标签：#datascience

Best-of-K eval selection bias: pick the highest-scoring config from K candidates on one eval set and that observed score is biased up. It reports the expected maximum of K noisy estimates, which beats the field mean whenever K exceeds one. The bias appears even when all K configs are truly equal, grows with K, and shrinks with n. Here is the version that bites you. Your bake-off ran a batch of prompts against one eval set, the top one came out ahead, and you shipped it. In production it does worse. That drop reads like bad luck, or drift, or a bad week. It is none of those. It is a number you could have computed before you shipped, and it gets larger the more candidates you tried. I ran a small script to make the gap concrete. Eight configs, one hundred eval items, and here is the catch: I made all eight configs truly identical , every one a fair coin at 50%. There is no real best. Nothing to tune. Then I let selection pick a winner anyway: config 0: 47/100 = 47.0% config 1: 50/100 = 50.0% config 2: 52/100 = 52.0% config 3: 52/100 = 52.0% config 4: 50/100 = 50.0% config 5: 50/100 = 50.0% config 6: 54/100 = 54.0% <- selected winner (argmax) config 7: 44/100 = 44.0% config 6: 54.0% (k=54 n=100 SE=4.98) config 2: 52.0% (k=52 n=100 SE=5.00) RANK: INDISTINGUISHABLE - gap 2.00 pp against 7.06 pooled SE = 0.28 SE < 2.0. Ranking "config 6" above "config 2" is NOT allowed. Held-out the winner on a fresh 100 items: 48/100 = 48.0%. Config 6 wins the bake-off at 54.0%. Then I asked the same eval-guard I use in the McNemar and rule-of-three pieces to rank config 6 against the runner-up. It refused: the gap is 0.28 SE, far under the two-SE bar, so INDISTINGUISHABLE . The ranker would not call config 6 the best. Selection did. And on a fresh held-out set the 54.0% falls back to 48.0%, toward the true 50% it was always going to be. TL;DR Picking the best of K configs by observed pass rate reports the expected maximum of K noisy estimates. The max of K exceeds the field mean, strict

The rule of three for evals says zero failures in N runs is a count, not a rate. With 0 failures in N independent runs, the exact 95% upper bound on the true failure rate is 1 - 0.05^(1/N) , which 3/N approximates. After 100 clean runs you still cannot rule out a 2.95% rate, about 1 in 34. Here is the reading that bites you. Your eval harness runs the agent 100 times, prints "0 failures," and the tile goes green. Someone screenshots it into the launch thread. The unspoken translation is "the failure rate is zero." It is not what the data says. I wrote a small script to make the gap concrete, so I ran a real gate over 200 deterministic agent runs first, counted honestly, and got the dashboard everyone trusts: gate: spend<=budget over 200 deterministic agent runs observed failures: 0 (distinct scenarios: 200) naive point rate : 0.00% binomial SE: 0.00 pp naive 95% CI : [0.00%, 0.00%] <- zero width: false certainty Look at the standard error. For a zero count the binomial SE is sqrt(0*1/200) , which is exactly 0, so the naive 95% interval collapses to [0.00%, 0.00%] . A zero-width confidence interval. The math is telling you it is completely certain, from 200 samples, that the true rate is precisely zero. That is obviously wrong, and it is the exact shape of every "all green" board I have ever trusted too much. TL;DR "0 failures in N runs" is an observed count, not a rate. The naive binomial SE of a zero count is 0, which is why a green board looks like proof and isn't. The honest number is the one-sided upper bound. With 0 failures in N runs, the 95% upper confidence limit on the true failure rate is 1 - 0.05^(1/N) . The rule of three, 3/N , approximates it and rounds the risk slightly up. At N=30 the bound is 9.50% (about 1 in 11). At N=100 it is 2.95% (1 in 34). At N=1000 it is 0.30% (1 in 334). Zero failures in 30 runs is compatible with a 1-in-11 true failure rate. To rule out a 0.1% rate at 95% you need about 2995 clean runs, not 50. "We ran it fifty times" and "

