claude is not properly instructed to select appropriate models per workflow phase, defaulting all worker agents to the same heavyweight model as the main session, which drives up costs unnecessarily. Critics and users alike view the current workflow setup as impractical and expensive, pointing to gaps in documentation and default configurations as the primary barriers to wider adoption.

0 comments Read more at DEV Community

ProgrammingHacker News ·

Developer runs 125B-parameter AI model on 16GB Mac using SSD streaming trick

A developer has released an open-source tool called Slotstream that enables running Qwen3-235B-A22B-Flash (a 125-billion-parameter model normally requiring over 100GB of memory) on Macs with as little as 16GB of RAM. The tool achieves this by using expert-offloading and SSD streaming to compensate for limited onboard memory. Built natively for Apple hardware using MLX and Swift, it can run the model at around 12 tokens per second on a 48GB Mac. Slotstream ships with an auto-mode that automatically balances memory usage against inference speed. The developer plans to add support for speculative decoding via the MTP module in a future update.

0 comments Read more at Hacker News