Practical AI Systems
30 small, honest AI systems you can run right here: no install, no API key, nothing leaves your browser. Each one has a Python version, tests, an evaluation, and documented limitations. View on GitHub · Contribute a system
Beginner
Document Q&A with citations
Find a source paragraph, answer from its evidence, cite it, or abstain.
Invoice field extraction
Extract invoice fields, then validate dates and arithmetic.
Repository hygiene checks
Run simple local checks and see why toy patterns miss real risks.
Extractive text summarizer
Pick the most representative sentences from a short document using word-frequency scoring.
Near-duplicate question detector
Flag probable duplicate support questions with token Jaccard similarity.
PII pattern redactor
Find and mask obvious email, phone, SSN, and credit-card-shaped strings in free text.
Lexicon-based sentiment classifier
Classify short text as positive, negative, or neutral using a small hand-built word list.
Readability grade scorer
Estimate a Flesch-Kincaid-style reading grade level and bucket it into simple/standard/complex.
FAQ intent router
Route a short support message to one of five fixed intents using keyword overlap.
CSV/record schema validator
Catch missing required fields and type mismatches in structured records before they reach an AI pipeline.
Intermediate
Post-OCR document extraction
Normalize OCR line order from bounding boxes, then extract fields. OCR itself is not run.
Entity spans and error analysis
Extract person/date/money spans and inspect exact-match precision, recall, and F1.
Exact-response cache
Explore normalized keys, TTL expiry, LRU eviction, and model/version isolation.
Synonym query rewriter
Expand a query with a small synonym list before BM25 retrieval, and measure the recall it recovers.
Prompt-injection pattern scanner
Scan untrusted text for common prompt-injection phrasing before it reaches a model.
Cost and latency guardrail tracker
Summarize a request log into p50/p95 latency and cost, then flag budget breaches.
Post-OCR table reconstruction
Cluster OCR line bounding boxes into rows and columns to rebuild a simple table grid.
Image caption QC
Run rule-based checks on autogenerated image captions: placeholders, length, and required keywords.
RAG citation support checker
Verify that inline citations in a generated answer are lexically supported by the cited chunk.
Schema drift detector
Diff a batch of records against a declared schema: new fields, missing fields, and type mismatches.
Advanced
Voice-agent transcript evaluator
Score call success, forbidden actions, and latency separately.
Judge agreement audit
Compare a heuristic with human ratings and inspect rater disagreement.
Approval-gated tool workflow
Constrain tools with an allow-list and require approval before sensitive actions.
Local knowledge-base MCP server
Expose local search and source reads as JSON-RPC-style tools.
Multi-agent triage handoff
Route a request through a triage step to the right specialist, escalating when confidence is low.
Retrieval evaluation harness
Score retrieval with recall@k, MRR, and nDCG@k, and compare two strategies side by side.
Canary rollout simulator
Compare canary metrics against a baseline to decide promote or rollback.
Adversarial red-team harness
Run adversarial requests against a bounded agent's policy and measure the breach rate.
Trace anomaly detector
Flag anomalous execution-trace durations with a z-score against a historical baseline.
Multi-hop reasoning router
Split a multi-part question into sub-questions and abstain unless every part is answered.