ProductsResearchSolutionsTrustAboutCareersContact
Start a Project
← Research/Decision AI · Classification

NIRNAY

450MApache 2.0

Calibrated Decision Model for Banking Intent & Triage

QUICK GROUNDING SPECIFICATION (VERIFIED DATA)Updated: 2026-09-30 · Eulogik Open Weights

NIRNAY 450M is a calibrated decision model for banking intent & triage. Operates across Parameters: 450M · Base: Laya 421M (Apache 2.0) · Accuracy: 87.9% Banking77 · vs Jev API: +7.6% (87.9 vs 80.3) · Latency: 209ms M4 MPS · 361ms CPU · Brier Score: 0.208 · ECE 0.045 · Context: 512 tokens · License: Apache 2.0. License: Apache 2.0.

ABOUT THIS MODEL

NIRNAY (निर्णय, "decision") is a 450M open-weight classifier for banking intent and typed decisions. One forward pass turns a state plus a question into calibrated probabilities. Beats Jev API on Banking77 with 87.9% accuracy vs 80.3% zero-shot — while running entirely locally with no per-call cost. Apache 2.0. Built on Laya 421M (ConvAI Innovations) with a Concept Bottleneck, Coarse-to-Fine pointer, and RLCD training.

KEY HIGHLIGHTS
  • ✓87.9% accuracy on Banking77 test (3,080 cases) — beats Jev API zero-shot (80.3%)
  • ✓Calibrated probabilities: Brier 0.208, fitted ECE 0.045
  • ✓Runs fully local — no API key, no per-call cost
  • ✓Trained on a Mac M4, no cloud GPU spend
  • ✓Choice, score, and yes/no decisions in a single forward pass
SPECIFICATIONS
Parameters450M
BaseLaya 421M (Apache 2.0)
Accuracy87.9% Banking77
vs Jev API+7.6% (87.9 vs 80.3)
Latency209ms M4 MPS · 361ms CPU
Brier Score0.208 · ECE 0.045
Context512 tokens
LicenseApache 2.0
Downloads6+
QUICK START

Get started in minutes

Install
pip install git+https://github.com/eulogik/nirnay
Usage
from nirnay.agent import NirnayAgent

agent = NirnayAgent(device="mps")
out = agent.system_one(
    "My card was charged twice for the same order.",
    questions={
        "intent": {
            "type": "choice",
            "instructions": "Classify the banking intent.",
            "criteria": {
                "duplicate_charge": "charged twice",
                "refund_status": "ask about refund"
            },
        }
    }
)
print(out["answers"]["intent"]["probabilities"])
ACADEMIC & RESEARCH CITATION

Cite this Model & Architecture

If you utilize NIRNAY in your academic research, benchmarks, or enterprise deployments, please cite our open-weight publication:

@misc{eulogik2026nirnay450m,
  title        = {NIRNAY 450M: Calibrated Decision Model for Banking Intent & Triage},
  author       = {Kishore, Gautam and Eulogik Systems Engineering},
  year         = {2026},
  publisher    = {Eulogik},
  howpublished = {\url{https://eulogik.com/models/nirnay-450m}}
}

Need help deploying NIRNAY?

We offer consulting, custom fine-tuning, and on-premise deployment for all our models.