API Pagination Extractor
Paginated API pulls are hand-coded each time - one generic extractor speaks offset, page, cursor, and link-follow, with 429 retry/backoff built in.
Move data in
It lives somewhere else and it has to land here, on a schedule, without losing rows.
API to Warehouse Connector
Every Monday, someone on the team downloads CSVs from five different SaaS tools. This replaces that.
Backfill Planner
Backfills are error-prone manual scripts - chunk the date range, persist every status, retry the flaky, quarantine the broken, and resume after a crash without repeating a success.
CSV to PostgreSQL Loader
Loading data into our database is a manual nightmare - copy-paste into SQL, fight with column types, pray nothing breaks.
CSV/JSON to Parquet Partitioner
Our data lake is a pile of CSVs - convert flat files into hive-partitioned, compressed parquet and prove the query speedup with a stopwatch.
Incremental / CDC Loader
Full reloads take hours every night - load only what changed with watermark extraction, key upsert, and soft-delete tombstones.
Streaming Window Aggregator
We can't compute rolling metrics live - event-time windows with watermarks in ~100 lines, with the same semantics Flink and Spark charge a cluster for.
Make raw values usable
The bytes arrived. What they MEAN is a decision, and every layer decides differently.
A Field Name Is an Identity, Not a String
headers["X-Request-Id"] is the wrong shape, and not for style reasons. RFC 9110 makes field names case-insensitive; RFC 9113 makes them lowercase on an HTTP/2 wire or the message is…
A File Has No Lines In It, a Splitter Makes Them
A file has no lines in it. It has bytes. Lines are produced by a splitter, and every runtime ships a different one. wc -l counts LF bytes.
A Numeric String Does Not Contain a Number, a Reader Assigns One
A numeric string does not contain a number. It contains characters. A reader assigns the number - and a reader is four things multiplied together: a symbol table, a grouping rule, a…
A Percentage Column Is an Apportionment
Round three equal rows to one decimal place and the column reads 99.9%. Every row is correct.
A String Does Not Contain a Boolean, a Reader Assigns One
A string does not contain a boolean. A reader assigns one - and a reader is three separate decisions: an accept table (which spellings mean true), a normalisation (case, whitespace…
Allocate a Total Instead of Rounding Its Rows
A rounding function returns a number. It cannot return the fact that the rows no longer add up.
Byte-Accurate Fixed-Width Parsing with a Pre-Flight Audit
Reading a fixed-width file is one line: pd.read_fwf(path, colspecs=...). That line decodes the record to a string and then slices it by character position.
CSV Cleaner
Turn a messy CSV export into a tidy, analysis-ready file - with an auditable report of every change.
Dedup & Survivorship Pipeline
Duplicate records everywhere - normalize the variants, cluster by match keys, and merge each cluster into a golden record with field-level survivorship and full provenance.
Duplicate File Finder
Find byte-identical files hiding under different names - efficiently - and reclaim the space. So a shared drive full of duplicates stops costing you storage.
Enumerate Every Viable Dialect Instead of Returning One
csv.Sniffer().sniff() returns a Dialect or raises csv.Error. It has no third answer.
JSON Flattener
Every API response is nested. Every warehouse table is flat.
Local-to-UTC Resolution with the Undecidable Cases Named
Normalising a local timestamp to UTC is one line: ts.replace(tzinfo=ZoneInfo(zone)).astimezone(utc).
Lossless SQL Type Inference from Raw Text
Everything imported as string. Typing the columns is easy - typing them without silently corrupting data is the actual job.
ORDER BY Name Is a Collation, Not an Order
ORDER BY name reads like a total order on a column. It is not. It is a collation applied to bytes, and a collation is three decisions the SQL never states: which sequence the characters are…
Render a GFM Table and Report What It Cannot Carry
Turning a DataFrame into a markdown table is one line. That line is also a lossy encoder, and the GitHub Flavored Markdown table extension has no error state to say so.
Report Every Reading a Conforming Parser Could Return
parse(text) -> timedelta cannot answer the question people actually have about a duration string, which is not "how long is this" but "does the parser at the other end of this pipe agree with me".