If you have written any optimization code, you have met linear programming even if nobody called it that. "Maximize output without blowing the resource budget" is an LP problem, and the classic algorithm that solves it is the simplex method. It is worth understanding not because you will hand-code it (you'll usually call a solver), but because knowing how it moves makes you far better at modeling problems for it. Here is the algorithm stripped down to its logic. The problem shape Every LP problem has three parts: an objective function to maximize or minimize, e.g. Z = 5x1 + 4x2 a set of linear constraints, e.g. 6x1 + 4x2 <= 24, x1 + 2x2 <= 6 non-negativity: all variables >= 0 Geometrically, the constraints carve out a feasible region (a polytope). The optimum always sits at a corner of that region. The simplex method is just a smart way of hopping from corner to corner, uphill, until there is no higher corner to move to. The algorithm as pseudocode build initial tableau (add a slack variable per <= constraint) loop: compute Cj - Zj for each column if all (Cj - Zj) <= 0: break # optimal reached pivot_col = column with most positive Cj - Zj # entering variable ratios = RHS / pivot_col entries (only positive entries) pivot_row = row with smallest non-negative ratio # leaving variable pivot(pivot_row, pivot_col) # elementary row operations return solution from final tableau That's it. Four moves per iteration: score the columns, pick the entering variable, run the ratio test for the leaving variable, pivot. Repeat until the optimality condition holds. A quick worked run Take Maximize Z = 5x1 + 4x2 subject to 6x1 + 4x2 <= 24 and x1 + 2x2 <= 6. Add slack variables s1, s2, build the tableau, and iterate. The optimum lands at x1 = 3, x2 = 1.5, Z = 21. Two pivots and you're done. Simple on paper until the numbers get ugly. Where humans (and debugging) actually break The algorithm is clean. The arithmetic is not. A single wrong entry in one pivot silently corrupts every table

Why Your WeChat Mini-Program Work Is Actually Relevant US data engineering interviews prioritize hands-on experience with data pipelines, SQL, and scalable storage. WeChat mini-programs, though often seen as front-end tools, generate and process substantial backend data. If you designed the logic behind user actions, session management, or real-time updates, you already have transferable skills. The key is to stop describing the mini-program as a "feature" and start describing it as a "data system." Recruiters don't care about WeChat's UI components; they care about how you moved, transformed, and stored data. The Core Translation: From Mini-Program to Data Engineering When rewriting your resume, map each mini-program task to a standard data engineering function: User interaction logic → Event-driven data pipeline (e.g., Kafka, AWS Kinesis, or custom queues) WeChat cloud functions → Serverless compute (e.g., AWS Lambda, Azure Functions) Database reads/writes via WeChat API → NoSQL or relational database operations (e.g., MongoDB, MySQL, DynamoDB) Session tracking or analytics → ETL pipeline extracting user behavior data into a warehouse Do not use WeChat-specific terms like "wx.request" or "Mini Program SDK" without explaining the data they handled. Replace them with equivalent English terms. A Quick Side-by-Side Translation Table WeChat Term US Data Engineering Equivalent WeChat Cloud Database Managed NoSQL database (like MongoDB Atlas) Mini Program Backend with Cloud Functions Serverless data processing (AWS Lambda) User login & session management Authentication pipeline with token storage Real-time message push Event-driven notification system Reshaping Your Bullet Points: Before and After Generic bullet points kill your chances. Here is a concrete rewrite that turns a typical WeChat mini-program description into a data engineering achievement. Before (Chinese-English hybrid, vague): "Developed a WeChat mini-program for e-commerce with 50k users. Used wx.request

I watched my sister paste a CSV into on of our LLM overlords and ask a question then reword it and get a different number. That's the problem with letting an LLM both read your question and trust it's math. ADA splits the two jobs. Drop in a CSV or Excel file, get a dashboard, anomaly detection, a forecast, and plain-English answers, all calculated locally using pandas (locally in the sense of never leaving the server and going to the AI apis.). The LLM layer that you can use (it's optional) but that only sees column names and never your rows, and the whole thing works with zero API key if need be (not the LLM part of course). MIT licensed, solo-built, open to contributors (in fact would need help to make it move out of LLMs and have a few open issues). I wish to make it as LLM free as possible as I want to make it deterministic. Live demo: automated-data-analyst.streamlit.app Repo: https://github.com/saineshnakra/automated-data-analyst

