Interfaze: Deterministic AI for OCR and Structured Data Extraction
Interfaze is an AI model built for tasks that require consistent, structured output: OCR, document extraction, web scraping, and data parsing. Where general-purpose LLMs are probabilistic and may format the same request differently across runs, Interfaze routes each input to task-specific encoders before passing structured data to a transformer for formatting. The result is reproducible output with per-word confidence metadata that developers can use to build fault-tolerant pipelines.
The problem with LLMs for structured extraction
LLMs are probabilistic by design. The same prompt can produce different phrasing, different structure, or different content across runs. For a document extraction pipeline that expects a specific JSON schema, an occasional introductory sentence like "Here is the JSON you requested:" before the payload breaks the parser. Temperature can be set to zero to reduce variation, but not eliminate it entirely.
For creative tasks this non-determinism is a feature. For tasks like extracting a specific financial figure, a date, or a document number, it is a reliability problem that requires extensive validation and retry logic.
Architecture
Interfaze uses a multi-stage pipeline rather than a single monolithic transformer.
Specialized encoders. When an input contains an image or document, it is routed to a Convolutional Neural Network (CNN) optimized for spatial recognition: pixel relationships, document layout, line detection, and text recognition. When input contains audio, it goes to a Deep Neural Network (DNN) stack trained for speech-to-text and speaker diarization. These encoders extract structured data (bounding boxes, text, timestamps, confidence scores) rather than passing raw data to the transformer.
Transformer orchestrator. The transformer receives structured, pre-processed output from the encoders. Its job is formatting: taking the encoder's structured data and producing the final output in the requested format (JSON, plain text, or a summary). Because it is not interpreting raw pixels or audio, it has a narrower task and makes fewer errors.
This separation means the model that "understands" images is specialized for images, and the model that "formats" output has clean, reliable input to work from.
Benchmarks
Interfaze introduced the Structured Output Benchmark (SOB), which differs from format-validity benchmarks by also checking whether the values inside a JSON object are correct, not just whether the JSON is syntactically valid. The benchmark provides the correct answer in context and measures accuracy of the extracted leaf values against ground truth.
Interfaze leads the SOB at 79.5% value accuracy. It also leads on OCRBench V2 (native OCR accuracy) and olmOCR (complex document processing). These benchmarks reflect the practical advantage of specialized encoders for perception tasks over general-purpose models.
Playground and integration
New accounts receive $20 in free credits at roughly $1.50 per million input tokens. The Playground at interfaze.ai provides a chat interface with a configuration panel on the right.
Relevant settings:
- Run Task: Selects the task type (Web Search, Scrape, OCR, Translate), routing the input to the appropriate encoder
- Temperature / Top P: For deterministic extraction, temperature near 0 reduces variation
- Code Example: The Playground generates the corresponding TypeScript or Python snippet for the current configuration, using the OpenAI SDK format for compatibility
OCR output structure
When processing a document with the OCR task, Interfaze returns a structured JSON object rather than plain text:
full_text: complete extracted text for the pageprecontext: arrays oflinesandwordswith extracted text per elementbounds: per-word and per-line pixel coordinates (top_left_x,top_left_y, and so on)average_confidence: confidence score from 0 to 1 for each recognized word and line
This metadata is what enables fault-tolerant pipeline design. Words with confidence below a threshold (for example, 0.5) can be automatically flagged for human review rather than silently passed to downstream systems.
Declassified FBI document test
Decades-old government document scans are a demanding OCR test: faded typeface, document noise, handwritten annotations, and heavy redactions. Applying Interfaze OCR to a sample of declassified FBI UFO files from the 1940s and 1950s illustrates the confidence scoring in practice.
On heavily degraded pages, many words return low confidence (below 0.4), which correctly signals unreliable extraction. On the same pages, clearly visible phrases return high confidence, showing the per-word granularity rather than a single page-level score. On pages with cursive handwritten annotations, the model attempts transcription and returns lower confidence scores than for typewritten text, consistent with the increased difficulty of the task. On cleaner typewritten pages, confidence scores are predominantly high and transcription is near-complete.
The practical takeaway for pipeline design: per-word confidence scores allow conditional routing. High-confidence extractions flow to automated processing; low-confidence regions trigger manual review or a secondary extraction attempt with a different prompt.
Final thoughts
Interfaze is most useful for extraction and parsing tasks where output reliability matters more than creative flexibility. The specialized encoder architecture, the SOB benchmark design, and the per-word confidence metadata all address the same problem: making AI-assisted extraction auditable and predictable enough to use in production without extensive retry and validation layers.
For document processing pipelines, data entry automation, or any system where a single incorrect extraction has downstream consequences, the combination of a specialized OCR model and structured confidence output is more practical than re-prompting a general-purpose LLM.
Pricing, documentation, and the Playground are at interfaze.ai.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.