Unsiloed AI

Market position

Unsiloed AI positions itself as an infrastructure or 'document layer' provider beneath enterprise AI and RAG applications, competing with general-purpose document parsing and OCR tooling. It explicitly names LlamaIndex, Gemini, Mistral and Unstructured.io as the solutions it benchmarks against. Claimed customers span Fortune 150 banks, NASDAQ-listed companies and early-stage startups, including 10+ YC companies, across finance, legal and healthcare. Reported headcount is 1-10 and disclosed funding in secondary sources is small, indicating an early-stage position in the market.

The company attributes its accuracy to a proprietary pipeline rather than off-the-shelf OCR or text-only LLM parsing. Stated components are: segmentation of documents into text, tables, images and plots by specialised models; heatmap-based chunking that derives 'pivot elements' (e.g. numbers and merged cells in tables) so related content such as a table spanning multiple pages stays in one chunk; a dual-stream representation that carries a data stream (extracted content) and a layout stream (indentation, alignment, clause/sub-clause hierarchy); and a domain-specific decoder that emits a required JSON schema or Markdown, informed by finance, healthcare and legal ontologies. A reinforcement-learning pipeline can train the decoder on internal terminology, and per-item confidence scores route low-scoring extractions into later fine-tuning jobs. The company claims it consistently outperforms LlamaIndex, Gemini, Mistral and Unstructured.io on public benchmarks; no benchmark figures are given in the sources.

Technology

Proprietary dual-stream vision models used alongside OCR-based models. The pipeline segments documents into text, tables, images and plots; applies heatmap-based chunking around pivot elements to keep related content together; runs chunks through parallel data and layout streams; and decodes into JSON schema or Markdown with domain-specific ontologies for finance, healthcare and legal. Outputs include per-field confidence scores, word-level citations and bounding boxes. An in-built RL pipeline trains the decoder on internal terminology, and low-confidence items are aggregated for fine-tuning. The system supports asynchronous processing for large documents and can be deployed fully air-gapped on-premise.

Go-to-market

Developer-led and self-serve: sign-up on unsiloed.ai for an API key with no credit card required, documentation with quickstart guides, an API reference and an interactive playground, plus an MCP connector for Claude users. This is combined with enterprise sales via a demo request on the homepage, direct founder contact channels (founders@unsiloed.ai, WhatsApp/iMessage) and support@unsiloed.ai. The YC launch post and YC company page serve as distribution to other startups, with 10+ YC companies cited as users.

Enterprise and startup AI teams in accuracy-sensitive domains — finance, legal and healthcare — plus teams building RAG pipelines and document automation. Cited customer types include Fortune 150 banks, NASDAQ-listed enterprises and early-stage startups, including 10+ Y Combinator companies.

Geography

Y Combinator and two aggregator profiles list the company's location as San Francisco, and all three advertised roles are based in San Francisco. Contact details in the launch post include a US phone number.

History

Sources give conflicting founding dates: one aggregator profile lists 2024 and another lists 2025. The company was co-founded by Aman Mishra and Adnan Abbas and appears in Y Combinator's Fall 2025 (F25) batch, with YC partner Nicolas Dessaigne. Its YC launch post, 'Unsiloed AI: Make Unstructured Data LLM-Ready,' is dated 31 October 2025. As of the sources, the company is listed as active with 1-10 employees and is hiring founding backend/infrastructure, ML research and go-to-market roles in San Francisco.

Risks & controversies

Public data on the company is sparse and inconsistent. One aggregator page carries a page title claiming $1M in funding while its body repeatedly states $500,000 raised; another lists $0 raised; founding year is given as 2024 by two sources and 2025 by another. Performance claims (outperforming LlamaIndex, Gemini, Mistral and Unstructured.io; millions of pages weekly for Fortune 150 banks) come from company-authored material and are not independently verified in the sources. Named customers are described only by category, not by name.