⚡ Gevva: Instant AI Decision Engine
Think of Gevva as a fast decision engine for AI. While generative models like ChatGPT slowly type words token-by-token (taking seconds), Gevva makes single-forward-pass decisions in ~15 milliseconds on modern GPUs. It verifies facts, catches hallucinations, inspects charts, routes user requests, and grades answers.
⚡ 15 Millisecond Reflex
Non-autoregressive forward pass. No typing delay or token stutter.
👁️ Vision & Text Combined
Verifies financial charts, supply tables, and invoices alongside text.
🏆 77.54 on JevBench Public
Scored 77.54 composite score on the open JevBench public evaluation suite.
🤖 Model Engine
Choose 'gevva-e2b-multimodal' for text + images/charts, or 'gevva-e2b' for text-only.
How this works: Provide a source document or upload an image (chart, receipt, invoice, photo). Then enter a statement to test. Gevva will instantly tell you if the statement is 🟢 TRUE (Supported), 🔴 FALSE (Contradicted / Hallucinated), or 🟡 UNCLEAR (Not enough info).
🔍 Verdict: Ready for input — click an example below to try!
⏱️ Latency: --
💡 Click any real-world example to test instantly:
Examples
| 🤖 Model Engine | 1. Source Evidence / Document (Text) | 2. Statement to Verify (Is this True or False based on the evidence?) | Context Image (Optional: Chart, Invoice, Table, or Scene) |
|---|
How this works: AI agents often need to decide what to do next when a user talks to them. Instead of waiting for a slow generative model to write out thoughts, Gevva looks at the incoming request and instantly picks the best tool or action in 15 milliseconds.
🎯 Routing Result: Ready for input
⏱️ Latency: --
💡 Click any real-world scenario to test:
Examples
| 🤖 Model Engine | User Request or Customer Message | Available Tools / Actions (One action per line) | Context Image (Optional) |
|---|
How this works: Did a student (or another AI) answer a test question correctly? Paste the question, the official answer key, and the student's answer. Gevva evaluates whether the response is factually accurate or contains mistakes in 15 milliseconds.
📝 Grading Result: Ready for input
⏱️ Latency: --
💡 Click an example to test correct vs incorrect grading:
Examples
| 🤖 Model Engine | Question / Problem Prompt | Official Answer Key / Reference Rubric | Candidate Answer to Grade |
|---|
🧠 The "Gut Reflex" (System 1) vs "Deep Thought" (System 2)
Psychologist Daniel Kahneman demonstrated that human thinking operates in two modes:
- System 1 (Fast & Intuitive): Recognizing a friend's face, dodging an incoming ball, or knowing 2+2=4 instantly in a few milliseconds.
- System 2 (Slow & Deliberate): Solving complex calculus, writing an essay, or planning an itinerary step-by-step.
+------------------------------------+ +------------------------------------+
| SYSTEM 1: GEVVA | | SYSTEM 2: CHATGPT / CLAUDE |
| - Non-autoregressive decision | | - Autoregressive text generation |
| - Evaluates everything in 1 pass | | - Generates token-by-token |
| - Latency: ~15 ms (GPU) | | - Latency: 2,000 - 5,000 ms |
| - Fast single-pass evaluation | | - High token-by-token latency |
| - Use: Routing, Hallucination | | - Use: Long essays, creative text |
| Guardrails, Fact Checking | | and multi-step reasoning |
+------------------------------------+ +------------------------------------+
Most AI agent systems waste significant time and compute calling large generative models just to make simple classification decisions (like "Should I route to this tool?" or "Did the answer contradict the source document?"). Gevva solves this by evaluating decisions in a single forward pass.
📊 JevBench Public Benchmark Results
Evaluated locally against the open JevBench Public Dataset (231 evaluation tasks across 18 families):
| Model | Parameters (Download) | Forward Latency | Public Accuracy | Public Composite Score | License |
|---|---|---|---|---|---|
Gevva e2b |
5.1B (10.2 GB) | 14–19 ms (RTX 5090) | 71.43% (47.75% Hard) | 77.54 |
Apache 2.0 |
Gevva e4b |
7.94B (15.88 GB) | 17–22 ms (RTX 5090) | 76.62% (54.95% Hard) | 77.28 |
Apache 2.0 |
Transparency Note: Measured locally on the 231-item open public split of JevBench. Not an official evaluation on the full private benchmark suite until evaluated by third-party maintainers.
Technical Architecture
- Backbone Architecture: Google Gemma 4 (
gemma-4-E2B-it/gemma-4-E4B-it), 5.10B / 7.94B total parameters (including 262K vocabulary embedding table and SigLIP vision tower). - Vision Encoder: Google SigLIP tower (native multimodal vision input).
- Context Budget: Native 128,000 token Rotary Position Embeddings (RoPE), fine-tuned on sequences up to 2,048 tokens.
- Pooling & Classification: Last non-pad token pooling over backbone hidden state with linear projection head to 3 calibrated logits (0=Contradiction, 1=Entailment, 2=Neutral).
- Calibration: Shipped with standard default temperature $T=1.0$.
Python SDK Usage
Install the official package from PyPI:
pip install gevva
Run instant predictions in Python:
from gevva import load
# Load model (automatically detects CUDA, MPS, or CPU)
engine = load("davidburhans/gevva-e2b")
# 1. Fact checking / Hallucination verification
probs = engine.predict([("The sky is blue today.", "The sky is blue.")])
print(probs) # [p_contradiction, p_entailment, p_neutral]
# 2. Tool & workflow routing
best_idx, scores = engine.rerank(
premise="Please refund my credit card.",
options=["process_refund", "track_shipment", "cancel_account"]
)
print("Chosen tool:", ["process_refund", "track_shipment", "cancel_account"][best_idx])
# 3. Answer grading
grade = engine.grade(
question="What is 2+2?",
reference="4",
candidate="The answer is 4."
)
print("Is Correct:", grade.is_correct, "Score:", grade.score)
- Hugging Face Model Hub:
- Text Model:
davidburhans/gevva-e2b - Deep Reasoning:
davidburhans/gevva-e4b - Vision & Multimodal:
davidburhans/gevva-e2b-multimodal - Training Dataset:
davidburhans/gevva-decisions
- Text Model:
- GitHub Repository: https://github.com/davidburhans/gevva