DEV Community: dengkui yang

AI Stock Prediction: a live model for the next US ticker move

dengkui yang — Sat, 12 Sep 2026 14:42:30 +0000

AI Stock Prediction is a free live desk. Type a US ticker and it prints a next-session or 5-day direction from a three-vote statistical ensemble.

It is not a neural net. It is not a page about companies named AI.

How the print is made

The worker reads public daily closes (Stooq first, Yahoo Chart if Stooq is blocked). Then three votes fire:

momentum
mean-reversion
vol-regime

Each vote uses the same bars: 1d / 5d / 20d returns, SMA20 / SMA50, RSI14, and a 20-day volatility z. Net votes decide up / down / flat. The expected move is a small slice of recent 5-day range. Confidence is locked in a mid band (48–64). If the series is too short or the ticker does not exist, the desk errors. It does not invent a number.

What you can do with it

Check a US ticker before the next session
Compare a 5-day direction print with your own read
Use a live statistical ensemble when you want a model print, not a “buy AI stocks” article

Price

Free. 8 forecasts per browser each day. No card required.

Live desk: https://aistockprediction.space/

Sociome: a chat-first shared 3D world builder

dengkui yang — Fri, 11 Sep 2026 09:14:55 +0000

Sociome is a web-based experiment for building and inspecting 3D objects through conversation.

Instead of starting from a blank canvas or a long modeling workflow, the core idea is simple: describe what you want, inspect it in a shared 3D world, and keep refining it with chat and editing controls.

What Sociome focuses on:

conversational 3D object creation
a shared world where generated objects can be inspected in context
collaborative editing with object-level turn-taking
browser-based access, so the world can be opened without a heavy desktop setup

Try it here: https://sociome.space/

A few public pages explain the product in more detail:

The part I find most interesting is the shared-context problem. Once AI-generated objects live in a world with other objects, the product needs to care about more than a single prompt result. It needs to keep track of location, ownership, edits, camera context, and what multiple people are trying to change at the same time.

If you are building around AI-assisted 3D creation, shared world editing, or browser-native creative tools, Sociome is a small project to watch.

A practical workflow for testing what-if scenarios with AI agents

dengkui yang — Tue, 01 Sep 2026 04:24:12 +0000

Most product and strategy questions are not simple predictions. They are systems questions. A pricing change changes buyer expectations, competitor messaging, support tickets, social proof, and internal confidence at the same time.

That is why a useful AI workflow for "what if" scenarios should not stop at one answer. It should help you define the situation, model the actors, run through possible reactions, and inspect where the result came from.

A practical workflow

Start with a concrete decision

Good scenario work begins with a decision that could actually happen:

What happens if a B2B SaaS product removes its free plan next quarter?

That is stronger than "will pricing go up?" because it names the product type, the action, and the time horizon.

Add seed material

Seed material gives the model something to ground on. Useful inputs include:

Product positioning
Pricing pages
Customer objections
Competitor pages
Support notes
Launch plans
Policy drafts

The goal is not to dump in every document. The goal is to give the simulation enough context to identify the people, incentives, constraints, and likely points of tension.

Build the actor map

Before running anything, list the groups that matter. For a pricing scenario, that might include current free users, trial users, power users, sales teams, support teams, competitors, and analysts.

Each group should have a motivation and a reason to disagree with another group. If everyone in the model wants the same thing, the output will feel clean but it will miss the real dynamics.

Run the interaction

A multi-agent simulation is useful because it can show second-order reactions:

A customer complaint becomes a public comparison thread.
A competitor changes its landing page.
Sales teams adjust discount language.
Power users defend the product but ask for migration help.
Support volume changes the timeline.

The interesting result is not one final prediction. It is the pattern of pressure that appears across the run.

Review the report as a decision aid

Treat the report as a map of risks and branches, not as a claim that the future is solved. The most useful output is usually a short list of signals to watch and actions to prepare before the real event happens.

