DEV Community: sharanjit singh

Why Maintaining Web Scrapers for Ecommerce AI Projects Breaks Down at Scale

sharanjit singh — Sun, 27 Sep 2026 03:29:12 +0000

Every developer building a price comparison tool, recommendation algorithm, or retail intelligence pipeline usually starts the same way: writing a quick web scraper in Python.

Using libraries like Beautiful Soup, Scrapy, or headless browsers like Playwright and Puppeteer, grabbing a few thousand product records feels trivial during early development.

However, once you scale that pipeline to track hundreds of thousands of SKUs across multiple retail domains, scraper maintenance quickly consumes your entire engineering week.

The Hidden Technical Debt of Scraping at Scale

The moment you push custom scrapers to production environments, you encounter systemic operational blockers:

Frequent DOM and Schema Shifts: Ecommerce websites deploy continuous front-end revisions and A/B layouts. A minor CSS class mutation or altered microdata attribute breaks parser rules silently, corrupting your downstream database tables.
Aggressive Fingerprinting & Anti-Bot Systems: Modern retail platforms enforce complex TLS fingerprint analysis, browser environment inspections, and Cloudflare challenge loops. Bypassing these consistently requires expensive residential proxy pools and complex headless browser spoofing.
Variant Attribute Normalization: Extracting raw HTML does not give you clean data. Standardizing nested parent-child variant structures, multi-currency values, and dynamic discount percentages requires significant transformation overhead before analytics can begin.
Bandwidth & Compute Waste: Running cluster instances to execute headless Chromium sessions just to parse dynamic JavaScript consumes substantial cloud compute.

Moving From Scraping to Pre-Structured Data

Recently, while evaluating ways to remove scraping infrastructure overhead from a data pipeline, I explored Datasets.store.

Instead of an on-demand proxy service or dynamic scraper API with fluctuating request credits, the platform functions as an off-the-shelf B2B data marketplace. They aggregate, normalize, and verify retail datasets covering:

Over 289 million product records across 80+ global platforms and 24 countries
SKU-level product specs, category hierarchies, and brand taxonomy
Historical price records, promotional discounts, and localized stock availability
Direct downloads in developer-friendly tabular formats like CSV and Parquet

Why Parquet and Direct Data Ingestion Win for AI & BI

If you are training custom retail LLMs, building dynamic pricing models, or feeding BI warehouses like Snowflake and BigQuery, file formats matter.

Querying raw scraped JSON dumps is slow and memory-intensive. Having normalized data delivered directly in Parquet format provides clear advantages:

Columnar Storage: Dramatically speeds up analytics queries using DuckDB, Polars, or pandas by reading only required feature columns.
Type Safety: Preserves explicit schema definitions, avoiding messy type errors when reading pricing floats and inventory integers.
Zero Scraper Maintenance: Eliminates the ongoing cost of proxy rotations, captcha solvers, and broken selector debugging.

How Do You Handle Large-Scale Retail Data?

For developers working with ecommerce catalogs or retail AI: do you still run your own scraping clusters, or have you transitioned to purchasing pre-collected, structured datasets to keep your engineering focus on core modeling?

I would love to hear your pipeline architectures and lessons learned in the comments below.

Structuring 100,000 Public Profiles: Building a Massive Creator Directory

sharanjit singh — Sat, 26 Sep 2026 05:12:53 +0000

Organizing massive amounts of unstructured data is a classic developer challenge. I recently helped manage submissions for a large scale project called Best OnlyFans Reviews, an independent editorial directory that indexes over 100,000 public creator profiles.

The primary goal of the project was to bring structure and transparency to the creator subscription economy. Instead of forcing users to rely on scattered social media links, this platform aggregates public data into a searchable, objective directory to help users make informed choices.

Core Technical Challenges and Features

Handling this volume of dynamic data required a focus on clear architecture and search performance. Here are the main components of the build:

Data Aggregation at Scale: Indexing and standardizing public metrics across more than 100,000 unique creator profiles.
Algorithmic Evaluation Framework: Implementing a structured six-factor scoring system to quantify profile clarity, update frequency, pricing versus perceived value, category accuracy, community engagement, and trust signals.
High Performance Search and Filtering: Building a responsive search interface that allows users to query specific niches and compare public monthly subscription rates without heavy database load times.
Dynamic Price Tracking: Managing high data volatility to accurately track standard monthly rates, active discount promotions, and verified free trial pages.