Report the Collisions a Sanitiser Creates
sanitise(name) -> str is the wrong shape, and not because of which characters are on the deny-list.
Report the Collisions a Slugifier Creates
slugify("Node.js at scale") and slugify("NodeJS at scale") both return nodejs-at-scale. Neither call is wrong. Neither raises.
Truncate to N Does Not Name an Operation
"Truncate to 20" does not name an operation. A truncator is three decisions - a unit of length, a boundary rule, and a policy for the piece it removes - and the integer 20 carries none of…
Prove it is right
Something changed upstream. Find out before a dashboard does.
Column Anomaly Detector
Bad values slip into columns and nobody notices until a dashboard breaks - scan every column and get triaged, explainable alerts before they do damage.
Data Contract Validator
Producers break schemas without warning - put a YAML contract in CI so breaking changes block the merge and violating data never lands.
Data Quality Scorecard
Stop guessing how bad your data is - get a 0-100 score, a letter grade, and a ranked list of what to fix first.
Dataset Snapshot Diff
Between yesterday's snapshot and today's, what actually changed? Nobody can answer without eyeballing two files - this diffs the same dataset over time on a key column and tells you exactly…
dbt Test Generator
No tests on our models - this profiles a table and auto-generates a paste-ready dbt schema.yml.
DQ Rules Engine
Our data-quality rules live in a wiki nobody enforces - by the time a bad batch is caught, it's already in a report.
Great Expectations Config Generator
Setting up data validation is tedious - so most teams skip it until something breaks in prod.
JSON Schema Validator
A FastAPI microservice that validates JSON payloads against a schema and returns every error at once - so broken API contracts surface before they corrupt data.
Null Heatmap & Missingness Report
df.isna().sum() answers the wrong question.
Schema Diff Tool
Database migrations break things silently - catch schema changes before they hit production.
Schema Registry
A producer changes a column type and three downstream jobs break silently - we find out from the dashboard, not before.
Source-to-Target Reconciliation
We copied a table from the source system into the warehouse and can't prove it matched - row counts drift and values silently differ.
Control who sees what
Who owns this column, who may read it, and can you show an auditor?
Business Glossary Manager
"Active user" means five different things to five teams, so every metric argument starts from zero - this gives each business term one owned, versioned definition and flags the governance gaps that let ambiguity creep back in.
Column-Level Lineage Parser
Table-level lineage tells you model B reads model A. It can't tell you that renaming orders.amount will silently break total_revenue three hops away.
Compliance Checker
Run a document against a policy ruleset before the auditor does.
Consent & Purpose Tracker
We process personal data for purposes the user never agreed to - and can't prove otherwise when asked.
Data Access Auditor
"Who can see the PII table?" takes a week of digging, so over-privileged and stale access piles up unnoticed - ingest your access grants and get triaged, explainable findings for a least-privilege review in seconds.
Data Dictionary Generator
Point it at a table, get a documented data dictionary - types, PII flags, descriptions.
Data Lineage Visualizer
Paste SQL, see what breaks when you change a table.
Data Retention Enforcer
We keep everything forever - storage bloats and we are one audit away from a finding for holding personal data past its purpose.
ERD Generator from SQL
"We don't have database diagrams" - paste your CREATE TABLE statements and get an instant ER diagram.
Lightweight Data Catalog
Nobody knows what tables and columns exist, what they mean, or who owns them - so every analysis starts by pinging three people on Slack.
Model Card Generator
"No model documentation." - a model with no card is a black box: nobody knows what it's for, what it trained on, or where it breaks.
PII Detector and Masker
"We accidentally share sensitive data in staging" - scan CSV data for PII and mask it before it leaves production.
Policy-Driven PII Redactor
"We masked the PII, so the extract is safe to share." Two things are usually wrong with that.
Privacy Policy Generator
Legal templates cost money and are still generic - answer a short questionnaire and get a GDPR/CCPA-aware policy draft plus a checklist of the compliance work the document can't do for you.
