PredictRAM Vedaniq | Analyst Language Model Docs

What Vedaniq Is

PredictRAM Vedaniq is PredictRAM's proprietary small language model for grounded financial intelligence. It is designed to deliver analyst-grade research, portfolio commentary, risk interpretation, and derivatives understanding using a multi-layered stack: CPT for financial language adaptation, SFT for workflow specialization, DPO for better response quality, JEPA for predictive regime-aware context, and deterministic engines for numeric accuracy.

This architecture allows Vedaniq to provide lower-hallucination, compliance-aware, and production-oriented financial analysis across research, portfolio, risk, and derivatives workflows.

Describe It As

a proprietary PredictRAM analyst language model
an India-first grounded financial reasoning system
a structured research drafting engine
a production analyst copilot for Phase 1 and Phase 2 workflows
a multi-engine financial intelligence stack combining CPT, SFT, DPO, JEPA, and deterministic engines

How Vedaniq Works

Vedaniq combines an open-weight reasoning base, financial adaptation, PredictRAM grounding, workflow orchestration, and production serving into one practical analyst stack.

1. Open-Weight Base Model

Current practical training path in this repo uses Qwen/Qwen2.5-7B-Instruct as the open-weight base.

2. Financial Adaptation

Fine-tuning and adaptation target valuation framing, business-quality interpretation, portfolio commentary, options explanation, and PredictRAM answer structure.

3. PredictRAM Grounding Layer

Grounding sources include PredictRAM market research reports, market data, OHLC summaries, company fundamentals, and macro context.

4. Workflow Orchestration

Vedaniq is served through workflow-specific routes such as grounded stock summary, portfolio copilot, options explainer, and education copilot.

5. Production Layer

Production handling includes managed GPU inference, artifact promotion, rollback, benchmark gates, and runtime health metadata.

Key architectural claim

Vedaniq is proprietary in its adaptation, grounding, workflow design, evaluation, and serving stack, even though the base-model lineage begins from an open-weight model.

Technology Used

The current Vedaniq stack combines model adaptation, grounded financial data hydration, and production serving around the existing PredictRAM application surface.

Model And Serving

Python
PyTorch
Hugging Face Transformers
QLoRA-style adaptation workflow
Flask inference surface
Gunicorn and systemd managed serving

Infrastructure And Product Layer

AWS EC2 GPU infrastructure
PredictRAM web application integration
internal artifact registry and rollback utilities
benchmark runners for Phase 1 and Phase 2
admin and app-user Vedaniq dashboards

Training Approach

Vedaniq is best described as starting from an open-weight instruction model and then being adapted for financial analyst tasks using domain-specific training data, structured output conventions, grounded context patterns, and production benchmark validation.

Start from an open-weight instruction model
Adapt it for financial analyst workflows and PredictRAM answer style
Merge or promote the trained inference artifact
Benchmark it against Phase 1 and Phase 2 analyst workflows
Expose only validated workflows into the production frontend

Use Cases And What Vedaniq Has Been Trained For

Analyst Workflows

grounded stock summaries
analyst-style research drafting
valuation commentary
OHLC-aware price action summaries

Phase 2 Copilot Workflows

portfolio commentary
options explanation and scenario analysis
education copilot outputs
low-hallucination grounded financial Q&A

PredictRAM Vedaniq ALM Architecture

The full architecture is broader than what should be shown on a public page. These are the few layers worth revealing because they create curiosity without exposing the whole operating model.

Control Plane Identity, governance, routing, approvals, audit

Product Plane Inference, retrieval, copilots, dashboards, risk workflows

↓

Multi-Model Router Fast router, main reasoning model, quant and code specialist lanes

Hybrid Intelligence Stack Language model adaptation, predictive state layer, deterministic finance engines

↓

Portfolio Context Fabric Unified cross-asset portfolio schema and feature pipelines

Risk and Compliance Core Policy packs, explainable checks, evidence-linked rule traces

Dashboard and Algo Studio Prompt-to-dashboard, prompt-to-backtest, approval-gated deployment

1. Control Plane and Product Plane Split

Vedaniq separates governance, policy, approvals, model routing, and audit from the user-facing inference and copilot plane.

