3 Startups · $7.1M+ · AI + Prediction Markets

What they're building

Every LLM application is a context problem. The dominant approach today is to pick the best monolithic transformer and throw your inputs at it. Interfaze's argument is that this is wasteful and often wrong: tasks like OCR on complex PDFs, web scraping from dynamic pages, multilingual speech recognition, and text classification require deterministic, high-accuracy outputs that general-purpose LLMs systematically underdeliver. Interfaze is built on a different architecture: a stack of heterogeneous specialised DNNs and small language models that handle perception and extraction, paired with a context-construction layer that crawls, indexes, and parses external sources into compact structured state, and an action layer that can browse, retrieve, execute code, and drive headless browsers. A thin controller sits on top and exposes a single OpenAI-compatible endpoint, deciding which specialised model to call and forwarding distilled context to a user-selected LLM for the final response. The benchmarks are notable: Interfaze-Beta hits 83.6% on MMLU-Pro, 91.4% on MMLU, 81.3% on GPQA-Diamond, and 90.0% on AIME-2025, with strong multimodal scores across ChartQA (90.9%) and AI2D (91.5%). The paper was accepted at IEEE CAI 2026.

Why this matters

The vibe of the current AI market is that bigger, more general models win everything. Interfaze's technical bet runs the opposite direction: for the tasks that actually show up in production, the tasks that companies pay for and that must be right 99% of the time, a specialised small model ensemble beats a monolith on cost, accuracy, and consistency. This matters to developers building real systems. A healthcare company using OCR to extract structured data from medical records cannot run on 90% accuracy. A government system scraping regulatory documents needs reproducible outputs. JigsawStack, the company's previous name under which they built the developer tooling platform, already had paying enterprise customers before YC. The pivot to Interfaze is a doubling-down on the core technical insight from that work: the bottleneck in AI applications is not generation, it is reliable extraction and perception. YC P26 is widely considered one of the strongest batches in the accelerator's history, which adds institutional validation.