Heads-up: this post references a tool I built. It's a genuinely useful walkthrough either way — the technique applies to any Airbnb scraping project. If you've ever tried to scrape Airbnb, you already know the two walls you hit: the pages are rendered by JavaScript, and Airbnb aggressively blocks datacenter IPs. Below is the reliable way to get clean Airbnb data in 2026 — listing prices, ratings, coordinates, and discounts — without running a headless browser or babysitting proxies. The key insight: Airbnb ships its data in the HTML You don't need to render the page. Every Airbnb search response embeds the full result set as JSON inside a <script id="data-deferred-state-0"> tag. Parse that and you get structured data straight away — no DOM scraping, no selectors that break on the next redesign. The path to the results is: niobeClientData[*][1].data.presentation.staysSearch.results ├── searchResults[] // ~18 listings per page └── paginationInfo.pageCursors[] // all page cursors, upfront Each listing carries a base64-encoded ID in demandStayListing.id (decode it, take the segment after the last colon, and you have the numeric listing ID for airbnb.com/rooms/<id> ), a price line with discounts, avgRatingLocalized ("4.95 (123)"), and GPS coordinates. The two gotchas Datacenter IPs get blocked. You need residential proxies. If a response comes back without the data-deferred-state marker, you've been served a bot check — rotate to a fresh IP and retry. ~270 result cap per search. Airbnb won't paginate past ~15 pages. To cover a whole market, split into narrower searches (by price band or neighborhood) and dedupe by listing ID. The no-code way If you'd rather not maintain proxy pools and parsers, I published an Airbnb Scraper on Apify that does exactly the above. Paste a location or a full Airbnb search URL (every filter is honored), and get flat JSON/CSV back. curl -X POST "https://api.apify.com/v2/acts/ethanteague~airbnb-scraper/run-sync-get-dataset-items?token=YOUR_TOKE

Every GEO ("generative engine optimization") tool, including ours until recently, sells some version of the same pitch: fix your robots.txt, add Schema.org markup, write FAQ schema, and AI engines will cite you more. We build one of these tools — Causabi scans sites for AI-crawler readiness and generates fix files (robots.txt, llms.txt, JSON-LD, FAQ blocks). As part of validating our own scoring weights, we ran the numbers on whether the score actually predicts getting cited. Short version: it mostly doesn't, once brand prominence is in the picture. What we measured We scored 44 domains on a 6-category on-site readiness algorithm: robots.txt (AI bots allowed or blocked) Schema.org (Organization/LocalBusiness JSON-LD completeness) FAQ schema (FAQPage markup, 3+ entries) content depth/structure brand/NAP signals freshness (dateModified, recency) Then we checked how often each domain actually got cited by an AI engine (Claude, via its web-search tool, one measurement window, a fixed prompt set per domain). What we found On-site score vs. citation rate: Pearson r ≈ -0.08, Spearman ρ ≈ -0.03. Functionally no correlation — if anything, a very slight negative one, which is more likely noise than a real inverse relationship at this sample size. 86% of the 44 domains got zero citations in the window, independent of their score. The domains that did get cited clustered almost entirely by brand prominence — well-known domains got cited at a noticeably higher rate (~0.16 of prompts) than everyone else (~0 for the rest of the sample), regardless of how well-optimized their markup was. Why I'm not overselling this n=44 is small. This is an internal validation exercise for our own product, not a peer-reviewed study, and I don't want it read as one. Specific caveats: Single engine (Claude) this round. Citation behavior differs meaningfully across ChatGPT, Gemini, Grok, and Perplexity — we haven't run the same check across all four yet. One time window, no longitudinal before/after.

Risk stratification sounds like a data-science buzzword until you have to build the thing. For a Medicare Advantage plan, it's a concrete pipeline: take a population of members, score each one's clinical and financial risk, and rank them so care management and documentation teams know who to touch first. Here's how I'd architect it. The core idea Population health risk stratification = scoring + segmentation. You compute a per-member risk signal, then bucket members into tiers (e.g., rising-risk, high-risk, catastrophic) so finite resources go where they move outcomes and revenue most. The mistake teams make is treating it as a single ML model. In practice you want a layered signal: a stable, explainable base (RAF + chronic conditions) plus optional predictive overlays. Explainability matters because care managers won't act on a black-box score, and auditors won't accept one. Step 1: Build the member feature record { "member_id" : "SYNTH-77310" , "age" : 73 , "hccs" : [ "HCC37_1" , "HCC85" , "HCC18" ], "raf" : 1.842 , "gaps" : [ "a1c_overdue" , "no_pcp_visit_180d" ], "utilization" : { "ed_visits_12m" : 3 , "inpatient_12m" : 1 } } The RAF here is your defensible, model-grounded risk anchor under CMS-HCC V28. Everything else is supplemental signal. Step 2: Score and tier def risk_tier ( member ): base = member [ " raf " ] util = 0.15 * member [ " utilization " ][ " ed_visits_12m " ] \ + 0.30 * member [ " utilization " ][ " inpatient_12m " ] score = base + util if score >= 3.0 : return " catastrophic " if score >= 1.8 : return " high " if score >= 1.0 : return " rising " return " stable " Keep the weights transparent and tunable. The point isn't a perfect model; it's a defensible, reproducible ranking your operational teams trust. Step 3: Make "rising-risk" actionable The tier that quietly drives the most ROI is rising-risk — members trending toward high cost who still have open documentation and care gaps. Surface their specific gaps (overdue labs, undocumented chroni

