Front-page modelcourtroom
Courtroom
Three opposing intents on one shared base model — Proposer (ADVOCATE), Defender (SKEPTIC), Judge (WEIGH) — review meaningful model outputs before they reach users.
Open model card →Model registry · 36 cards · Apache 2.0 + MIT
Proprietary model. Open-licensed foundations. Owned by Nilfee.
We don't ship a closed-weight model in your inference path. Apache 2.0 and MIT only — by policy. Every model has a public card: training data, drift baseline, version history, error bounds, and the routes that call it.
9
in production
live today
2
shadow-staged
document_vlm · Qwen2.5-VL-7B
23
in development
memory pipeline · gated
0
closed-weight in path
by policy · forever
Two models that frame every answer
Courtroom and Hallucination Check.
These aren't features. They're the constitution. Every consequential answer Mukodi prints passes through both before reaching your screen.
Front-page modelhallucination-check
HallucinationCheck
Sentence-level entailment check that asserts every factual sentence in a model output is grounded against a retrieved source passage before delivery.
Open model card →License policy · written in stone
Apache 2.0 and MIT. Forever.
No Llama, GPT, or Claude in the inference path. Mukodi's FinAi model runs only on Apache 2.0 / MIT foundation weights. The foundations are open-licensed; the fine-tuning, training data, and orchestration are Nilfee's proprietary work. The one model with a tainted-weight predecessor — sentiment_analyzer — is flagged on its card with a clean replacement in flight.
Apache 2.022
MIT8
Tainted-weight1· flagged · replacement pending
Closed-weight0· by policy · forever
Apache 2.0MITOne taint, flagged
- anomaly_detector
Anomaly Detector
Real-time transaction anomaly / fraud detection — Isolation Forest + LSTM Autoencoder + Z-Score
Isolation Forest + LSTM Autoencoder + Z-Score - bill_forecaster_v1
Bill Forecaster (TimesFM 2.0)
Payer Room — quantile forecast of upcoming bill amounts + due dates (p10/p50/p90 over 30d/90d) via TimesFM 2.0 LoRA fine-tune.
TimesFM 2.0 (500M params, LoRA fine-tune, PyTorch) - bill_payment
Bill Payment
RL agent for bill-payment scheduling. Q-Learning over payment history and cash-flow state. Continuous online learning.
Q-Learning - budget_guardian
Budget Guardian
RL agent for budget-anomaly alerts. Thompson Sampling over user spending patterns. Continuous online learning.
Thompson Sampling - cashflow_forecaster_v1
Cashflow Forecaster (TimesFM 2.0)
Planner Room + financial_health_score augmentation — quantile cashflow forecast (p10/p50/p90 over 30d/90d) via TimesFM 2.0 LoRA fine-tune. Credit-path consumer.
TimesFM 2.0 (500M params, LoRA fine-tune, PyTorch) - commodity_forecaster_v1
Commodity Forecaster (TimesFM 2.0)
Public market-data forecaster — quantile commodity-price forecast for enumerated non-precious commodities (p10/p50/p90 over 30d/90d) via TimesFM 2.0 LoRA fine-tune.
TimesFM 2.0 (500M params, LoRA fine-tune, PyTorch) - courtroom
Courtroom
Three opposing intents on one shared base model — Proposer (ADVOCATE), Defender (SKEPTIC), Judge (WEIGH) — review meaningful model outputs before they reach users.
Qwen 3.6-27B (Apache 2.0) - credit_risk_scorer
Credit Risk Scorer
Probability-of-default credit risk model — stacked ensemble (XGBoost + LightGBM + LogisticRegression)
XGBoost + LightGBM + LogisticRegression stacked ensemble - document_vlm
Document VLM
Vision-language model for document understanding (OCR + extraction) — Qwen2.5-VL-7B with LoRA fine-tuning
Qwen2.5-VL-7B + LoRA - financial_health_score
Financial Health Score
Composite financial wellness score (0-100) — rule-based sub-scores + XGBoost calibration
Rule-based + XGBoost calibration - forex_forecaster_v1
Forex Forecaster (TimesFM 2.0)
Public market-data forecaster — quantile INR-pair forex rate forecast for enumerated currency pairs (p10/p50/p90 over 30d/90d) via TimesFM 2.0 LoRA fine-tune.
TimesFM 2.0 (500M params, LoRA fine-tune, PyTorch) - forex_remittance
Forex Remittance Predictor
Forex rate predictor (LSTM + XGBoost ensemble, per-pair).
LSTM + XGBoost ensemble (per-pair) - forgery_detector
Forgery Detector
PR-trio of audit roadmap.
LightGBM meta on 6 forensic layers - fraud_detection
Fraud Detection
Supervised fraud classifier (LightGBM, isotonic-calibrated). PR-trio (2026-05-02) un-deprecated the PR #14 stub by adding the missing trainer + consumer.
LightGBM (cross-model enriched) - fuel_forecaster_v1
Fuel Forecaster (TimesFM 2.0)
Public market-data forecaster — quantile fuel-price forecast for enumerated Indian cities (p10/p50/p90 over 30d/90d) via TimesFM 2.0 LoRA fine-tune.
TimesFM 2.0 (500M params, LoRA fine-tune, PyTorch) - goal_attainment_forecaster_v1
Goal Attainment Forecaster (TimesFM 2.0)
Saver Room — quantile forecast of goal attainment progress (p10/p50/p90 over 30d/90d) via TimesFM 2.0 LoRA fine-tune.