Subject Access Request Extractor
A subject access request arrives with one email address. Finding the rows is easy.
Know when it breaks
The pipeline fails at 3am. The question is whether anyone finds out before the meeting.
Cron Job Monitor
A dead-man's-switch for scheduled jobs - heartbeat in, email out when a job goes silent. So you find out batch jobs failed before your users do.
Data Freshness Monitor
Dashboards show stale data and nobody notices - this tool catches it first.
Data Pipeline Monitor Agent
"We discover pipeline failures from angry users" - an agent that catches failures, slow runs, row-count drops, and silent staleness before anyone downstream notices.
Database Health Dashboard
One screen of database visibility - cache hits, latency, locks, replication lag - scored red/amber/green so problems surface before users notice.
Environment Variable Checker
A pre-flight check for environment variables - declare what your app needs, validate it before you ship, and fail the deploy if anything's missing or malformed.
Incident Response Agent
Incident playbooks exist but nobody follows them - this agent executes the runbook step-by-step and escalates when a remediation step fails.
It Answered From Last Quarter's Document - the Index Read 89% Fresh While 27% of Answers Quoted a Changed Fact, and the Dashboard Ranked a Dearer, Worse Reindex Schedule Above a Cheaper, Better One
The dashboard said the index was 89% fresh, and 27% of the answers it served quoted a fact the source had already changed.
It Is Slow and Nobody Knows Where - Per-Hop p99s Summed to 792 ms Over an SLO the Chain Missed by 51, the LLM Owned 83% of the Mean and the Reranker 52% of the Tail, and Three Config Fixes Beat a Smaller Model on p99 While Saving 6x Less Mean
Every hop of the RAG chain has its own latency dashboard, and adding them up said we were 792 ms over the SLO.
It Started Refusing Valid Requests - the Dashboard Read the Upgrade as a Two-Thirds Cut in Refusals While True Refusals Rose 39% and Over-Refusal Doubled, and a Campaign Week Doubled the Rate With Both Errors Unchanged
The dashboard said the upgrade cut refusals by two thirds. The model was refusing 39% more requests than before, twice as many of them legitimate - it had started saying "I'm not able to"…
Log Parser and Alerter
Stop grepping logs by hand. Parse mixed-format logs into structured records, see the severity breakdown and top errors, and fire alerts when errors spike.
Metric Anomaly Alerter
We notice metric drops late - this watches a metric and fires the day it moves, with three independent detectors.
Model Drift Detector
Models degrade silently in production - this watches a live window against the training reference via PSI and alerts before accuracy quietly rots.
Pipeline SLA Monitor
A pipeline that lands late, runs slow, goes stale, or arrives thin usually breaks a consumer before anyone on the data team notices.
Find out what is in it
A new dataset landed. Two hours of profiling before you can say anything about it.
Auto-EDA Dashboard
Drop in a CSV, get the whole first-look in seconds - shape, missingness, dtypes, distributions, correlations, and the quality problems worth a human's attention.
Correlation & Multicollinearity Explorer
"Which features relate?" - a correlation heatmap plus VIF to catch the redundant, collinear features that quietly wreck a model.
Customer Segmentation Tool
Find your natural customer groups with KMeans - auto-picks the number of segments, names each one, and shows how distinct they really are.
Outlier Explainer
"Which rows are weird, and why?" - Isolation Forest finds the anomalies; per-feature z-scores explain each one.
PCA/UMAP Explorer
"Too many features to see" - project high-dimensional data down to 2D and actually look at its structure.
Survey Results Analyzer
Turn a raw survey export into a type-aware summary in seconds - NPS, Likert scores, choice breakdowns, and open-text sentiment.
Topic Modeling Tool
"What are these documents about?" - discover the themes in a text pile automatically, no labels required.
User Feedback Analyzer
We have 10K reviews and no insights - this turns raw feedback into a sentiment split, ranked complaints, ranked praises, and the one thing to fix first.
Turn it into a number people act on
One metric, three dashboards, three answers. Define it once and serve it.