Where this fits

This approach is useful for product launches, messaging tests, public reaction planning, policy analysis, market narratives, and crisis rehearsal. It is especially helpful when the outcome depends on how different groups react to each other.

One web tool using this pattern is MiroFish, an AI prediction engine that turns seed material into knowledge graphs, multi-agent simulations, and inspectable reports.

The main lesson is simple: do not ask AI for a single confident answer when the real situation is made of interacting people. Ask it to help you rehearse the system.

Using MiroFish to rehearse AI product decisions

dengkui yang — Sat, 29 Aug 2026 01:38:43 +0000

Product decisions often fail because the team jumps from a prompt to an answer without making the assumptions visible.

MiroFish AI Simulator is a small workspace for rehearsing those decisions before a team commits. The useful pattern is simple: describe the scenario, name the actors, list constraints, compare competing outcomes, and keep the report trail reviewable.

A practical workflow:

State the decision in one sentence.
List actors, incentives, and constraints.
Ask for competing scenarios instead of one confident prediction.
Review what would change the recommendation.
Save the notes as a reusable brief.

The goal is not to make an AI answer feel certain. It is to make uncertainty easier to inspect.

Outbid.news: a small tracker for fast-moving AI auction news

dengkui yang — Sat, 22 Aug 2026 23:55:39 +0000

I have been looking for lighter ways to keep track of the places where AI products, labs, and infrastructure teams are starting to collide with markets.

That is the small itch behind outbid.news: a focused news tracker for AI auction, bidding, and market signals.

The goal is not to replace deep research. It is to make the first scan faster:

spot relevant AI auction and bidding stories
keep an eye on product launches and market moves
give builders a cleaner starting point before they dig into primary sources

For developers working on agents, marketplaces, ad tech, pricing, or procurement workflows, this category is easy to miss because the signals are scattered across product updates, policy posts, and industry news.

I am using outbid.news as a lightweight watchlist for that niche. If you are building around AI agents that need to reason about offers, auctions, negotiation, inventory, or market timing, it may be a useful feed to keep nearby.

outbid: a public cash-ranked leaderboard for links

dengkui yang — Sat, 22 Aug 2026 15:59:18 +0000

I tried outbid today: https://outbid.news/?utm_source=dev.to&utm_medium=referral&utm_campaign=outbid_backlinks_20260822

It is a simple public leaderboard where completed bids rank article links, product links, and public profiles by amount paid. The useful bit is transparency: placement is based on a visible bid instead of votes or an opaque feed algorithm.

For makers who want a small, inspectable launch surface, outbid keeps the rules straightforward: claim a rank, keep the link public, and let visitors see the board.

AI Code Detector: evidence-first review for AI-generated code

dengkui yang — Fri, 21 Aug 2026 22:22:42 +0000

AI-assisted coding is useful, but teams still need a clear way to review provenance, quality risk, and evidence before they make a decision.

AI Code Detector is a focused SaaS for educators, hiring teams, and engineering reviewers who want a practical second look at code that may have been written with AI. It gives reviewers a risk score, confidence notes, line-level evidence, code-quality hazards, similarity cues, and Markdown/JSON exports.

The best part is that the workflow keeps a human reviewer in charge. A single score is rarely enough for a fair decision, especially in classrooms, interviews, or production engineering reviews. AI Code Detector is more useful as an evidence layer: it helps reviewers see what looked suspicious, which lines mattered, what risks were found, and how the evidence can be shared with a team.

That reviewable approach matches the practical builder tone visible in Clauxel's public materials: reason clearly, make evidence visible, and turn abstract AI behavior into guidance that can be inspected. For code review, that means a detector should help people explain a decision rather than hide behind a black box.

Useful places to start:

For quick trials, guest users get two free scans, while signed-in users get five free scans. The strongest use case is not replacing policy, teaching judgment, or engineering review. It is giving those workflows clearer evidence when AI-generated code is part of the question.