TimesFM 2.0 (500M params, LoRA fine-tune, PyTorch) - hallucination-check
HallucinationCheck
Sentence-level entailment check that asserts every factual sentence in a model output is grounded against a retrieved source passage before delivery.
DeBERTa-v3-base-NLI (MIT) - intent_classifier
Intent Classifier
Multilingual intent classifier for the multi-agent orchestrator. v0 = regex (live, fallback).
DistilmBERT v1 (regex v0 fallback) (Apache 2.0) - llm_advisor
LLM Financial Advisor
LoRA fine-tuned LLM for Indian financial advice (Qwen 3.6-27B; Apache 2.0)
Qwen 3.6-27B (default) / Qwen 3-8B (fallback) + LoRA (Apache 2.0) - loan_default_predictor
Loan Default Predictor
Loan default probability for underwriting — XGBoost + Cox Proportional Hazard + LSTM ensemble
XGBoost + Cox Proportional Hazard + LSTM ensemble - macro_finbert
Macro News Sentiment (FinBERT)
Frozen off-the-shelf finbert scoring news headlines for macro sentiment gauge.
ProsusAI/finbert (Apache 2.0) - memory_importance_scorer
Memory Importance Scorer
LightGBM regressor scoring importance of user memories at write time.
LightGBM regressor + ONNX - memory_sanity_classifier
Memory Sanity Classifier
DeBERTa-v3-base binary classifier gating memory writes.
Rule-based v0 / DeBERTa-v3-base v1+ - p2p_p2m_classifier
P2P / P2M Classifier
Peer-to-peer vs peer-to-merchant classifier. PR #14: retain (consumer exists). Followup: ml-tabular Q3 trainer or deprecate.
LightGBM (binary) - prediction_engine
Prediction Engine
Thompson Sampling RL agent for next-best-action prediction over user-behavior signals. Continuous online learning — no one-shot training run.
Thompson Sampling - rag_embedder_minilm
RAG Embedder (MiniLM multilingual)
Frozen multilingual sentence-transformer encoding chunks + queries to 384-dim vectors.
paraphrase-multilingual-MiniLM-L12-v2 (Apache 2.0) - rag_reranker_crossencoder
RAG Reranker (CrossEncoder)
Frozen cross-encoder (default BAAI/bge-reranker-v2-m3, Apache-2.0) re-scoring the top-k RAG candidates before feeding the agent.
BAAI/bge-reranker-v2-m3 (MIT) - savings_optimizer
Savings Optimizer
RL agent for savings-plan optimization. PPO-based policy learning over user financial state. Continuous online learning.
PPO - scribe
Scribe shadow
Per-turn auto-capture pipeline. Pre-classifier abstains 90 percent of the time; for the remaining 10 percent the extractor proposes 5 factual MemoryTypes bound to a strict JSON schema.
IndicBERTv2 (Apache 2.0) + Qwen 3-MoE (Apache 2.0) - scribe_signal_classifier
Scribe Signal Classifier
LightGBM binary classifier gating which conversation turns trigger Scribe LLM extraction. PR 2 ships with RuleClassifierONNX shim filling the LightGBM slot (Option E); PR 7 trains the real model.
LightGBM binary (ONNX) + RuleClassifierONNX shim fallback (MIT) - sentiment_analyzer
Sentiment Analyzer
4-class financial-chat sentiment classifier (distressed/negative/positive/neutral) serving /ai/ask via inline CPU ONNX int8.
IndicBERTv2-LoRA + ONNX int8 - spending_forecaster_v1
Spending Forecaster (TimesFM 2.0)
Coach Room + spending_predictor hybrid — quantile spending forecast (p10/p50/p90 over 30d/90d) via TimesFM 2.0 LoRA fine-tune.
TimesFM 2.0 (500M params, LoRA fine-tune, PyTorch) - spending_predictor
Spending Predictor
Time-series spending forecast (7d, 30d, 90d horizons) — XGBoost + LightGBM + ARIMA ensemble
XGBoost + LightGBM + ARIMA ensemble - stt_indicwhisper
STT (IndicWhisper Hindi)
Frozen AI4Bharat IndicWhisper for Hindi STT. Serves /voice/* endpoints.
ai4bharat/indicwhisper-hindi (Apache 2.0) - transaction_categorizer
Transaction Categorizer
Target (F1 convergence): hierarchical L1/L2 categorization via LightGBM + enrichment. CURRENTLY SERVED: in-memory scikit-learn TF-IDF + XGBoost flat ~27-category bundle (categorizer_v8), loaded from disk.
LightGBM - tts_indictts
TTS (IndicTTS)
Frozen AI4Bharat IndicTTS multilingual TTS serving the voice pipeline.
ai4bharat/indic-tts (Apache 2.0)
How to read this page
Stage and license, defined.
Stage
Production means a route is live and serving users. Shadow means the model runs in parallel without affecting decisions. In-development means it is wired but gated. Dormant means the framework ships disabled by a feature flag.
License
Every foundation model in the inference path is Apache 2.0 or MIT. Frozen-external models are loaded from upstream Apache/MIT weights and never re-trained in-house. The tainted-weights case (sentiment_analyzer) ships under a public replacement plan on its own page.