Design + Product Thinking: NYC’s Path to Reliable AI AI delivers value when it’s useful, trusted, and operational. For city services that affect millions, those qualities don’t happen by accident — they come from applying design thinking (who the service is for, how it’s used) together with product thinking (what outcome we’re trying to achieve and how we operate over time). This article explains why hiring designers and product managers matters for NYC’s digital and AI initiatives, summarizes the city’s PIT Crew program, and outlines how Flamelit applies outcome-focused delivery in the public sector. Why design and product roles matter Designers and product managers have distinct but complementary responsibilities that reduce common AI delivery failures: Designers (Design Thinking): center human needs, prototype user flows, and validate that interfaces and decision workflows are understandable and accessible. They surface usability and trust issues early, preventing technically accurate models from becoming unusable in practice. Product managers (Product Thinking): define the measurable outcomes, prioritize use cases, align stakeholders, and manage the lifecycle from discovery to ongoing operations. They ensure work is evaluated against mission impact, not just technical metrics. Together they prevent common failures: building technically impressive models that nobody trusts, deploying brittle systems without human review, or shipping features with unclear ownership that decay in production. PIT Crew and NYC hiring context NYC’s PIT Crew program is a city initiative designed to attract and staff product, engineering, and design talent for public service projects. It’s a practical recognition that public-sector digital transformation needs people skilled in user research, product management, and delivery. Read more about the PIT Crew and how it works here: https://www.nyc.gov/content/pitcrew/pages/ (open in a new tab). Hiring programs like PIT Crew help create the c

If you’ve ever struggled with CALCULATE() or wondered why SUMX() behaves differently from SUM() , this guide is for you. DAX (Data Analysis Expressions) is the language that powers Power BI , Analysis Services , and Power Pivot — enabling dynamic calculations, filtering, and time intelligence. Below is a categorized cheat sheet of essential DAX functions , plus examples showing how to use each in real-world Power BI scenarios. Filtering & Context These functions control how filters are applied and evaluated in your calculations. Function Example Description CALCULATE() CALCULATE(SUM(Sales[Amount]), Region[Name] = "Nairobi") Changes filter context to calculate total sales for Nairobi. FILTER() FILTER(Sales, Sales[Amount] > 10000) Returns a table filtered by condition. ALL() CALCULATE(SUM(Sales[Amount]), ALL(Region)) Ignores filters on Region. REMOVEFILTERS() CALCULATE(SUM(Sales[Amount]), REMOVEFILTERS(Region)) Removes filters from Region. VALUES() VALUES(Customer[City]) Returns unique list of cities. SELECTEDVALUE() SELECTEDVALUE(Product[Category], "All") Returns selected category or “All” if none. TREATAS() TREATAS(VALUES(Temp[City]), Customer[City]) Applies one table’s values as filters on another. KEEPFILTERS() CALCULATE(SUM(Sales[Amount]), KEEPFILTERS(Product[Category] = "Electronics")) Keeps existing filters and adds new ones. ALLSELECTED() CALCULATE(SUM(Sales[Amount]), ALLSELECTED(Region)) Respects user selections in visuals. ALLEXCEPT() CALCULATE(SUM(Sales[Amount]), ALLEXCEPT(Sales, Sales[Year])) Removes all filters except Year. Aggregation Summarize or aggregate data across rows or columns. Function Example Description SUM() SUM(Sales[Amount]) Adds all sales amounts. AVERAGE() AVERAGE(Sales[Amount]) Calculates mean value. COUNT() COUNT(Customer[ID]) Counts non-blank entries. COUNTROWS() COUNTROWS(Sales) Counts rows in a table. DISTINCTCOUNT() DISTINCTCOUNT(Customer[ID]) Counts unique customers. MIN() MIN(Sales[Amount]) Finds smallest sale. MAX() MAX(Sales[Amo

