The LLM plans.
The browser engine executes.
In-Browser Model Context Protocol (MCP) tools and deterministic Smart Recipes draft an optimized DAG from anonymized statistical metadata. DuckDB-WASM executes every transformation zero-copy inside your browser Web Worker. In Local Mode, raw data never leaves your device.
(0 Bytes Egress)
1. Parsing Raw Dataset locally in browser memory...
(DuckDB-WASM + WebGPU)
In-Browser Ingestion & Conversion
Structured tables (CSV, JSON, Excel, Parquet) and unstructured documents (PDF invoices & tables) parse directly inside your browser tab into DuckDB-WASM and Apache Arrow. Zero cloud uploads.
Client-Side Metadata Profiling
DuckDB-WASM queries columnar structures via Web Workers in milliseconds. Statistical distributions, cardinality, and null percentages are calculated without sending raw cell data outside your machine.
BYOK MCP Planning & Recipes
Your chosen LLM (OpenAI, Anthropic, Gemini, Groq) or 1-click deterministic Smart Recipes construct a rigorous execution DAG using in-browser Model Context Protocol tools. The agent plans; you review.
WebGPU Similarity & Clustering
High-dimensional categorical clustering, KNN imputation, and entity deduplication run client-side using WebGPU compute shaders and WASM math kernels. Zero cloud compute required.
Deterministic Execution & Audit
The browser engine executes vectorized SQL recipes directly in Web Workers. Every cleaning rule, schema mutation, and column operation is chained into a tamper-evident SHA-256 cryptographic audit manifest.
The pipeline engine is deterministic.
But real-world data is chaotic.
You can't use standard exact-matching on messy, human-entered data. It breaks. And relying on generic LLMs to compare millions of rows is impractically slow, expensive, and prone to hallucinations.
We need a better way to cluster reality.
Centers = 8
The Silver Ratio.
Proportion 1 : 2.414 · KNN & Fuzzy Clustering
Standard strict-matching fails on human-entered data. RTIO embeds categorical variables into high-dimensional vectors. Entities like "Apple Inc" and "Apple, Inc." are resolved client-side via hardware-accelerated nearest-neighbor (KNN) cosine similarity in WebGPU and WASM math kernels.
STATUS: RESOLVING ENTITIES CLIENT-SIDE
ACCELERATION: WebGPU Shaders + WASM SIMD
COMPLEXITY: O(log N) Vector Space
DISTANCE METRIC: COSINE SIMILARITY
The AI only plans — it never touches your data · LLM receives metadata only · Physical zero egress in Local Mode
Complete Observability & Audit
Every pipeline execution appends to a tamper-evident SHA-256 cryptographic audit manifest stored in IndexedDB. Inspect exact SQL queries, column lineage, before/after statistical diffs, and data quality scores at each node — fully exportable for compliance and auditing.
ENGINE: DuckDB-WASM SIMD + Web Worker
12:04:01 - Parse: sales_records.csv → DuckDB-WASM in-tab
12:04:01 - Columnar profile: 38ms — metadata extracted, 0 bytes egress
12:04:01 - Quality scan: 24ms — 5 issues, score 68/100
12:04:01 - Vectorized DAG: 18ms — dedup + date_std + outlier filter
12:04:01 - WebGPU KNN impute: 41ms — 18 missing values filled
12:04:01 - Audit manifest: 2ms — SHA-256 chain sealed
[OK] Total: 123ms. Network: 0 bytes. Memory: 14MB tab heap.
Competitive Landscape
Architected differently from the ground up.
Comparing RTIO against legacy enterprise data preparation platforms, desktop open-source tools, and API orchestrators.
| Dimension | RTIO | Alteryx Designer Cloud | OpenRefine | n8n |
|---|---|---|---|---|
| Compute Location | Browser Tab (DuckDB-WASM) | Cloud Server or Desktop | Local Java Desktop | Node.js Server / Docker |
| Data Privacy | Local Mode: 0 bytes egress (Never leaves browser). Cloud: Ephemeral | Data in vendor cloud (Designer Cloud) or local | Local machine, no cloud | Raw payloads pass through server worker |
| Pricing | Free unlimited local compute | Starter: $250/user/mo ($3,000/yr); Enterprise quote-only | Free Open Source | Free self-hosted; Cloud: $20-$500/mo |
| Large Dataset Handling | DuckDB columnar vectorized; stream queries via File System API | Desktop local RAM; Cloud runs queued | Memory-bound Java heap; crashes on high cell counts | Node.js heap bound; documented OOM on large JSON |
| Offline Mode | Yes (PWA + OPFS) | Desktop only | Desktop only | No |
| UI Paradigm | Visual DAG + Virtualized Spreadsheet Grid | Visual DAG Canvas | Table Grid only | Node Graph (API-oriented) |
| AI Integration | In-Browser BYOK MCP Planner + 1-Click Smart Recipes | Add-on AI Assistant (Paid add-on) | None | LLM Action Nodes |
* Note: Alteryx pricing sourced from Alteryx.com public Starter tier ($250/user/mo billed annually), Sept 2026. Alteryx Designer Cloud was formerly known as Trifacta / Google Dataprep.
Architecture FAQ
System Mechanics
How is RTIO different from a standard LLM wrapper?
What is the client-side Parquet-first pipeline?
How does RTIO handle large datasets in the browser?
How does RTIO handle missing data imputation and outlier detection?
What cleaning operations are supported?
What file types and data sources can I ingest?
What is Entity Resolution and how does it work client-side?
Ready to clean data the right way?
Free tier includes unlimited local compute. No credit card required.