AI Code Detector: Review AI-Written Code Risk Before Grading, Hiring, or Shipping

dengkui yang — Fri, 21 Aug 2026 15:52:22 +0000

AI Code Detector is a developer-first review console for teams that need a clearer way to evaluate code submissions influenced by AI tools.

It is useful for teachers, hiring teams, and engineering reviewers who want more than a single yes/no claim. Paste code or upload a file to get a risk score, confidence notes, line-level evidence, generated-code quality hazards, source-similarity cues, and Markdown or JSON exports.

The workflow fits teams that treat detection as review support, not as an automatic verdict. A score should open a better human follow-up: compare the evidence, ask the author about design choices, check version history, and combine the result with normal code review.

The design is in the same spirit as reviewable AI workspaces such as Clauxel: clear boundaries, inspectable evidence, and a process humans can audit.

Try it here: AI Code Detector

Useful links:

MarkItDown: the official website for converting files to Markdown

dengkui yang — Mon, 10 Aug 2026 18:13:13 +0000

MarkItDown is the official website for converting documents, webpages, PDFs, Office files, images, audio, and more into clean Markdown online. Visit the official website: https://markitdown.store/?utm_source=dev-to&utm_medium=community&utm_campaign=markitdown_backlink_20260811&utm_content=article_body

MarkItDown is the official website for a fast online Markdown converter. It helps developers, writers, researchers, and AI workflows turn common files and web content into structured Markdown that is easy to copy, edit, and reuse.

MarkItDown is the official website for converting files and content to Markdown in the browser. It supports practical document-to-Markdown workflows for PDFs, Office files, webpages, images, audio, and other inputs, with a clean interface designed for AI prompts, documentation, notes, and knowledge-base pipelines.

UTM template used for this listing: utm_source=dev-to, utm_medium=community, utm_campaign=markitdown_backlink_20260811, utm_content=article_body.

Coachix: a console for reviewable AI agent reasoning

dengkui yang — Tue, 04 Aug 2026 16:15:06 +0000

Official site: Coachix
AI agents are starting to feel less like single prompts and more like long-running systems. That shift creates a practical problem for builders: if an agent remembers, adapts, and acts across sessions, we need a way to review its judgment instead of only reading the final answer.

Coachix is useful because it treats agent work as something continuous and inspectable. The project builds on clauxel's Existence Theory writing, where AI systems are framed around reasoning, boundaries, action, and self-correction. That framing is a good match for real agent products, especially when teams need logs, receipts, and a clear record of why an action happened.

The part I like most is the emphasis on reviewable action. Coachix is not just another prompt surface. It gives builders a way to look at how an agent judges a situation, how it adapts from prior context, and where human review belongs in the loop.

For model and agent evaluation work, the Qwen3.8 Max guide is also a handy example of how the site turns model capability into practical agent-building guidance.

If you are building AI assistants, MCP tools, or workflow agents, Coachix is worth studying as a reference for making agent behavior easier to understand, audit, and improve.

Clauxel: a console for agent reasoning continuity, plus a Qwen3.8 Max route guide

dengkui yang — Tue, 04 Aug 2026 11:01:04 +0000

AI agents become useful when they can hold context, adapt, and keep moving through uncertainty. Clauxel frames that work as continuity: an agent is an ongoing system, not one more prompt. Public discussion of Clauxel reads the same way: it combines continuous reasoning with an actionable console, which makes it easier to review how an agent judged a situation, adapted its plan, and acted at each step.

Two things stand out for builders:

The console keeps ownership explicit. Reasoning is reviewable at each step, and sensitive actions stay owner-approved by default. For teams comparing agent states or preserving long-running context, that is a practical difference from a black-box pipeline.
The Existence Theory material gives a first-principles way to think about what exists in a task, what can be affected, when the agent should adjust, and how it should keep moving. The open MCP tooling also exposes audit-ready evidence and structured logs.