series: Prediksi Survival Klub Debutan Kenapa Project Ini? Setiap musim, klub yang promosi ke liga top (Premier League, La Liga, dst.) menghadapi risiko besar: sekitar 2 dari 3 klub yang naik biasanya kembali terdegradasi di musim pertama mereka. Saya penasaran — bisakah performa di beberapa laga awal musim memberi sinyal dini soal peluang klub tersebut bertahan? Ini jadi project portofolio pertama saya sebagai data scientist yang baru mulai (0-1 tahun pengalaman). Saya sengaja pilih topik yang saya suka (sepak bola) supaya prosesnya tetap enjoyable, bukan cuma "tutorial project" generik. Rencana Project Pertanyaan utama: Berdasarkan performa 8 laga pertama musim debut, seberapa besar peluang klub promosi bertahan hingga musim berikutnya (tidak degradasi)? Data yang dipakai: football-data.co.uk — data hasil pertandingan tiap musim sejak 1993/1994 Wikipedia (halaman musim liga) — daftar klub promosi & klasemen akhir musim Tech stack: pandas , requests untuk data collection scikit-learn untuk modeling (mulai dari Logistic Regression sebagai baseline) imbalanced-learn untuk handle class imbalance Streamlit + Plotly untuk dashboard interaktif Deploy ke Streamlit Community Cloud Timeline (Build in Public) Saya bikin timeline ini publik supaya ada tekanan yang sehat untuk benar-benar menyelesaikannya, bukan cuma jadi ide yang menguap: Checkpoint Target Tanggal Yang Harus Selesai Part 1 (post ini) 11 Juli 2026 Kickoff, rencana, environment siap Part 2 15 Juli 2026 Dataset jadi, push ke GitHub Part 3 17 Juli 2026 EDA selesai, insight awal Part 4 24 Juli 2026 Model final dipilih + evaluasi Part 5 31 Juli 2026 Dashboard live di Streamlit Cloud Part 6 (final) 8 Agustus 2026 Project selesai, recap lengkap Tantangan yang Sudah Saya Antisipasi Data leakage — fitur harus dihitung dari laga awal musim saja, bukan seluruh musim, biar model beneran memprediksi bukan "menyontek" hasil akhir Dataset kecil — kemungkinan hanya ~60-100 sampel klub, jadi saya mulai dari model sederhana (Lo

Adapted from an appendix of my MS thesis. Markov Chain Monte Carlo Almost as soon as computers were invented, they were used for simulation. Markov chain Monte Carlo (MCMC) was invested as Los Alamos, Metropolis et al (1953) simulated a liquid in equilibrium with its gas phase. Their tour de force was the realization that they did not need to simulate the exact dynamics, they only needed to simulate some Markov chain with the same equilibrium distribution. The Metropolis algorithm was widely used by chemists and physicists, but was not widely known among statisticians until after 1990. Hastings (1970) generalized the Metropolis algorithm, and simulations following his scheme are said to use the Metropolis-Hastings (MH) algorithm [1]. A special case of the MH algorithm was introduced by Geman et al (1984) discussing optimization to find the posterior mode rather than simulation. Algorithms following their scheme are said to use the Gibbs sampler. It took some time for the spatial statistics community to understand that the Gibbs sampler simulated the posterior distribution, thus enabling full Bayesian inference of all kinds. Gelfand et al (1990) made the wider Bayesian community aware of the Gibbs sampler, and then it was rapidly realized that most Bayesian inference could be done using MCMC, whereas very little could be done without MCMC. Green (1995) generalized the MH algorithm as much as it could be generalized [1]. Theoretical Foundations A sequence X 1 ​ , X 2 ​ , … of random elements of some set is a Markov chain if the conditional distribution of X n + 1 ​ given X 1 ​ , … , X n ​ depends on X n ​ only. The set in which the X i ​ take values is called the state space of the Markov chain. A Markov chain has stationary transition probabilities if the conditional distribution of X n + 1 ​ given X n ​ does not depend on n . This is the main kind of Markov chain of interest in MCMC. The joint distribution of a Markov chain is determined by the following [1]. The ma