Building a reliable discovery engine requires a strong focus on data accuracy, transparent methodology, and fast query resolution. By focusing entirely on public metadata, the platform serves as a practical, objective research tool for adult audiences seeking context before committing to subscriptions.

You can check out the live directory, search filters, and pricing methodology directly at Best OnlyFans Reviews.

Found an open-data project mapping OnlyFans economics with free CSVs and a JSON API

sharanjit singh — Sat, 26 Sep 2026 04:33:59 +0000

I recently came across an interesting open-data project while looking into public financial data feeds, and I wanted to share it here for developers, data analysts, and researchers: OnlyFansStatistics.com.

Most numbers floating around the creator economy are either locked behind steep enterprise paywalls or based on anecdotal claims. What caught my eye about this site is that it treats creator-economy metrics like open software and reproducible data.

Where the data comes from

The project pulls primary numbers directly from official UK Companies House regulatory filings submitted by Fenix International Limited (the parent company). They triangulate those figures with creator-economy research panels and business press reporting, then re-audit the whole dataset monthly against new filings.

They even have a public audit and methodology breakdown available at onlyfansstatistics.com/methodology.

What they provide for devs and data people

Instead of just rendering static text summaries, the creators turned the entire database into usable developer and analyst assets:

A live JSON master feed providing every published figure alongside source IDs and audit revision timestamps.
13 downloadable CSV datasets ready for spreadsheet modeling, Pandas dataframes, or SQL ingestion.
16 custom charts provided in vector SVG, standard PNG, and 1200x630 social card formats.
Over 50 individual data pages breaking down platform gross revenue, creator earnings distribution, user growth curves, and geographic splits.
A free 1.5 MB media press kit with curated figures for editorial and academic teams.

Free and open reuse

Everything on the site (the raw CSV files, SVG charts, and JSON endpoints) is released for free embedding, research, and application development under standard attribution.

If you are building creator-economy tooling, testing data visualization pipelines, or training models on real-world subscription platform datasets, it is worth checking out:

Website: https://onlyfansstatistics.com/

Engineering Interactive Hardware: Real-Time Feedback and AI Architecture in Humanoid Companion Robotics

sharanjit singh — Sat, 26 Sep 2026 03:38:23 +0000

Building interactive hardware requires solving hard synchronization problems between physical sensors, microcontrollers, and conversational AI software. Over the past few years, the consumer companion and humanoid robotics sector has shifted away from purely static mechanical molds toward responsive cyber-physical systems.

In this overview, we look at the core architectural layers required to build responsive companion robotics: sensor acquisition, heating regulation, and low-latency voice pipeline integration.

1. Distributed Sensor Pipelines and Tactile Input

Realistic physical interaction requires distributed tactile arrays embedded beneath flexible polymer surfaces like medical silicone or thermoplastic elastomer (TPE).

Typical sensor pipelines use capacitive or piezoresistive sensor grids mapped across distinct touch zones:

Microcontroller tier: Low-power ESP32 or ARM Cortex microcontrollers poll capacitive sensor arrays via I2C or SPI at 50Hz to 100Hz.
Threshold filtering: Signal smoothing algorithms filter out ambient capacitance shifts caused by room humidity or temperature drift.
Event dispatch: Normalized touch events trigger local reactions (such as activating local actuators) or broadcast payloads to an onboard central logic unit.

2. Closed-Loop Thermal Regulation

Unlike industrial robotics, companion robotics prioritize natural surface temperatures. Simulating human body temperature (36.5°C to 37.5°C) requires closed-loop PID control loops:

Heating elements: Flexible polyimide heating films are routed along internal alloy skeletal frames.
Thermal monitoring: NTC thermistors provide continuous surface feedback to prevent overheating and protect silicone layers.
Power management: Pulse-width modulation (PWM) governs current draw to maintain safe thermal plateaus without degrading battery runtime.

3. Voice Interaction and Edge-to-Cloud AI Pipelines

Voice interactivity requires a pipeline capable of sub-second turnarounds. The pipeline splits execution across local edge hardware and remote cloud endpoints:

Local wake-word detection: Edge neural processing units (NPUs) process lightweight keyword models locally to preserve battery life and privacy.
Streaming speech-to-text (STT): Audio streams send chunks over WebSocket connections directly to low-latency transcription services.
Context orchestration: Conversational models maintain user state and personality prompts, returning streaming audio responses via neural text-to-speech (TTS).

