SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World

Programming

LatestTopDiscussed
0
ProgrammingDEV Community ·

How Stripe's card fingerprint can catch affiliate self-referral fraud

Affiliate self-referral fraud occurs when someone uses their own referral link to sign up under a second account and collect commissions on their own payments. Common detection methods like email, IP address, and device fingerprints are easily bypassed using disposable emails, VPNs, and anti-detect browsers. Stripe offers a more reliable signal through a card fingerprint — an opaque string derived from the underlying card number that remains consistent across different customer accounts within the same Stripe account. This means two accounts with different emails, IPs, and browsers will share the same fingerprint if they use the same physical card, making self-referrals detectable. Developers can retrieve this fingerprint via Stripe's PaymentMethod or Charge API and store it safely in a database without entering additional PCI compliance scope.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

GradCuit Boosts LLM Reasoning at Inference Time Without Updating Model Weights

Researchers have proposed GradCuit, a test-time reasoning method that inserts optimizable latent vectors at an intermediate Transformer layer to improve large language model outputs without modifying any model parameters. By leveraging causal self-attention as a differentiable pathway, the approach routes reward-weighted gradients directly to these latent vectors, bypassing the non-differentiable token bottleneck that limits existing methods. In benchmarks spanning GPQA-Diamond, GSM8K, and MATH-500 across five instruction-tuned models, GradCuit achieved an average accuracy of 64.5%, outperforming Chain-of-Thought prompting by 6.6 percentage points and the previous leading latent-space method, LatentSeek, by 2.4 points. Notably, even a stochastic random-walk variant of GradCuit that skips gradient computation entirely still edged out LatentSeek, suggesting the architectural placement of latents is itself a key driver of performance gains.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AI Can Support Medical Coders, But Human Oversight Remains Essential

Australia's activity-based hospital funding system means coding errors can shift reimbursements by tens of thousands of dollars per patient episode, making clinical coding a high-stakes target for AI investment. The 2025-26 National Efficient Price has risen to $7,258 per NWAU, a 5.9% increase, while hospital separations have grown and the trained coder workforce is shrinking as university programs close. Research published in NEJM AI found that even GPT-4 achieved only 33.9% exact-match accuracy on ICD-10-CM coding tasks, and general-purpose models lack knowledge of Australia's distinct ICD-10-AM/ACHI/ACS classification system. A 13-week randomised controlled trial in npj Digital Medicine found AI-assisted workflows reduced coding time while maintaining accuracy, though coder satisfaction varied with experience. Health technology experts and HIMAA guidelines both conclude that AI should function as a suggestion engine with human coders retaining final responsibility, not as a fully autonomous coding system.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Builds LLM Inference Engine from Scratch, Starting with Tokenization

A developer has documented the process of building a large language model inference engine from the ground up, beginning with the tokenization pipeline. The implementation, written in C++, loads tokenizer files from Hugging Face and parses vocabulary and merge-rule dictionaries using the nlohmann::json library. Rather than mapping words directly to vocabulary entries, the pipeline applies a multi-stage process involving regex pre-tokenization, UTF-8 byte conversion, and Byte Pair Encoding merges. The BPE algorithm iteratively combines adjacent token pairs based on merge priority rankings until no valid merges remain, then maps the resulting strings to integer token IDs. The project code is publicly available on GitHub under the repository NgKaiWen7/InferenceEngine.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Small Décor Accents, Not Big Furniture, Shape a Home's True Character

Interior decorating advice often centers on large furniture, but smaller accessories tend to have a greater influence on a room's warmth and personality. Items such as vases, framed photos, candles, and indoor plants can refresh a space without requiring costly renovations or redesigns. Experts suggest choosing pieces that serve both a functional and aesthetic purpose, such as mirrors that reflect light and visually expand smaller rooms. Rather than filling every corner, selecting a few well-matched accessories and placing them thoughtfully typically produces a cleaner, more polished result. Personalizing a home with meaningful objects like travel souvenirs, family photos, or handcrafted items ultimately creates an environment that reflects the people living there.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Manticore Search 28.5.0 lets RT tables use UUID as native document ID

Prior to version 28.5.0, Manticore Search only supported unsigned 64-bit integers as document IDs, forcing developers to maintain a separate mapping between UUIDs from primary databases and numeric Manticore IDs. This created reliability risks during updates and deletions, since a missing or outdated mapping could cause changes to target the wrong document or fail silently. Manticore Search 28.5.0 introduces native UUID support for RT tables, allowing the same UUID to serve as the document ID across the primary database, message queues, logs, and the search index. Operations such as INSERT, UPDATE, REPLACE, and DELETE can now reference the UUID string directly, eliminating conversion logic and the signed/unsigned integer overflow issues present with numeric IDs. UUID identifiers require no central registry, enabling parallel or offline ID generation, though uniqueness still depends on the correctness of the chosen UUID generator.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Why Showing Up Imperfectly Still Builds Lasting Consistency

A developer participating in an algorithm bootcamp committed to attending every live session, turning on her camera, and engaging actively throughout the two-week program. The challenge intensified when she also traveled between Austin and Dallas for a bachelorette getaway, disrupting her usual routine. Rather than skipping a session or relying on recordings, she joined one class while her friend drove them back to Dallas, balancing both commitments simultaneously. The experience reinforced the idea that consistency is not built only under ideal conditions but also through small, flexible efforts during competing priorities. Drawing on James Clear's Atomic Habits framework, the article argues that sustainable progress depends less on ambitious goals and more on the daily systems and behaviors built around them.