I run a small AI trends site, and three weeks ago I started doing something simple: every day, snapshot the star count of every repo that crosses my GitHub trending scan for AI. No judgment, no curation, just append-only rows in a database. 611 repos and 2,671 data points later (June 19 to July 10), the picture of what's actually growing looks pretty different from what my feeds told me was hot. Here's what the data says. Before publishing this I re-checked every number below against GitHub's live API. Star counts drift by the hour, so treat them as of July 10. The top 10 risers, by raw stars gained Repo Gained Window From → To calesthio/OpenMontage +30,253 21 days 5,899 → 36,152 DeusData/codebase-memory-mcp +20,483 19 days 7,516 → 27,999 mattpocock/skills +19,053 15 days 137,485 → 156,538 obra/superpowers +16,887 20 days 232,908 → 249,795 NousResearch/hermes-agent +14,896 21 days 197,297 → 212,193 Panniantong/Agent-Reach +14,334 14 days 34,780 → 49,114 usestrix/strix +13,243 12 days 26,363 → 39,606 addyosmani/agent-skills +12,685 21 days 63,156 → 75,841 asgeirtj/system_prompts_leaks +11,720 21 days 43,415 → 55,135 msitarzewski/agency-agents +11,055 10 days 118,241 → 129,296 Windows differ because I only hold snapshots for the days a repo appeared in my scan; each row states its own real window. Three things in this data genuinely surprised me. 1. "Skills" are eating agent frameworks Four of the top ten are not agent frameworks. They are collections of packaged expertise that plug into an existing agent: obra/superpowers (still compounding at roughly 840 stars a day on a 250k base), mattpocock/skills, addyosmani/agent-skills, msitarzewski/agency-agents. A year ago this table would have been full of new frameworks. Now the framework layer looks settled and the growth is in what you load INTO the agent. The moat moved from orchestration code to encoded judgment. 2. The sharpest climbs are applications, not infrastructure The steepest sustained climb from a newcomer in

A friend messaged me a photo of a sealed booster box last month with one question: "worth it?" He'd already decided, really. The chase card in that set was all over his feed, so the box felt like a good deal. I asked him to send me the pull rates instead of the hype, and we spent twenty minutes turning "worth it?" into something we could actually compute. That exercise is a small, self-contained data problem. It's also a good example of how a clean-looking model can hand you a confident number that doesn't survive contact with reality. If you like building little estimators, this one is worth doing carefully, because the interesting part isn't the formula. It's everything the formula assumes. The formula is the easy part Expected value of a box is a weighted sum. Each card you can pull has a probability and a market value, and you multiply the two across every slot the box gives you. That's it. Undergrad probability. Here's a stripped-down version for a hypothetical set. I'm using made-up numbers so nobody mistakes this for real pull data — the point is the shape of the computation, not the specific set. # One "hit slot" in a box: probabilities cover the full outcome space. # Values are illustrative market estimates in USD. hit_table = [ { " name " : " Alt-art chase " , " p " : 0.0125 , " value " : 180.00 }, { " name " : " Secret rare " , " p " : 0.030 , " value " : 45.00 }, { " name " : " Full-art rare " , " p " : 0.100 , " value " : 8.00 }, { " name " : " Standard hit " , " p " : 0.400 , " value " : 0.55 }, { " name " : " No notable hit " , " p " : 0.4575 , " value " : 0.06 }, ] assert abs ( sum ( row [ " p " ] for row in hit_table ) - 1.0 ) < 1e-9 ev_per_slot = sum ( row [ " p " ] * row [ " value " ] for row in hit_table ) hit_slots_per_box = 36 # e.g. one meaningful slot per pack ev_box = ev_per_slot * hit_slots_per_box print ( f " EV per slot: $ { ev_per_slot : . 2 f } " ) # $4.65 print ( f " EV per box: $ { ev_box : . 2 f } " ) # $167.31 The box costs $150 sea

Every big social network locks audience data behind auth walls and anti-bot systems. Bluesky went the other way. The AT Protocol is open by design, so public profile data (bios, follower counts, full follower and following lists) is queryable through a documented API without logging in. The whole surface is basically two endpoints: GET https://api.bsky.app/xrpc/app.bsky.actor.getProfile?actor=HANDLE GET https://api.bsky.app/xrpc/app.bsky.graph.getFollowers?actor=HANDLE&limit=100 There's also getProfiles for batching 25 handles per call. Follower lists paginate with a normal cursor , which still works on the graph endpoints. Search is a different story, cursor pagination 403s there now, but that's a topic for another post. For one-off lookups, curl is honestly all you need. Where it gets tedious Bulk. Thousands of profiles, follower exports that run into six figures, weekly snapshots for tracking. Pagination, rate-limit backoff, and stitching the pages together is boring code that has to run reliably. I packaged that part as an Apify actor: Bluesky Profile Scraper . Paste handles or profile URLs, optionally turn on follower/following export, and you get JSON or CSV back with a sourceProfile field linking each follower record to the profile it belongs to. $2 per 1,000 records, runs on a schedule if you want snapshots over time. What people use this for Vetting an influencer's real audience before paying them. Exporting who follows a competitor and what their bios say. Charting follower growth from weekly runs. And enrichment: find who's talking about you with a mentions monitor , then profile those authors to see their actual reach. Bluesky is the only major network right now where any of this is straightforward and stable. Worth using while it lasts.