Platforms operating at the intersection of consumer robotics and adult companion customization, such as Sex 'n Dolls, highlight how these mechanical skeletons, heating elements, and conversational software components come together into finished products.

4. Hardware and Software Integration Challenges

Developers building interactive humanoid companions face several engineering hurdles:

Structural fatigue: Internal steel skeletons must allow multi-axis posing without putting excessive mechanical tension on internal wiring harnesses.
Latency ceilings: If sensory feedback or voice responses take longer than 800ms, user immersion breaks.
Privacy isolation: Companion electronics must isolate local audio and diagnostic logs to guarantee complete data confidentiality for the end user.

As edge computing chips become smaller and generative audio models become faster, companion robotics will continue evolving from static figures into fully responsive interactive systems.

Why Most Self-Improvement Software Fails: The Engineering Case for Full-Stack Life Systems

sharanjit singh — Fri, 25 Sep 2026 06:01:27 +0000

As developers, we know the cost of isolated microservices that do not communicate with each other. When data pipelines are fragmented, systems break down.

Yet in personal development, almost all software operates in extreme isolation:

Fitness apps track lifting volume and calorie counts.
Budgeting apps track personal balance sheets.
Productivity apps track daily task completion.
Social apps manage networking and contacts.

None of these systems pass state to each other. In reality, biological and psychological systems are deeply linked. Poor sleep degrades next-day executive function and code output. Low physical energy tanks social confidence. Chronic financial stress elevates cortisol, causing cognitive fatigue and sleep issues.

When one subsystem crashes, the whole stack suffers.

The Problem With Modern Productivity Stacks

Most users try to patch this problem by stitching together five different mobile apps and twenty unverified supplements recommended by fitness influencers.

The core issues:

Unverified Marketing Claims: Consumers spend money on compounds and protocols with zero clinical trial backing.
Siloed Data: Working out hard does not help if recovery and sleep protocols are completely ignored.
Lack of Personalized Starting Points: Most productivity apps throw generic dashboards at users without diagnosing their primary bottleneck first.

A New Pattern: Integrated Performance Systems

I recently came across an interesting platform tackling this architectural challenge called Super Achiever Club.

Instead of treating health, wealth, and social development as separate entities, the platform models them as a single connected system:

Evidence-Based Health Engine: Rather than pushing generic wellness advice, the platform evaluates clinical evidence behind supplements (such as creatine, Tongkat Ali, ashwagandha, magnesium, and omega-3), scoring products on ingredient clinical backing, proper dosage, and formulation transparency.
Wealth and Decision Frameworks: Practical modules for business, personal finance, investing, and repeatable daily execution habits.
Social and Communication Dynamics: Frameworks to build real-world confidence, negotiation, and interpersonal communication.
AI Coach Diagnostic: An assessment quiz that maps a user's current baseline and builds a tailored starting point rather than offering one-size-fits-all advice.

Key Takeaway for Builders

Building tools for human performance requires the same systems thinking we apply to software architecture: holistic integration always beats fragmented micro-tools. When products evaluate clinical evidence objectively and connect interrelated domains, users make better decisions and build lasting habits.

What are your thoughts on integrated self-improvement tools versus dedicated single-purpose apps?

Deconstructing Niche E-Commerce UX: How ATLAS 1 Maps Alternative Aesthetics

sharanjit singh — Fri, 25 Sep 2026 05:17:10 +0000

When building an e-commerce platform for mainstream fashion, standard UX patterns work reliably. You group items by department, filter by color, size, and material, and show a standard grid of product cards.

However, building for niche subcultures like techwear, warcore, and cyberpunk fashion introduces unique information architecture challenges.

Shoppers in these communities rarely search for a simple garment in isolation. Instead, they look for visual coherence, functional hardware integration, and a specific narrative aesthetic.

I recently spent time analyzing ATLAS 1, a platform focused on futuristic clothing, tactical streetwear, and alternative aesthetics, to examine how they structured this specific problem.