Cohen's d Is Not the Effect. It Is One Summary of It, and the Summaries Disagree
"d = 0.5, a medium effect" is read as a fact about how big something is - so this holds d exactly fixed, changes nothing but the shape of the distribution, and asks the other effect-size metrics whether they agree.
Cohort Analysis Tool
"Retention analysis takes our analyst 2 days" - turn a raw event log into a cohort retention heatmap in seconds.
Dashboard Spec Generator
Dashboards built from vague asks - this reads the data's shape and recommends the right charts, with the reasoning.
Funnel Analyzer
See exactly where users drop off - biggest leak called out, step-by-step conversion, and a segment comparison.
Inline SVG Trend Marks That Report Their Own Scale
Adding a sparkline column to a table is one line of pandas. That line autoscales every row to its own min and max, which is what makes a row that grew 3.5% and a row that grew 3.5% off a…
KPI Tracker
Execs ask for the same metrics every Monday - this turns a metric time series into the answer automatically.
KPI Tree / Driver Decomposition
"Why did revenue move?" - this splits a KPI's change across its drivers exactly, with no residual.
Metric Catalog & Ownership
Which metrics exist, and who owns them? A registry with tiers, dependencies, and automated governance that flags the metrics nobody is minding.
Metrics Layer / Semantic Definitions
Every dashboard computes revenue differently - a metrics layer makes one governed definition the single source of truth.
Natural Language to SQL
Non-analysts can't query the warehouse - NL→SQL lets them ask in English, behind guardrails that keep it read-only and in-schema.
Period-over-Period Metric Diff
"Conversion is up 5% this week!" - but is it a real shift or just noise? Metric Diff pairs a period-over-period delta with a significance test so you only act on changes that are real.
Pivot Narrator
Nobody reads the pivot table.
Sales Forecast Dashboard
Get sales forecasts out of someone's spreadsheet - a reproducible forecast with a confidence band and an honest accuracy check.
Scheduled Report Sender
Weekly reports sent by hand - this defines a report once on a cron and previews every scheduled send (a dry run that never emails).
Self-Serve Data Explorer
Analysts get pinged for every number - a self-serve explorer lets anyone pivot, aggregate, and filter without SQL or a ticket.
SQL Optimizer Advisor
Slow queries, no idea why? This lints SQL for the anti-patterns that quietly force full table scans.
Decide what is actually true
The line went up. Whether that means anything is a separate question.
A Measured Lift Is a Lift at an Age
The test said +10%, the launch delivered +2%, and the chart everyone blames it on cannot tell you whether the effect faded or whether you measured different users each week.
A Normality Test Is a Power Curve, Not a Gate
Shapiro-Wilk says your data is not normal, so you abandon the t-test. Across 63 measured cells that gate fires 29 times on a t-test that was already correct, and the substitute it sends you…
A p-value Is a Random Variable, and the Spread of That Variable Is the Result
"p = 0.049" gets written up as a finding, but nobody has ever watched the same study run twice - so this replays one true effect forty thousand times and reports what the p-value actually does.
A Segment You Found Is Not a Segment You Named
The test won overall, then somebody sliced it by country, device, plan and tenure and found the segment it hurt - and on a launch that helped every single user equally, that slicing invents a harmed segment 40% of the time, with a number the…
A Split Is a Hypothesis, and Passing Its Test Is Not Evidence
The experiment dashboard says the split came out 49.3 / 50.7. Somebody decides that is close enough to even and reads the result.
A Split Test Measures a Transfer Between the Arms
The A/B test said +26%. It shipped to everybody and nothing moved. Nobody made a mistake: a split test measures the difference between the arms, and the decision needed the total.
A Stopping Rule Is Part of the Test
Somebody opens the experiment dashboard every morning and ships the variant the day it goes green.
A/B Test Calculator
Stop eyeballing significance. Enter your A/B numbers, get a proper two-proportion z-test - p-value, confidence interval, and a clear verdict - plus a sample-size planner.