Introduction I used to think "connecting to a database" was one skill. Turns out it's two: connecting to a database chilling quietly on your own laptop, and connecting to one living in the cloud, behind a login, in this case, an SSL certificate that will not let you in until you treat it with respect. This week I did both. Same tool (Power BI), same dataset, two very different vibes. Grab a coffee, here's the full walkthrough local PostgreSQL first, then Aiven's cloud version, side by side, screenshots and all. Part 1: Local PostgreSQL → Power BI Step 1 : Create a schema Nothing fancy, just giving my table a home: CREATE SCHEMA powerbi ; Step 2 : Import the dataset Right-click the new schema → Import Data in DBeaver, point it at your CSV, and let the wizard do its thing. Step 3 : Check the table landed properly A quick peek at the columns to make sure nothing got mangled on the way in. Step 4 : Connect Power BI In Power BI Desktop: Get Data → Database → PostgreSQL database. In the Server field, type localhost (or 127.0.0.1 ) and your database name. localhost Choose Import , hit OK, and log in with your local username and password. Click Load . That's it. That's the whole local experience. Part 2: Aiven PostgreSQL (Cloud) → Power BI Now for the part that actually taught me something. Step 1 : Grab your connection details Everything you need lives on Aiven's Overview page: Host, Port, Database name, User, SSL mode. Your service URI will look something like this (don't worry, this isn't a real password, Aiven masks it in the console): postgres : // avnadmin : •••••••• @ pg - xxxxxxxx - yourproject . c . aivencloud . com : 22016 / defaultdb ? sslmode = require Step 2 : Import the dataset into Aiven Same DBeaver wizard as before, just pointed at the Aiven connection instead of local. CREATE SCHEMA powerbi ; Step 3 : Aiven's certificate. Download the CA cert from the Overview page: Now here's the part that actually tripped me up: Power BI's PostgreSQL connector doesn't ha

Why Your SOE Resume Needs a Structural Overhaul Chinese state-owned enterprises (SOEs) often have deep hierarchical structures and a culture of collective achievement. But Western tech companies want to see individual impact, autonomy, and data-driven results. Continuing to lead with your former employer's prestige or your rank (e.g., "Senior Engineer Grade 7") wastes valuable space. The solution: reshape every section to answer the question "What did you personally accomplish with data?" The Core Shift: From Hierarchy to Impact In a Chinese SOE resume, it's tempting to list departments you led or teams you oversaw. In a Western senior data scientist resume, focus on the problems you defined, the algorithms you deployed, and the revenue, cost savings, or user metrics that improved. For example, instead of "Led the data analytics team of 10 people," write "Designed and deployed a demand-forecasting model that reduced inventory costs by 15% (¥12M annually)." Three Resume Sections That Require Full Rewriting Professional Summary: From 'Accomplished Engineer' to 'Data Science Leader' Start with your total years of experience, your technical stack, and the types of business problems you solve. Example: "Senior Data Scientist with 10+ years applying machine learning to supply chain and logistics. Expertise in Python, TensorFlow, and Spark. Reduced operational costs by 15-30% through predictive models deployed at [SOE name]." Work Experience: From Role Descriptions to Metric-Driven Bullets For each role, list 3-5 bullets. Every bullet should have a verb, a task, a technology (if relevant), and a quantified result. Avoid vague phrases like "responsible for." Use specific numbers: "Improved forecast accuracy from 70% to 85% by building an ensemble of ARIMA and XGBoost models." Education & Certifications: Emphasize Transferable Skills Your Chinese degree is fine, but add relevant certifications (AWS, TensorFlow, Coursera) to show adaptability. Consider a "Technical Skills" se