Here are a few UX and architectural decisions that stand out from an engineering and design perspective:

Dual Aesthetic Universe Architecture: Rather than relying strictly on standard categories like tops or bottoms, the catalog splits into two distinct visual themes: SYSTEM (dark, utilitarian, tactical, and dystopian gear) and ASCENSION (experimental, sculptural, and avant-garde cuts). This taxonomy immediately orients the user based on aesthetic intent rather than just item types.
Complete-Look Composition over Standalone SKUs: One of the biggest friction points in subcultural fashion is assembly. Without guidance, matching multi-pocket utility pants with an appropriate technical jacket or modular harness can be overwhelming. Organizing products into 25+ curated complete looks across 12 distinct sub-genres makes styling accessible directly from the catalog.
Guided Discovery via the ATLAS Coach: Implementing an interactive recommendation flow (the ATLAS Coach) helps narrow down garments according to specific body fit, aesthetic preferences, and functional goals. This bridges the gap between static browsing and interactive styling assistance.
Asset Optimization for High-Detail Hardware: Specialized techwear relies heavily on hardware details like magnetic buckles, waterproof zip seams, and multi-strap configurations. Balancing full-bleed, high-resolution imagery with quick initial load times requires disciplined asset compression, CDN delivery, and responsive image loading across mobile and desktop viewports.

For developers and designers working in high-concept retail spaces, treating product discovery as world-building rather than just inventory filtering is a great way to drive engagement.

You can check out how the catalog and navigation are put together over at ATLAS 1.

How do you approach multi-layered catalog taxonomies when building niche e-commerce storefronts?

Building Beyond Static MP3s: How Adaptive Audio and AI Reshape Emotional Wellness Apps

sharanjit singh — Thu, 24 Sep 2026 11:11:06 +0000

When building digital wellness or meditation apps, the conventional engineering pattern has been fairly static: upload a library of high-bitrate MP3 files to S3 or a CDN, pair them with a clean media player in React Native or Swift/Kotlin, and let users pick from a list.

While that approach works for basic utility, it runs into an immediate UX bottleneck: habituation. The human brain recognizes repetitive acoustic patterns very quickly. Once a user hears the identical loop or tone a handful of times, its calming effect degrades, and the session begins to feel artificial.

The Challenge of Responsive Sound

To create genuine nervous system regulation, sound needs to behave more like a dynamic environment than a pre-rendered playback file. This is the core architectural challenge tackled by U4RIA, an emotional wellness platform designed around a proprietary dynamic audio engine.

Instead of fixed loops, an adaptive audio architecture requires:

Multi-track procedural layering that modulates acoustic frequencies in real time
Seamless asset streaming and caching so background transitions remain stutter-free
Low-latency state management to adapt ambient soundscapes to real-time check-ins

Structured Interventions with the SOS Wellness Wheel

A major trap in mental health software is menu fatigue. When a user opens an app while dealing with acute stress or insomnia, navigating deep category trees causes cognitive friction.

U4RIA handles this through a structured framework called the SOS Wellness Wheel across 13 core modules. Instead of browsing hundreds of recordings, users map their current friction point to receive immediate, targeted interventions:

BreatheSync exercises for guided nervous system regulation
Frequency-based restorative soundscapes
AI-personalized daily affirmations and sleep journeys

Scaling and Community Traction

A dynamic sound engine must perform smoothly across diverse hardware profiles without draining the device battery during long nighttime sleep sessions. Overcoming these real-world mobile constraints has helped the platform scale organically:

Reached over 200,000 organic downloads across 160 countries
Maintains a 4.9 rating on iOS and a 5.0 rating on Android
Covered in publications like Wellbeing Magazine, Business Today Global, and MSN

For developers and founders building in the digital health space, moving intelligence into the audio generation layer rather than treating sound as static media opens up massive product possibilities.

You can inspect the platform on the U4RIA official hub or test the mobile implementation via the U4RIA download page.

Solving the AI Context Problem: How Byblos AI Bridges Product Specs and Marketing

sharanjit singh — Fri, 18 Sep 2026 06:11:25 +0000

If you have tried using AI agents to build or scale software projects, you probably hit the same wall: isolated context.

You use one tool to draft a PRD, user stories, and feature specs. Then you switch over to another tool to work on branding or launch copy, only to spend twenty minutes copy-pasting and re-explaining the core architecture and value proposition. Context fragmentation slows teams down more than writing the specs manually.

