I implemented a neural network in SQL
Running neural networks in SQL on climate data proves the relational array model works.
Data tools and analytics from Show HN — visualization, ETL pipelines, databases, and data science utilities.
Running neural networks in SQL on climate data proves the relational array model works.
Public fraud detection across 227M Medicaid claims—1,860 flagged providers, zero false positives yet.
1.3M case analysis proves systemic sentencing bias with rigorous controls; visceral real-time cost counter.
Autonomous discovery of dynamic laws and world structure in Elementary Cellular Automata. Atlas of 20 worlds, 7 cycle laws, fragility spectrum. Preprint: https://doi.org/10.5281/zenodo.20516375
69x causal compression on period-15 CA oscillators with deterministic law discovery.
350x faster GPU Bloom filter with academic paper backing the performance claims.
Google Trends for Hacker News lets you settle editor wars with 18 years of data.
Fluid-stochastic network clearing on real UN trade data makes geopolitical shocks feel terrifyingly concrete.
Bills + AI summaries + bias-rated news + politician posts all in one.
AI-extracted names from 150k Roman inscriptions mapped across the empire.
PostgreSQL extension for BM25 relevance-ranked full-text search. Postgres OSS licensed.
Native Postgres BM25 with ORDER BY content <@> 'query' operator syntax.
Explore 230 global music charts from 1940–2025 with a Back to the Future UI.
ACORN-1 pushes WHERE filters into HNSW traversal, fixing pgvector's post-filter limitation.
Reading a paper means opening a PDF, then hunting separately for the code, the citations, whether it replicated, and what genes/drugs it touches. I spent a few months trying to fix that. Two parts: 1.
8.5M papers in a WebGL map with MCP server access for AI agents.
Finally, a history archive that lets you search by meaning instead of just keywords.
Visualizes colonial extraction arcs no other tool maps this explicitly.
Aggregates nested counters into time buckets at write time, skipping event storage entirely.
Real-time MCP sentiment tracker when everyone's debating if the protocol is dead or thriving.
Stunning 3D galaxy view makes exploring YC's entire portfolio actually fun.
Tracks a promise made in chapter 8 through 92 chapters of narrative drift.
Infers invisible submarine cable paths using A* pathfinding between ocean-crossing hops.
Recovers metadata for deleted videos across 1.5B indexed entries since 2005.
SirixDB is an an embeddable, bitemporal, append-only database system and event store, storing immutable lightweight snapshots. It keeps the full history of each resource. Every commit stores a space-efficient snapshot through structural sharing. It is log-structured and never overwrites data. SirixDB uses a novel page-level versioning approach.
Structural sharing below page level beats ZFS block copying and event replay.
Git-like branching for columnar data with DuckDB-beating benchmarks from pure JVM.
GPU-accelerated layout renders 7M articles with shortest-path finding between any two.
Agent-guided compilation handles merged cells and multi-level headers LLMs choke on.
Datalog engine that finally beats SQL on graph queries without the Datomic tax.
30,000 AI responses revealing systematic bias patterns across 100 models.
207 papers buried in government records, now searchable with full context extraction.
Distributes DuckDB queries across BEAM nodes when Polars maintenance became unsustainable.
1.36B articles queryable in milliseconds solves a decade of researcher scraping pain.
Zero-copy Arrow queries from DuckDB to Snowflake without ODBC row-column conversion overhead.
Open-source streaming SQL engine written in Rust using Apache Arrow and DataFusion. Supports continuous queries, temporal stream joins, tumbling/session windows, and CDC/Kafka connectors. Lightweight, embeddable, and sub-microsecond latency
Sub-microsecond streaming SQL via zero-alloc hot path; genuine advancement over SQLite+DataFusion.
Reactive SQL graph in the browser beats Tableau for local prototyping.
Turning 184,000 reviews into 'build vs skip' verdicts inside Claude Code is genuinely clever.
Proving an open math conjecture using nothing but SQL cross-apply queries is wild.
3,000 years of literary citation networks visualized with WebGPU; LLM pipeline resolves messy bibliographic data.
7,560 runs proving cheaper models beat expensive ones on production OCR tasks.
Stunning 3D globe tracking 725 AI clusters and 80M H100 equivalents globally.
WebGPU renders millions of galaxies in-browser with real-time density correction filters.
Runs billions of rows locally on your machine, bypassing cloud warehouse costs entirely.
Turning 98 live seismic feeds into a global 'quietest place' tracker is a genuinely clever data mashup.
Petal visualization of population density is an instantly graspable data viz win.
24M spending rows searchable instantly; missing data but fills a real transparency gap.
Beats utility forecasts on 6 of 7 RTOs using only public EIA data and open models.
Monte Carlo sampling inside a spreadsheet beats collapsing uncertainty into single scalars.
Column-oriented analytics for SQLite, designed for fast local analytical queries with benchmarked speedups of up to 130,000x over standard SQLite scans on selected workloads.
Columnar storage inside SQLite delivers 130,000x speedup on aggregation scans.
BM25 search directly on S3 Parquet files without ingesting data into Elasticsearch.
Local-first hybrid vector database. HNSW + BM25 + fusion search. Single binary, no cloud.
Single-binary Rust DB fusing HNSW and BM25 without cloud dependencies or API keys.
591 projects