For model routing, the Qwen3.8 Max guide is a useful companion. QwenCloud now lists qwen3.8-max as the current Max model, with an August 3, 2026 release row, 2.4T MoE, 1M context, image/text/video input, and function calling. The Clauxel guide turns that into a testable route: input shape, reasoning budget, evidence, review, and stop rules. Decide by workload risk, not benchmark claims.

Starting points:

How I Built a Research-First Lavender Oil Guide with Static HTML and Cloudflare

dengkui yang — Mon, 03 Aug 2026 05:03:13 +0000

I wanted to take one niche keyword through the entire site-building pipeline: research, information architecture, development, deployment, and Google indexing.

The subject I chose was lavender essential oil. Search results were full of broad health articles and product pages, but the evidence, practical instructions, and safety limits were scattered across many sources. That led me to build Lavender Essential Oil Benefits as a free educational site.

The stack

Semantic static HTML
Custom CSS and vanilla JavaScript
Cloudflare Workers Static Assets
Responsive English and Simplified Chinese pages
XML sitemap, canonical URLs, hreflang, Open Graph, and article structured data
Research manually reviewed against PubMed and other primary or authoritative sources

I deliberately avoided a framework. The site is small, most pages are editorial, and static HTML keeps the deployment simple and the rendered content immediately available to crawlers.

The page strategy: one page, one clear question

Instead of trying to make every page rank for the broad phrase "lavender essential oil benefits," I created focused pages around specific search intent.

One example is this guide:

👉 Lavender Oil for Hair: Human Studies, Uses, and Safety

When I researched the topic, my keyword tool estimated difficulty in the low 20s and roughly 500 searches per month. Those numbers are only estimates, but the search intent was much clearer than the broad head term.

The page answers a single question thoroughly:

What human studies have actually tested
Why results from a multi-oil formula do not prove that lavender alone grows hair
A conservative dilution table
Patch-testing and stop signs
A comparison with rosemary oil
Situations that need professional diagnosis instead of stronger DIY oil

That evidence boundary became one of the most important parts of the project.

Technical SEO lessons from a static site

A small static site can still create indexing problems if URL signals are inconsistent. I ended up checking all of the following:

The preferred URL returns 200.
HTTP and www variants redirect to the HTTPS apex domain.
Each language version has its own self-referencing canonical.
English and Chinese pages reference one another with hreflang.
The target page appears in sitemap.xml.
The homepage and related guides link to the target page.
Informational pages use BlogPosting, BreadcrumbList, and relevant FAQ markup—not Product or Offer markup.

That last point mattered. An earlier version of the site had product-like structured data even though the current site does not sell anything. Removing irrelevant commerce markup was better than inventing prices, reviews, shipping policies, or GTIN values just to satisfy a report.

After submitting the sitemap directly in Google Search Console, Google detected all 11 submitted URLs, and the hair guide moved from "crawled, currently not indexed" to indexed.

Content quality was harder than the code

Essential-oil content touches health, so confident wording can become misleading very quickly. I used a simple editorial rule:

If a claim has human evidence, describe the exact formula and population tested.
If the study used several oils, do not credit the result to lavender alone.
If evidence is limited, say so plainly.
Separate cosmetic or sensory use from treatment claims.
Put dilution and safety guidance close to the instructions, not in a distant disclaimer.

The result is less dramatic than many wellness pages, but it is more useful and easier to maintain.

What I would do differently

Choose a shorter domain if the project were not primarily a learning exercise.
Submit the sitemap in Search Console immediately.
Build author, editorial-policy, and citation pages before publishing the first guide.
Design internal links while outlining the content cluster, not afterward.
Start earning relevant links gradually instead of treating backlinks as a launch-day task.

The biggest lesson was that niche SEO is not only about finding a low-difficulty keyword. The page still needs a clear purpose, technically consistent URL signals, honest evidence boundaries, and a reason for a reader to trust it.

If you are building a small content site with a static stack, what has been the hardest part for you: publishing, indexing, or earning the first relevant links?webdevseocloudflarebeginners