Interfaze wants an AI model that stops guessing

Ask a general-purpose language model to read a scanned invoice and it will hand you an answer with total composure - and, some of the time, total fiction. It might invent a line item, drift a decimal, or confidently transcribe a number that was never on the page. For a chatbot, that is a rounding error. For a developer wiring that output into an accounting system, it is a support ticket, or worse. Interfaze is a company built on that gap.

Based in San Francisco and part of Y Combinator’s 2026 batch, Interfaze makes an AI model aimed squarely at what it calls deterministic tasks - the backend jobs that need the same answer every time and a way to check it. Optical character recognition on a messy PDF. Pulling structured data off a web page. Turning an hour of audio into a transcript with speaker labels. Classifying an image. These are not glamorous problems. They are the plumbing of most real software, and they are exactly where the current generation of models tends to leak.

01 / THE PITCHReliability as a feature you can bill for

The core idea is almost stubbornly practical. Instead of asking one enormous model to be brilliant at everything, Interfaze leans on task-specific small models - proven specialists for OCR, object detection and speech recognition - and stitches them into a transformer. The company’s framing, published in a paper accepted at the IEEE Conference on Artificial Intelligence in 2026, is that the future of applied AI is built on those small, sharp models rather than ever-larger generalists.

What developers actually get back is the interesting part. Interfaze returns confidence scores, bounding boxes and metadata alongside its answers. That means a team can write ordinary rules - flag anything below 90% confidence, route low-certainty scans to a human - and trust that the numbers underneath are real. The model is doing the reading; the developer keeps the judgment.

02 / WHO IT IS FORThe people who ship the boring middle of the stack

Interfaze’s customer is a developer, not an end user. It sells to the engineers building document-processing pipelines, extraction tools, transcription features and scrapers - the B2B and SaaS teams for whom “the AI was mostly right” is not a shippable state. The predecessor product, JigsawStack, gives some sense of the appetite: since its 2024 beta it processed more than ten million API requests, and Interfaze’s own site points to early customers such as Aegis and RMH.

The problems it takes on are the ones that quietly consume engineering time. Every OCR integration eventually meets a document it cannot parse. Every scraper breaks when a site changes its markup. Every transcription job stumbles over crosstalk and accents. Interfaze’s bet is that teams would rather pay per token for a model that handles these cases - and reports how sure it is - than maintain a fragile stack of their own.

03 / HOW IT WORKSSpecialists fused into a transformer

Under the hood, Interfaze describes a native hybrid architecture: deep neural networks - the convolutional and detection models that computer vision has trusted for years, plus speech models - are fused directly into a transformer decoder through a shared embedding space. Rather than calling those specialists as opaque outside functions, the model can see their outputs and metadata as part of its own reasoning. For a document, a parser splits the file into pages and blocks, OCR and layout models turn each page into structured tokens, and a small retriever surfaces only the segments a question actually needs.

The company has been candid that this is early. It calls the current release a beta prototype and has flagged the honest trade-offs of the approach - token costs from repeated context, slower inference than a bare specialist model, and the hallucination risk any transformer carries. A production version is the stated next step. That willingness to publish the limitations, rather than only the benchmark wins, is itself a tell about how the team works.

04 / THE DIFFERENCEWhy not just use GPT or Gemini?

It is the obvious question, and Interfaze answers it two ways. The first is accuracy on the narrow tasks it targets: the team reports its prototype outperformed a base Gemini model on vision-heavy work and exceeded Claude Opus 4 on several benchmarks. Benchmarks are benchmarks, and worth reading with care, but the direction of the claim is the point - beat the generalists where specialization matters.

The second answer is about verifiability. General models will give you a clean paragraph; they will not readily give you a calibrated confidence number and a bounding box you can put a rule against. Interfaze layers in configurable guardrails too - NSFW filtering and fourteen safety categories covering violence, hate, privacy and more - which is the sort of control that matters when you are pushing model output into a regulated workflow rather than a group chat.

05 / DISTRIBUTIONThe quiet trick of being OpenAI-compatible

Interfaze’s smartest move may be the least technical. Because the model speaks the OpenAI chat API, a developer can try it by changing a base URL and a key - no rewrite, no migration, no lock-in to overcome first. It works with the OpenAI, Vercel AI and LangChain SDKs across TypeScript and Python. That lowers the cost of trying a specialist model to almost nothing, which is how a small team gets a foot in the door of stacks already built around the incumbents.

The tooling is bundled too. Interfaze ships a headless browser and proxy scraper, web search, a secure code sandbox and retrieval as part of the model’s reach - the glue code teams usually assemble and babysit themselves. Priced at $1.50 per million input tokens and $3.50 per million output, with caching included, it is a usage-based business rather than a seat-based one: you pay for the work the model does.

06 / THE TEAMFrom JigsawStack to a rebuild at the model layer

Interfaze is led by Yoeven D. Khemlani, its CEO and co-founder, alongside co-founder Harsha Vardhan Khurdula. Khemlani is a second-time founder - previously CTO and co-founder of the Southeast Asian prop-tech startup Stayr/Staytion - who moved into AI infrastructure with JigsawStack, a suite of small-model APIs for backend developer tasks. Khurdula brings years of computer-vision and reinforcement-learning research and several peer-reviewed papers.

The lineage matters. Interfaze is not a fresh idea grafted onto a pitch deck; it is the same team going one layer deeper. Having run a developer-API company to ten million requests, they concluded the real constraint was the models themselves - and rebuilt from the architecture up. The company backs the thesis with research rather than only marketing, publishing its design at IEEE CAI 2026. To date it has raised about $1.5 million across pre-seed and seed rounds, with Antler and Ada Ventures among the backers.

07 / WHERE IT FITSA wedge, not a frontier race

Interfaze is not trying to out-scale the frontier labs, and that is the whole strategy. It sits in the practical middle of the market - between general LLMs pressed into service for extraction, the document-AI and OCR platforms, and scraping tools like Firecrawl and ScrapeGraphAI. Its wedge is a single, tool-equipped API that treats reliability and verifiable output as the product, not an afterthought.

Whether that wedge widens depends on the production model living up to the beta’s promise, and on developers deciding that a confidence score is worth switching for. But the framing is clear-eyed. In a market loud with claims of general intelligence, Interfaze is making a narrower, more testable one: that the software everyone actually ships runs on boring, reliable tasks - and those deserve a model of their own.