I recently spent time digging into how Byblos AI approaches this problem, and their shared project brain architecture is worth a look for developers and tech founders.

A Unified Project Brain

Instead of treating tasks as disconnected one-off prompts, Byblos AI runs specialized agents against a single, continuous project context:

AI Product Manager: Structures technical requirements, feature roadmaps, and target user personas.
AI Brand Manager: Translates product decisions into unified messaging, core value propositions, and visual guidelines.
AI Campaign Manager: Generates multichannel launch assets and promotional workflows that pull directly from active product specs.

When you tweak a feature or pivot a roadmap priority, you do not have to rewrite prompts for your marketing workflows. The state persists across agents.

Data Security and the Upcoming Co-CTO

Two technical aspects stood out:

Data Privacy by Design: The platform is fully GDPR compliant and hosted on EU cloud infrastructure, making it viable for teams concerned about data sovereignty.
Technical Pipeline (Co-CTO): They are rolling out a Co-CTO suite designed to handle technical execution. This will integrate AI Developers, automated QA Engineers, DevOps agents, and code generation routines directly into the same shared state.
Human in the Loop: They offer optional human expert reviews as on-demand add-ons if you need an experienced engineer or strategist to validate decisions.

If you are curious about how autonomous agent orchestration can eliminate context-switching overhead across product and growth pipelines, check out Byblos AI.

How are you currently synchronizing state and project memory between technical planning tools and marketing systems?

Applying LLMs to High-Density Time-Series Telemetry: A Look at TuneWorks

sharanjit singh — Thu, 17 Sep 2026 06:18:54 +0000

Working with high-frequency time-series data is notoriously challenging for standard Large Language Model architectures. A typical standalone automotive ECU records dozens of channels—manifold pressure, ignition angle, wideband lambda, throttle position, coolant temperature, and injector duty cycle—at sample rates anywhere from 20 Hz to 100 Hz.

A single 20-minute track session or chassis dyno pull generates hundreds of thousands of CSV rows. Passing that raw payload directly into an LLM context window is practically impossible due to token limits, latency, and hallucination risks.

I recently came across TuneWorks, an interesting platform tackling this problem specifically for Haltech ECU logs. Rather than attempting to feed raw CSV telemetry directly to a model, it highlights an effective architecture pattern for domain-specific engineering tools.

The Engineering Problem: Telemetry Noise vs Signal

Traditional datalog viewers like MegaLogViewer or desktop scatter plot engines force tuners to write custom formulas and filters manually to find lean spikes or intermittent misfires.

To bridge conversational AI with this dense data, a platform must handle several preprocessing steps:

Pre-filtering and Anomaly Detection: Isolating specific operational windows (such as wide-open throttle pulls or abrupt AFR spikes) before passing structured summaries to the LLM.
Context Injection: Telemetry data is meaningless without physical context. An engine running lean on tip-in might be expected depending on manifold design, or it could signal a failing fuel pump. Correlating telemetry with car modifications, suspension settings, and ambient conditions provides the necessary ground truth.
Side-by-Side Channel Scrubbing: Marrying tabular and visual chart overlays with natural language answers so the user can visually verify AI recommendations against the raw sensor trace.

A Sensible Model for Niche Developer & Engineering Tools

Beyond the technical data pipeline, TuneWorks takes a practical approach to distribution and monetization in a niche market. Instead of locking drivers into a high recurring monthly subscription for software they might only use on track weekends, it operates on a pay-as-you-go credit system starting at $10.

For developers building vertical AI applications, time-series sensor ingestion is one of the most promising frontiers. The key takeaway from tools like TuneWorks is that 90% of the solution lies in smart data downsampling, normalization, and relational build context before the model ever sees a prompt.

Has anyone else here built interfaces for conversational time-series querying? How do you balance client-side data downsampling with server-side LLM context limits?

Solving the "Cheap Gas" Paradox with Real Detour Math Instead of Map Pins

sharanjit singh — Sun, 13 Sep 2026 07:39:44 +0000

Most navigation and fuel-finding apps treat gas prices as a basic spatial pin problem. They take a radius around your current coordinates, fetch the cheapest listed prices, and slap pins on a map.