AIC Ranking Plus an Absolute Goodness-of-Fit Test
Fitting ten distributions and ranking them by AIC is four lines of scipy. The ranking will always produce a winner - including when the data came from something that is not on the list.
ANOVA Says Something Differs, and the Five Ways to Ask Which Pair Name Different Pairs
The ANOVA came back significant and now someone has to say which groups differ - and the five usual ways of finding out name different pairs on the same data.
Before-and-After Data Is Not Two Groups, and Analysing It as Two Throws the Correlation Away
We measured the same stores before and after, then tested them as two separate groups - and a change the data clearly shows came out as "no significant difference".
Crosstab & Chi-Square Tool
"Compare groups in survey data" - contingency table, chi-square test, effect size, and the exact cells driving the difference.
Four Procedures Are Called the T-Test, and They Do Not Have the Same Error Rate
"We ran a t-test" does not name a procedure. On 50 users against 10 with unequal spread, the textbook t-test calls a coin-flip significant 29% of the time while claiming 5%.
Four Tests on One 2x2, and the Test You Picked Decided the Launch
0 of 20 converted in control, 4 of 20 in the variant: two standard tests say it moved, two say it did not - and which one you ran decides the launch.
Mann-Whitney Is Not a Robust t-test, It Is a Test of a Different Hypothesis
Two defensible tests, one dataset, both significant, opposite conclusions - because "just use Mann-Whitney" swaps the question, not just the method.
No Significant Difference Is Not No Difference, and the n That Makes It One Is Computable
The test found no significant difference and everyone read it as "the two are the same" - but a small study says that about a difference that matters two times in three.
Overlapping Error Bars Are Routinely Significant, and on Paired Data the Rule Fails the Other Way
"The error bars overlap, so it is not significant" throws away up to half of the real results on the chart - and on paired data it fails the other way.
Parallel Trends Is an Assumption, and the Test for It Has a Power
We could not randomise this one, so we compared the change in the group that got the policy against the change in the group that did not, and checked that the lines looked parallel beforehand.
Sample Size & Power Calculator
"How many users do we need?" gets asked at the start of every experiment and answered by vibes.
Statistical Test Advisor
"Which statistical test do I use?" - describe the data, get the right test (parametric or not), and run it.
Testing an Assumption and Then Choosing a Test Is One Procedure
"We checked the variances first." Run Levene, then pick Student's or Welch's. Across 14 measured designs that two-stage move never once beat simply using Welch, and on four of them it…
The Control Group Is a Fitted Object, and Its Credential Is the Objective Function
One market got the intervention and there was never going to be a control group, so a weighted average of the other markets is fitted to stand in for one - and the pre-period plot everybody publishes as proof was chosen to look good, which is why…
Two Analysts, One Dataset, Two Verdicts - a p = 0.05 Result Is Evidence for the Null Once n Is Large
One analyst reported "significant, p = 0.047", the other ran a Bayesian test on the same numbers and reported "strong evidence of no effect" - and nobody could say which of them was wrong.
Variance Reduction Is a Bet on a Correlation You Already Collected
The test needs six weeks, and somebody says CUPED will halve it. The first claim is arithmetic.
Learn from it
Fit something, then find out honestly whether it is better than doing nothing.
A Backtest Number Is an Estimate, and the Protocol Decides How Noisy an Estimate
"The forecast was excellent in-sample" - a single train/test split reports one draw of an origin, and the model that won it was chosen by it.
A Prediction Interval Is a Claim About a Frequency, at a Horizon, in a Regime
"The point forecast is useless" - so the chart gets a 95% band, and the band's claim is never checked, because the check almost nobody runs has almost no power.
Baseline Model Ladder
"We got 0.87 AUC." Compared to what?
Batch Scoring Service
Scoring new data shouldn't be a manual copy-paste ritual.
Churn Predictor
Spot at-risk customers before they leave - ranked risk scores, churn drivers, and an honest model-quality check.
Class Imbalance Toolkit
"Our fraud model ignores the minority class." - a classifier that predicts "not fraud" every time scores 98% accuracy and catches zero fraud.
