Unsiloed AI 籌集 $7.5M,讓企業文件適用於 AI
Unsiloed AI Raises $7.5M to Make Docs AI‑Ready
Unsiloed AI 籌集 $7.5M 種子輪融資,開發處理企業文件、供 LLM 使用的技術。公司稱,其系統在涵蓋 542 份金融文件的專有測試集上達到 99.78% 準確率,在 olm OCR-Bench 上得分 88.0。服務採按頁收費,每頁 1 美分起。
Funding Overview and Market Context
In the fall of 2025, Unsiloed AI closed a $7.5 million seed round, with Felicis Ventures acting as lead investor. The round also featured participation from Y Combinator, Orange Collective, Uncorrelated Ventures, and notable angels such as Aarthi Ramamurthy and Beth Turner. For a seven‑person startup, the capital injection is a decisive signal that the market is hungry for reliable, enterprise‑grade document‑to‑LLM pipelines.
The broader AI financing landscape has been volatile. Recent headlines—like Nvidia’s tightened AI financing rules after its CEO’s exit—highlight investors’ growing caution around speculative AI bets. Yet, targeted funding for niche, high‑impact use cases remains robust, as illustrated by Yvonne Bajela’s $17 M fund for EU AI startups. Unsiloed AI’s raise slots neatly into this selective funding trend, positioning the company as a specialist rather than a broad‑stroke AI play.
Technical Architecture of Unsiloed AI
Unsiloed AI’s core product tackles the “silo” problem: complex enterprise documents (financial statements, legal contracts, technical manuals) that contain tables, charts, and non‑linear layouts. Traditional large language models (LLMs) treat such inputs as flat text, leading to hallucinations and misinterpretations—a risk underscored by CEO Aman Mishra:
“The problem with LLMs is they’re inherently probabilistic. There is going to be some deviation induced. There is going to be some hallucination involved, and that has a massive negative impact downstream.”
To mitigate this, Unsiloed AI employs a multi‑layered pipeline:
Dual‑Stream Vision Models
- Text Stream: Tokenizes raw OCR output, preserving linguistic context.
- Image Stream: Generates image tokens for visual elements (charts, diagrams).
- This separation allows the system to “see” a chart as a visual object while still understanding surrounding narrative text. Mishra notes, “In one of the streams you have the text tokens, in the other stream you have the exact image tokens. This is something that we have spent a lot of time architecting.”
Heat Map‑Based Chunking
Unsiloed AI runs a relevance heat map across each page, flagging high‑value regions (e.g., a balance‑sheet table) for deeper processing. Low‑importance sections are down‑sampled, conserving compute and reducing latency.
Specialized Decoder
A domain‑aware decoder maps extracted entities to industry‑specific vocabularies. For finance, it distinguishes “EBITDA” from “EBIT”; for law, it respects “indemnification” clauses. This decoder dramatically cuts the false‑positive rate that plagues generic OCR pipelines.
Traceability Layer
Every datum returned to the downstream LLM is annotated with its source coordinates (page, bounding box). This provenance is crucial for audit‑heavy sectors such as banking and legal services.
Accuracy Benchmarks and Competitive Landscape
Unsiloed AI reports 99.78 % accuracy on a proprietary test set of 542 financial documents. On the public olm OCR‑Bench benchmark, it scores 88.0, eclipsing Reducto (66.0) and Unstructured (39.9). These numbers are not merely academic; they translate into tangible risk reduction for downstream AI agents that rely on extracted facts.
Competitor Snapshot
🔹 --------------------
• Benchmark Score (olm OCR‑Bench): --------------------------------
• Notable Strength: ------------------
🔹 *Unsiloed AI*
• Benchmark Score (olm OCR‑Bench): 88.0
• Notable Strength: Dual‑stream vision, traceability
🔹 Reducto
• Benchmark Score (olm OCR‑Bench): 66.0
• Notable Strength: Fast batch processing
🔹 Unstructured
• Benchmark Score (olm OCR‑Bench): 39.9
• Notable Strength: Broad API ecosystem
🔹 Microsoft Azure Form Recognizer
• Benchmark Score (olm OCR‑Bench): —
• Notable Strength: Deep integration with Azure services
🔹 Amazon Textract
• Benchmark Score (olm OCR‑Bench): —
• Notable Strength: Scalable cloud‑native service
While Microsoft and Amazon provide massive scale, they lack the fine‑grained traceability and industry‑specific decoding that Unsiloed AI emphasizes. This differentiation is especially relevant for regulated clients.
Business Model, Pricing, and Early Adoption
Unsiloed AI adopts a pay‑per‑page model starting at one cent per page. The low entry price lowers the barrier for pilot projects, while the pricing granularity aligns with enterprise budgeting cycles.
Early Customers
- A major bank – uses Unsiloed AI to ingest quarterly earnings reports, feeding verified figures into risk‑assessment LLMs.
- A leading law firm – extracts clause metadata from contracts, enabling rapid compliance checks.
- Bolna AI – a voice‑AI startup that leverages structured data to generate spoken summaries of financial briefs.
The seed round will fund three primary initiatives: hiring additional researchers and engineers, expanding the sales organization, and building proprietary, domain‑specific training datasets. With a team of seven, the company is poised to scale its engineering capacity without diluting its focus.
Industry Impact and Future Outlook
Why It Matters
Enterprise AI adoption hinges on data quality. When LLMs ingest noisy or mis‑parsed documents, downstream decisions—credit approvals, legal risk assessments, strategic recommendations—can be compromised.
Read the full breakdown originally published at https://ltdeveloperblogs.github.io/posts/felicis-leads-75m-bet-on-unsiloed-ai-to-make-enterprise-documents-usable-by-ai-agents/
來源:DEV Community · dev.to