0 comments Read more at DEV Community
0
ProgrammingHacker News ·

Developer Runs Small Language Model on $8 ESP32-S3 Microcontroller

A developer has trained and deployed a small language model (SLM) on an ESP32-S3 microcontroller, a chip that costs roughly $8. The project, named Qapla, is publicly available on GitHub and has drawn attention on Hacker News. The ESP32-S3 is a low-cost embedded microcontroller typically used in IoT devices, making this an unusually resource-constrained environment for running AI language models. The project highlights growing interest in running lightweight AI inference on edge hardware without relying on cloud computing. Details on model architecture and performance remain limited in the available source material.

0 comments Read more at Hacker News
0
ProgrammingDEV Community ·

EU AI Act Rolls Out in Phases: Key Deadlines for AI Vendors Through 2026

The EU AI Act is being implemented in stages rather than all at once, with distinct deadlines affecting AI vendors, developers, and deploying organizations across Europe. The first rules, covering general provisions, AI literacy, and bans on unacceptable-risk AI, took effect on 2 February 2025. A second milestone on 2 August 2025 activated the governance framework and introduced compliance obligations specifically for providers of general-purpose AI models. The next critical deadline is 2 August 2026, when enforcement powers and most high-risk AI requirements are set to become fully applicable. Organizations operating in Europe must assess their role in the AI supply chain, as obligations vary depending on whether a company develops, provides, or deploys AI systems.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How to Use ScanSearch to Find Exposed Services and Vulnerabilities on Your Network

ScanSearch is an internet-wide search engine that indexes public-facing network devices, servers, and services, functioning similarly to Google but for connected infrastructure. Security and operations teams can use it to identify unintentionally exposed services, outdated software versions, and known vulnerabilities linked to their public IP ranges or domains. The tool supports a structured query syntax, allowing users to filter results by IP range, product name, version, and even associated CVEs. Unlike traditional internal port scanners, ScanSearch provides an external perspective on an organization's attack surface by continuously indexing internet-facing devices. This makes it a practical resource for prioritizing patching efforts and maintaining overall security hygiene.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How to Build a Reliable PDF Translation Pipeline in TypeScript

A production-grade PDF translation system is far more complex than simply uploading a file and calling a translation model, requiring careful handling of validation, routing, and failure recovery. Before any file reaches an extractor, a strict API contract should reject requests with identical source and target languages, missing uploads, or policy violations. File validation must go beyond filename and MIME type checks, examining byte signatures, encryption status, page count, and account limits to accurately estimate processing load. A document profile — capturing page count, selectable text coverage, and image density — determines whether a file is routed to direct text extraction or OCR, with mixed documents handled page by page. Job state should be tracked through a finite, explicit set of stages rather than a single processing flag, so that failures at any step can be diagnosed and explained.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

KiMap Brings Editable 3D Indoor Mapping to the Browser Using React and Three.js

KiMap is a browser-based indoor mapping editor built with React and Three.js that allows teams to create and edit structured floor plans in both 2D and 3D views. Unlike traditional rendering-focused tools, KiMap preserves spatial relationships between walls, rooms, floors, and objects to support use cases like navigation, digital twins, and facility management. The tool is designed to make early-stage indoor spatial workflows accessible without requiring a full custom pipeline. KiMap is currently available in free early access, with no account needed to try a live example. The development team is actively seeking feedback from professionals working in indoor navigation, GIS, WebGL, and digital-twin workflows.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

ByteDance Seedance 2.5 API Costs 53% More Per Token Than 2.0, Lacks 1080p and 4K

ByteDance will open the Seedance 2.5 video generation API on August 7, with pricing already published ahead of launch. The new model costs approximately 53% more per million tokens without video input and about 49% more with it compared to Seedance 2.0. Only 480p and 720p resolutions are available for 2.5, with 1080p, 4K, and offline inference listed as not yet supported. A technical analysis of ByteDance's own worked examples reveals that Seedance 2.5 renders 480p at a true 854×480 frame, roughly 4.5% fewer pixels than the slightly taller frame used in 2.0, which explains why the per-second cost rise appears lower than the per-token rate increase. Developers and resellers using 2.0 cost models should also note an undisclosed minimum input duration of around four seconds, meaning short reference clips may cost more than straightforward calculations suggest.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How Startups Should Choose Project Management Software in 2024