Classification Threshold Explorer
A model outputs a probability. A decision needs a cutoff. Almost everyone ships 0.5 - a number nobody chose, which quietly assumes a false positive and a false negative hurt exactly the…
Data Leakage Detector
The model scores 0.99 in cross-validation and falls apart in production. That gap is usually leakage.
Feature Engineering Pipeline
Every new ML project you rewrite the same impute / scale / one-hot boilerplate, get the column lists slightly wrong, and leak test statistics into training.
Feature Importance Explainer
"Stakeholders don't trust the model." One importance number is easy to doubt - three that agree is evidence.
General Time Series Forecaster
"Forecast any metric" - Holt-Winters (trend + seasonality) in pure numpy, with honest backtest accuracy.
Hyperparameter Tuner
We grid-search by hand - nudge n_estimators, re-run a cell, keep whatever looked better.
Mini Model Registry
"Which model is actually in production?" should have a one-word answer.
Monotone WOE Binning with a Defensible IV
Binning a feature and printing its Information Value takes twenty lines. Knowing whether the number you just printed means anything is the actual job - and the conventional 0.02 / 0.1 / 0.3…
Probability Calibration Checker
A "0.9" from your model should mean it happens 90% of the time. Often it doesn't.
Recommendation Engine
"We have no 'you may also like'" - item-item collaborative filtering, with an explanation for every recommendation.
Text Classification Trainer
"Auto-tag our support tickets" - train a TF-IDF + logistic-regression classifier and route new text by category.
Train/Eval Leaderboard
Model selection is ad-hoc - train a few classifiers, eyeball one accuracy number, ship whichever looked best on one lucky split.
Get answers out of documents
The answer is in a 200-page PDF nobody will read.
Chunking Strategy Tester
Bad chunks = bad answers. Prove which chunking strategy retrieves best - on evidence, not vibes.
Competitive Intel Summarizer
Turn scattered competitor notes into one comparison matrix + strategic takeaways.
Contract Clause Extractor
Turn a 3-day legal read into a 30-second clause + risk triage.
FAQ Generator from Docs
Stop answering the same question daily - turn your docs into a ready FAQ.
JSON Schema Inference from LLM Outputs
Your LLM pipeline returns JSON. Nothing checks it. json.loads is not a contract.
Knowledge Base Builder
Turn scattered docs into a searchable, cited knowledge base - before the expert leaves.
Meeting Notes Summarizer
"Meetings produce no actionable output" - paste the transcript, get TL;DR + decisions + action items + open questions.
Named Entity Extractor
"Pull the names, orgs, money, and dates out of this text" - structured entities, no heavy model download.
Near-Duplicate Finder for a RAG Corpus
A duplicate in a vector index does not cost you storage. It costs you the answer.
PDF Q&A Bot
"Nobody reads the 200-page policy manual" - upload it, ask questions, get cited answers.
Resume Screener
Rank a stack of resumes by skills fit - a triage aid, not a hiring decision.
Semantic Search Engine
Search by meaning, not just exact words - what Ctrl+F can't do.
Structured Output Enforcer
Stop your pipeline from crashing when the LLM returns almost-JSON.
Check the AI is any good
It sounds right. Sounding right is not a measurement.
A Golden Set Is a Sample of Last Year's Traffic, and Reweighting Cannot See What It Never Sampled
We labelled 240 real conversations once, called it our eval set, and a year later it still said 90% while customers were getting 84%.
A Prompt CI Gate Is a Test With a False-Alarm Rate, and Zero Tolerance Is Always Red
Our prompt CI said eight cases broke on a change that fixed a typo, so we stopped reading it - and then a real regression shipped.
Agent Evaluation Dashboard
"We don't know if our agents are actually good." Run a fixed eval suite against two agent versions and get pass rate, quality, latency, and per-category regressions - numbers instead of vibes.
An Unchanged Score Is a Net, and McNemar Is Quieter on a Swap Than on No Change
We upgraded the model, the eval score didn't move, so we shipped it - and the refunds team said it had stopped working.
