shadowApache 2.0textretail-facing

LLM Financial Advisor.

LoRA fine-tuned LLM for Indian financial advice (Qwen 3.6-27B; Apache 2.0)

Foundation · Qwen 3.6-27B (default) / Qwen 3-8B (fallback) + LoRA (Apache 2.0)

Marketing claim

Every foundation model in the inference path is Apache 2.0 or MIT — no Llama, no Claude, no GPT, no Gemini.Note ·All foundation weights Apache 2.0 or MIT; one distilled adapter (sentiment_analyzer) is undergoing replacement — see /models/sentiment_analyzer.

What it does

The honest description.

LoRA fine-tuned LLM for Indian financial advice (Qwen 3.6-27B; Apache 2.0)

Try it (mock)

Run the model on your own input.

Demo runs in your browser against a mocked edge — no signup, no production data. The same shape as the live route.

Ask Saraswati (mock)

Loading the demo…

Architecture

How it works.

User financial query with multi-language support (Hindi, Tamil, Urdu, Kannada, Bengali, Marathi, Telugu, English). Indian financial domain context + user state features. Natural-language financial advice. Trained in DPO format (chosen vs rejected response pairs) covering savings, investments, debt, budgeting tailored to Indian market.

Family

Qwen 3.6-27B (default) / Qwen 3-8B (fallback) + LoRA

Foundation

Qwen 3.6-27B (default) / Qwen 3-8B (fallback) + LoRA (Apache 2.0)

Training data & drift

What it learned from, and how we watch it.

Source

Inputs: User financial query with multi-language support (Hindi, Tamil, Urdu, Kannada, Bengali, Marathi, Telugu, English). Indian financial domain context + user state features.

Drift baseline · current quarter

Drift baseline not yet documented for this model.

What it's NOT
  • Qwen 3.6-27B (Apache 2.0; replaced Llama-4-Scout 2026-05-02).
  • PR-A2 registered LLMLoRATrainer.
  • PR-C made run() GPU-ready DORMANT behind LLM_ADVISOR_TRAINING_ENABLED (default False): flag OFF -> run() raises NotImplementedError byte-identically; flag ON on a GPU host -> real LoRA pipeline, fail-closed via ImportError->NotImplementedError if torch/peft/trl absent (NEVER fabricates metrics).
  • trainable + self_healing STAY false (PR-C is wiring, NOT activation).
  • Remaining gates: (1) real IndianFinQA corpus (Phase 8 C), (2) GPU, (3) category=llm self-healing strategy (exemption VETOED), (4) LicenseComplianceWarden Qwen-variant sign-off, (5) CT human two-step + eval-gate license pin.
  • Flip trainable+self_healing only after all five.
Who calls this model

Across the house and the API.

Internal routes

No public routes documented yet.

Version history · immutable once published

Every release, traceable.

v12026-05-11catalog YAML migration (stage: shadow)catalog-yaml-migration

Data Mukodi holds about you is subject to DPDP §12 erasure (use /consent in the app), and any concern about this model can be raised via [email protected] under DPDP §13 grievance, or via /security.

← Back to all models