1. The Problem It Solves Logistic Regression is used when the outcome is a category rather than a number . Most commonly, it's used for binary classification , where the answer is either Yes or No , True or False , or 1 or 0 . Typical business problems include: Will a customer churn? Is this transaction fraudulent? Will a customer click an ad? Will a loan default? Is an email spam? Will a machine fail in the next 24 hours? Unlike Linear Regression, we're not trying to predict a continuous value. Instead, we're predicting the probability that an event belongs to a particular class. For example: A customer may have an 82% probability of churning . The business can then decide whether that probability is high enough to trigger an intervention. 2. Core Intuition Imagine you're trying to predict whether a customer will cancel their subscription. Suppose the only feature you have is how many times they opened your app this month. If you use a straight line like Linear Regression, the predictions quickly become unrealistic. A very active customer might end up with a -20% chance of churn . A completely inactive customer could end up with 140% . Probabilities obviously can't work like that. To fix this, Logistic Regression takes the linear equation and passes it through a mathematical function called the Sigmoid Function . Instead of producing a straight line, it creates an S-shaped curve . No matter how large or small the input becomes, the output always stays between 0 and 1 . That makes it perfect for probability estimation. 3. The Mathematical Model The model first calculates a linear score. Instead of using that score directly, it passes it through the Sigmoid function. Where: z = linear score p̂ = predicted probability The final output is always between 0 and 1 . For example: 0.08 → Very unlikely 0.32 → Low risk 0.65 → Moderate risk 0.94 → Very high probability Businesses can then choose a decision threshold. For example: Probability ≥ 0.50 → Predict Churn Probability

I recently completed an exploratory data analysis project on the NHANES (National Health and Nutrition Examination Survey) dataset from Kaggle. It's a real-world health survey collected by the CDC covering body measurements, lifestyle habits, and demographic data from thousands of US adults. In this article I'll walk you through exactly what I did — from loading and cleaning the data all the way to running statistical tests — and share what I found along the way. The Dataset The dataset has 5,735 rows and 28 columns , but for this project I focused on 8 columns that were relevant to the questions I wanted to answer: Column Description smoking Has the person smoked at least 100 cigarettes? gender Male or Female age Age in years education Highest level of education weight Weight in kg height Height in cm bmi Body Mass Index Step 1 — Loading and Selecting Columns import pandas as pd import numpy as np import matplotlib.pyplot as plt import seaborn as sns db = pd . read_csv ( ' NHANES.csv ' ) data = db . loc [:, ( ' SEQN ' , ' SMQ020 ' , ' RIAGENDR ' , ' RIDAGEYR ' , ' DMDEDUC2 ' , ' BMXWT ' , ' BMXHT ' , ' BMXBMI ' )] data = data . rename ( columns = { ' SEQN ' : ' id ' , ' SMQ020 ' : ' smoking ' , ' RIAGENDR ' : ' gender ' , ' RIDAGEYR ' : ' age ' , ' DMDEDUC2 ' : ' education ' , ' BMXWT ' : ' weight ' , ' BMXHT ' : ' height ' , ' BMXBMI ' : ' bmi ' }) One thing worth knowing about NHANES: all the columns come in as numeric codes. 1 means Male, 2 means Female. 1 means the person smoked, 2 means they didn't. You have to map these to readable labels before doing any analysis, otherwise your charts are meaningless. Step 2 — Cleaning the Data Drop the ID column and remove nulls data . drop ( ' id ' , axis = 1 , inplace = True ) data . dropna ( inplace = True ) This brought us from 5,735 rows down to 5,406 — about 6% lost, which is acceptable. Remove outliers using the IQR method The IQR (Interquartile Range) method flags values that fall too far outside the middle 50% of

The opening kickoff of Super Bowl LVII was still three weeks away when sharp bettors began their work. While casual fans were scrolling through prop bets and debating quarterback matchups on social media, a handful of disciplined bettors with sophisticated models were already identifying the first exploitable edges. Within hours, sportsbooks registered the shift: Kansas City opened at -2.5, but sharp action pushed the line to -3. By game day, it had settled at -2.5 again after public money flooded in on the Chiefs. This seemingly minor dance of numbers contains profound lessons about market efficiency, behavioral psychology, and where consistent value in sports betting actually exists. The difference between sharp money and public money isn't merely a matter of skill—it's a window into how financial markets process information in real time. For researchers, data scientists, and anyone interested in understanding how markets function under uncertainty, betting lines offer a peculiar advantage: instantaneous, objective outcomes. You can know within hours whether your hypothesis was correct. This article explores what line movement data reveals about market inefficiencies, the methodology behind detecting them, and what this teaches us about information asymmetry in competitive markets. The Hidden Market Beneath the Surface Most bettors see a line and make a decision: is this price fair or favorable? But they miss the crucial information happening before they ever see that number. Sportsbooks don't set lines based on game probability—they set them based on where they predict the money will flow. This distinction transforms betting markets into fascinating research subjects. Consider the structure: A sportsbook's primary goal isn't prediction; it's profit through balanced exposure. They're market makers, not forecasters. When sharp bettors arrive first with informational advantages, they move the line. When public money arrives later with no informational advantage but