The issue with this UX pattern is that it leaves the computational heavy lifting to the driver. If station A is $3.40 per gallon but sits 8 miles away, and station B is $3.43 per gallon two blocks away, driving to station A is an automatic net loss. Between traffic lights, idle time, and vehicle fuel consumption during the detour, the $0.03 delta vanishes immediately.

I recently tested an app called GasMeUp that handles this problem the way an engineer would expect it to be handled: by computing total door-to-door cost instead of showing raw prices.

How the Calculation Works

Instead of making drivers do mental calculations while driving, the app runs a multi-variable calculation:

Pulls live, licensed fuel rates for surrounding stations
Evaluates true driving distance and estimated detour duration
Incorporates the vehicle fuel tank capacity to determine the actual break-even threshold
Ranks the stations and selects a single "Best Pick"

A 3-cent difference means almost nothing on an 11-gallon tank, but it matters on a 30-gallon truck. Factoring vehicle specs directly into the ranking algorithm eliminates false bargain detours.

Solving Highway Off-Ramp Traps

The feature that stood out most to me is the dedicated Highway Mode. Anyone who has done long road trips knows the first-exit trap: stations sitting right off the interstate ramp often mark up their fuel because they rely on driver fatigue and uncertainty.

On freeways across all 50 states, the app detects the route and scans stations ahead exit by exit. Crucially, it prices them based on how far off the ramp they actually sit. This gives drivers visibility into whether it is worth pulling off now or cruising to the next exit.

Clean Utility Architecture

From a product and monetization perspective, it is refreshing to see a mobile utility that stays away from the typical ad-tech playbook:

No user accounts or login friction
No social feeds or engagement loops
Zero ad networks or location tracking brokers

It runs on a straightforward subscription model ($1.99/month or $17.99/year with a 7-day trial) to support the live data feed costs without selling user location history.

If you are interested in pragmatic utility tools or do a lot of driving, check it out here: GasMeUp

Curious if anyone else has built or used route-optimization tools that factor opportunity cost into spatial navigation.

How to Pick the Best AI Chatbots and Companions Without the Noise

sharanjit singh — Tue, 08 Sep 2026 04:18:53 +0000

Over the last year, conversational AI has shifted well beyond standard syntax helpers and productivity assistants. We have seen a massive surge in character-based AI, immersive roleplay bots, and dedicated virtual companions.

The main headache right now is discovery. Every platform claims to offer human-level reasoning, endless context windows, and realistic voice integration. In practice, however, the landscape is fragmented:

Some platforms invest heavily in long-term memory and coherent context retention.
Others lean into open character creation and massive community prompt libraries.
Many prioritize uncensored creative writing, while others enforce strict guardrails.
Pricing ranges from standard monthly subscriptions to confusing credit systems that drain fast when using voice models.

Testing each platform individually is both expensive and time-consuming. I recently found a lightweight tool called Chat Bots Compared that tackles this discovery problem from a practical angle.

Instead of presenting an overwhelming directory with dozens of options, the site guides you through an interactive six-question quiz. It analyzes your priorities across conversational realism, character customization, voice features, community setups, and pricing models, then matches you with the platform that fits your exact workflow.

For anyone looking to compare AI companions or research where consumer AI chatbots currently stand without wasting hours on sign-ups, check out Chat Bots Compared. It offers clean comparisons that help you cut straight through the promotional noise.

How do you usually evaluate companion and roleplay models? Are you paying more attention to raw context length, voice fidelity, or moderation boundaries?

Found a clean, free tool to convert JSON and CSV to Excel (includes an API)

sharanjit singh — Sat, 29 Aug 2026 05:50:30 +0000

I frequently need to take raw JSON payloads from APIs and hand them over to product managers or clients as readable spreadsheets. Finding a decent JSON to Excel converter that does not demand an account registration or look sketchy regarding data privacy is surprisingly difficult.

I recently came across JSON Support and it has been incredibly useful.

Here is why it is worth checking out:

It is completely free and requires zero signups.
They guarantee privacy by not storing any of the pasted data or uploaded files.
It provides a developer API with a free tier for automated data conversions.
It supports reverse conversion so you can turn Excel or CSV files back into JSON format.
It includes inbound webhooks to easily plug into no-code automation workflows like Zapier.
There is a built-in JSON validator and formatter for quick cleanups.

If you are building data pipelines or just need to quickly format an API response into an .xlsx file for your team, it is a great utility to have bookmarked.