Few-Shot Example Selector
Stop shipping the same static few-shot examples for every input - pick the ones that actually resemble the query.
Hallucination Detector
Flag the claims your RAG answer makes that the retrieved context never supported.
LLM Cost & Token Tracker
No more surprise API bills - log every call, roll spend up, watch the budget.
LLM Guardrail Filter
Model outputs unsafe / off-topic text? Put cheap deterministic checks around it.
LLM Model Router
Stop paying frontier-model prices for "what's the capital of France?"
Prompt Linter
Prompts are the only part of an LLM system that ships with no review gate.
Prompt Registry & Versioning
Prompts get pasted into random .py files and Slack threads - no history, no diff, no rollback. This gives them the same discipline as code and models.
RAG Evaluation Harness
You changed the chunking or swapped the embedding model. Did RAG get better or worse? "It feels better" is not an answer - this harness gives you numbers.
Semantic Response Cache
Stop paying for repeated similar queries - reuse an answer when the meaning matches.
The Answer Was Right and the Path Was Not, and Strict Reference Match Reported -1 for a -31 Change
Our agent eval said the refund bot got the answer right, and it had sent the refund to the wrong order.
The Citation Does Not Say That - Citation Coverage Rose 24 Points While the Cited Passages Supported 8 Fewer Claims
We told the model to cite every sentence, citation coverage went from 74% to 98%, and the passages it cited supported 8 points fewer of its claims.
The Fact Was in the Context, and the Model Did Not Read It - Context Recall Said 83% of Needed Facts Were in the Window, 34% of Queries Lost One That Was There, and a 12k Window Scored +9.9 on Recall and -10.6 on Answers
Context recall said 83% of the facts the answers needed were in the window, and 34% of the queries lost a fact that was there - more than were lost to retrieval, and the dashboard counts every one of them as a success.
Token & Cost Estimator
Price an LLM feature before you build it - not after the invoice lands.
Make it run itself
The task is small, correct, and done by hand every single week.
Accessibility Checker
We don't test for accessibility until the end - paste HTML and get WCAG-referenced findings, severity levels, and a 0-100 score in seconds.
Airflow DAG Generator
Writing DAG boilerplate is repetitive - describe the pipeline in ~30 lines of YAML and get a validated, ready-to-deploy Airflow DAG.
Auto-README Generator
Scan any repo into a structured profile and generate a README - by template offline, or with Claude for richer prose. Wire it into GitHub Actions to auto-open a docs PR on every push.
Code Review Agent
"PRs sit unreviewed for days" - an agent that gives every PR an instant first pass so reviewers focus on logic, not nitpicks.
Customer Ticket Router
"Tickets go to the wrong team 40% of the time" - an agent that routes tickets by weighted keyword signal, and escalates to human triage instead of guessing when confidence is low.
Daily Standup Bot
One place for the team's standup. Everyone submits yesterday / today / blockers; the bot rolls it into a single digest with blockers surfaced first. No more scrolling Slack threads.
dbt Model Generator
Writing dbt staging models from scratch is repetitive boilerplate - paste your SQL DDL and get production-ready dbt artifacts instantly.
Email Draft Agent
"I spend 2 hours daily on email" - an agent that drafts replies, checks its own work against the original ask, and revises before handing it back.
Markdown to PDF Report
Write the weekly report in Markdown, get a branded, print-ready PDF - cover header, styled tables, page numbers. No more hand-formatting in Word.
Onboarding Checklist Agent
"New hires miss steps, ramp-up takes 3 months" - a multi-step agent that resolves onboarding step dependencies and flags what's overdue or blocked before it cascades.
Read a Cron Line the Way Cron Reads It
0 0 13 5 is the line somebody writes for Friday the 13th. It fires on 61 days a year instead of 3, because when both day fields are restricted cron takes their union.
Report Generation Agent
Monthly reports take 2 days of copy-paste - a team of role-specialized agents turns a table of metrics into a ready-to-send report in seconds.
Report the Arrival Process a Backoff Creates
backoff(attempt) returns a number of seconds. That is the whole interface, and it is the wrong shape for the problem.