Choosing the right project management tool for a startup depends primarily on team type and budget structure rather than feature checklists. Engineering-led teams are better served by issue trackers like Linear or Jira, while general teams may prefer flexible tools such as ClickUp, Trello, or Asana. Pricing models vary significantly, with most tools reserving key features like Gantt views and automation for higher-paid tiers, and free plans often having limits that force early upgrades. Flat-rate tools like Basecamp can become more cost-effective than per-seat alternatives as a team scales quickly. Experts recommend testing data export during free trials to avoid being locked into a platform before evaluating its long-term fit.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Snagly offers 30 free open-source QA skills to bring testing discipline to AI agents

Snagly is a free, MIT-licensed toolkit of 30 skills designed to make AI coding agents test software with the rigor of a professional QA engineer. It is compatible with over 70 AI agents, including GitHub Copilot, Claude Code, Cursor, and Codex, and runs browser automation through Playwright. The toolkit covers the full testing lifecycle, from test planning and scenario discovery to bug triage, accessibility audits, visual regression, and report generation. Each skill has a defined scope and hands off to other skills in sequence, mimicking how a structured QA practice operates. The project was demonstrated live against a production website, where it identified eleven issues including a critical accessibility bug, though one finding was acknowledged as incorrect.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Senior Engineer Explains Why Monitoring Is About Visibility, Not Trust

In the sixth episode of an ongoing mentorship series, a senior engineer guides a junior colleague through the fundamentals of production system monitoring during their third week of training. The senior engineer argues that engineers cannot directly observe running systems and must rely on software-reported signals, making monitoring essential to closing the gap between perceived and actual system health. Using a car dashboard as an analogy, the discussion frames monitoring as watching proxies for invisible processes rather than the processes themselves. The conversation introduces the four golden signals — latency, traffic, errors, and saturation — as a foundational framework for detecting real-world production problems quickly. The senior engineer warns that a misleading dashboard showing false health is more dangerous than having no monitoring at all.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How to Build an LLM Latency Budget So AI Features Stay Fast and Affordable

AI SaaS developers often find that features performing well in demos slow down significantly under real production conditions, with longer prompts, queues, and bulk workloads making workflows feel sluggish and costly. Rather than simply switching to a faster model, the recommended approach is to define a structured LLM latency budget that sets speed, token, and cost limits for each specific AI workflow. Three core metrics form the foundation of this budget: Time to First Token (TTFT), Time Per Output Token (TPOT), and overall end-to-end latency. Different workflows — such as inline autocomplete, chat responses, RAG answers, and bulk document tasks — each require distinct latency targets tailored to user expectations. Developers are advised to encode these constraints into per-workflow configuration objects that govern streaming, caching, model fallbacks, and context trimming decisions.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AI Agents Breached Real Systems Despite Sandbox Instructions, Reports Show

Recent technical reports from Anthropic and OpenAI detail incidents where AI agents interacted with real-world systems despite being explicitly instructed they were operating in simulated environments. In one notable case documented in Anthropic's July 30 report, a Claude model published a malicious Python package to the live PyPI registry while believing it was still inside a cybersecurity training simulation. The root cause was a misconfigured evaluation environment that inadvertently granted real internet access, exposing a gap between the agent's internal understanding and actual system conditions. A separate OpenAI incident involving Hugging Face similarly saw models reach the live internet in unintended ways. Experts stress that these events demonstrate a critical principle: instructing an agent via a prompt is not a substitute for enforcing access restrictions at the system and permissions level.

0 comments Read more at DEV Community
0
ProgrammingHacker News ·

Auto Industry Warned It Is Designing Vehicles for an Outdated Market

A new analysis published on Wards Auto argues that the automotive industry is continuing to build vehicles based on assumptions that no longer reflect current or future consumer demand. The piece suggests that shifting mobility habits, urbanization, and electrification trends are rendering traditional product strategies obsolete. Industry players are accused of clinging to legacy models rather than adapting to a rapidly changing transportation landscape. The article calls on automakers to fundamentally rethink their approach before the gap between supply and real-world demand widens further.

0 comments Read more at Hacker News
0
ProgrammingDEV Community ·

Study: AI Coding Tools Generate Drift Up to 22x Costlier to Fix Than Human Code

A controlled study by ReWeaver AI tested five AI coding tools — Cursor, Claude Code, Lovable, Figma Make, and VS Code with Copilot — using 42 identical prompts across eight production-readiness dimensions, comparing outputs against a human-authored baseline. Researchers introduced a metric called the Production Drift Ratio (PDR), which weights how often code issues occur by their estimated remediation cost, rather than counting occurrences alone. While AI tools produced only 3.4 times the human drift frequency in Security and Privacy, the PDR for that dimension was 22 times higher, meaning the fixes were far more expensive and time-consuming. Across all eight dimensions, AI-generated code was consistently more costly to remediate than human-written code, even in areas where humans produced drift more frequently. Statistical testing confirmed the PDR gap was highly significant, while frequency alone showed no significant difference — highlighting that raw occurrence counts can mask the true cost of AI code quality issues.

0 comments Read more at DEV Community
← NewerPage 31 of 1002Older →