Slack Q&A Agent
People ask the same questions in #general every day - this agent answers them from a knowledge base via RAG, and escalates anything it isn't confident about to a human instead of guessing.
SQL Query Formatter
A FastAPI microservice that turns unreadable SQL into a clean house style - and lints it for danger (DELETE with no WHERE) and style (SELECT , implicit joins).
Choose, and be able to defend it
The number is on the screen. Someone still has to decide, and later justify it.
A Decision Log Is an Instrument, and an Instrument Has a Scoring Rule
A decision log without a prediction attached is a diary: it records what was chosen and can never say whether choosing it was any good.
A Guardrail Is a Constraint, and a Constraint Has a Power
Every experiment review has a checklist of counter-metrics. Somebody confirms none of them moved, and the change ships.
A Lead Is a Claim About a Lag, and a Lag Has to Be Estimated
Revenue lands a month after anyone could have done something about it, so somebody builds a table of candidate upstream metrics, correlates each against revenue at a few lags, sorts by the biggest number and puts the top three on a dashboard.
A Proxy Metric Is a Bet That a Correlation Survives Being Optimised
A proxy metric gets picked because it correlates with the outcome somebody cares about. It becomes a target. It goes up.
A Target Is a Method Plus a Claim About the Future
Somebody asks what next quarter's number should be, and six people in the room have six defensible answers.
An Invoice Is a Joint Cost, and a Joint Cost Has No Unique Owner
Finance forwards the warehouse bill and asks which team spent it. The question sounds like accounting and it is not.
An Ordering Is Not a Schedule
Somebody asks what to do first, so you score the backlog - RICE, or WSJF, or cost-of-delay-over-duration - and the meeting ends with a ranked list. But a list is not what you pay.
Capability Map - What This Covers for a Team
A catalog is only useful if somebody arriving with a problem can find the thing that solves it.
Expected Value Is a Number, Not a Decision
Somebody asks which option to take, so you build the spreadsheet: probability times payoff, one row per option, pick the biggest.
Feature Prioritization Tool
We argue about priorities without data - RICE turns "I feel strongly about X" into a rankable, defensible number.
OKR Tracker and Advisor
OKRs get set and forgotten - this tracks progress against pace and flags which key results are behind while there's still time to act.
Product Roadmap Visualizer
Roadmaps are PowerPoints that go stale - generate the roadmap from data instead: a Gantt-style timeline, grouped by lane, colored by status, with a "today" line.
Sprint Retrospective Generator
Retros are unstructured and repetitive - feed in the sprint numbers and get grounded observations and concrete action items in your team's format.
Startup Idea Validator
We build before we validate - this scores an idea across five dimensions, builds a Lean Canvas, and hands you the cheapest experiment to test your riskiest assumption first.
The Exercise Works, the Matrix That Scores It Cannot Rank
A pre-mortem is cheap and it works: assume the project already failed, then write down why. What happens next is the problem.
User Story Generator
Writing good user stories takes practice - paste raw feature ideas and get INVEST-scored stories with Given/When/Then acceptance criteria.
When Should a Human Take Over? - 'Escalate Below 90%' Flooded a 700/Day Team and Cost More Than Never Escalating, Calibrated Per-Segment Pricing Lost to One Cutoff Because Capacity Binds, and a Same-Skill Model With Louder Confidence Cut Escalations to 1.2% While the Bill Rose 27%
We told the AI agent to hand anything under 90% confidence to a person. It sent 3,655 tickets a day to a team that can review 700, and the day cost more than letting the agent decide…
The spine underneath
The control plane the tools plug into: identity, permissions, audit, connectors, orchestration.
Gateway + login
Authentication - who are you?RBAC + app registry
Authorization - what can you open?Audit log
Provable who-did-what-when, append-onlyConnector layer
One safe place for data-source credentialsMount a real app
Wrap an existing build behind the shellOrchestration (Airflow)
Govern scheduled pipelines (OSS)Nothing matches that. Try a symptom rather than a tool name, or clear the filter.