Benefit: compliance and reliability without slowing live workflows.

2. Multi-Model Inference Router

A routing layer decides whether a prompt should stay lightweight, go to the main reasoning path, or use a quant and code specialist path.

Benefit: lower latency for simple tasks and better quality for complex ones.

3. Hybrid Intelligence Stack

Vedaniq is designed so the model explains, deterministic engines calculate, and predictive layers add regime and scenario context.

Benefit: better financial accuracy with more useful narrative context.

4. Multi-Asset Portfolio Context Fabric

A shared context fabric is intended to unify exposures, collateral, liquidity, concentration, and mandate tags across asset classes.

Benefit: one portfolio copilot with consistent risk language across books.

5. Risk and Compliance Copilot Core

Policy packs, rule checks, and evidence-linked compliance traces are central to the architecture rather than added later.

Benefit: governance by design, not bolt-on governance.

6. Dashboard and Algo Studio Layer

Prompt-driven dashboard generation, backtest generation, and gated execution candidates are part of the long-term operating model.

Benefit: faster path from analyst intent to usable tooling.

Behind the scenes, the broader architecture also covers predictive analytics, latency optimization, fallback systems, security controls, and evaluation metrics. Those layers stay implicit here while the visible story stays product-led.

Vedaniq ALM

What Vedaniq ALM Is

Vedaniq ALM is the first practical shipping form of the PredictRAM SLM effort. It is the analyst assistant profile of the broader Vedaniq family.

Its role today is to act as:

an India-first analyst assistant
a grounded equity and macro reasoning model
a research drafting engine
a lower-hallucination financial Q&A surface

Features Already Implemented

Vedaniq ALM is the most concrete implemented product layer in this area. Based on the current repo state, the following are already implemented or explicitly marked complete:

Phase 1 Implemented Capabilities

grounded stock summary generation
valuation commentary
OHLC trend and price-action commentary
deterministic price query handling from PredictRAM market data instead of unnecessary GPU generation
macro context injection
risk metric grounding including CAGR, Sharpe ratio, volatility, alpha vs NSEI, VaR 95 percent, and max drawdown
Full Terminal deployment at /python-ide/full-terminal
prompt history and persistence for Vedaniq Lab usage
Phase 1 prompt validation suite

Phase 2 Implemented Capabilities

portfolio copilot mode
options explainer mode
education copilot mode
routing logic that detects portfolio, options, education, and price-style prompts
context builders for portfolio positions and options context
Phase 2 benchmark runner and model comparison harness

Product and Serving Behaviors Already Implemented

GPU and CPU inference lanes
budget-aware routing and fallback behavior
deterministic direct-answer path for price-only requests
health checks for Vedaniq services
production wiring guidance for single-host or split-host deployment
artifact registry and rollback path for inference artifacts
JEPA shadow contract endpoint in the inference service

JEPA-Related Implementation Already Present Inside Vedaniq

slm/vedaniq_jepa_shadow.py defines a JEPA shadow contract
the GPU inference service exposes /api/vedaniq/internal/jepa-shadow-contract
the contract supports regime label, confidence, nearest historical neighbors, volatility state summary, correlation state summary, stress tags, and artifact metadata

Current JEPA status

This means JEPA is not yet the primary reasoning engine, but a shadow-mode integration path already exists.

Future Plans for Vedaniq ALM

Near-Term Plan

harden Phase 1 grounded analyst workflows for controlled and early production
expand Phase 2 portfolio, options, and education workflows without breaking grounding discipline
keep deterministic engines as the numeric source of truth
maintain cost discipline through GPU caps, CPU fallback, and cached context reuse

Model Quality Plan

move from single-stage QLoRA-only adaptation to CPT plus SFT plus DPO
improve India-specific financial fluency using corpora from fundamentals, OHLC-derived text, annual reports, transcripts, and regulatory materials
improve preference quality through reviewer-ranked response pairs

Product Architecture Plan

evolve into a multi-engine system with separate reasoning, deterministic quant, predictive analytics, and governance layers
support analyst copilot, portfolio copilot, F&O copilot, and prompt-to-dashboard workflows
add risk engine outputs, options analytics, compliance checks, and predictive state context
integrate JEPA as a companion predictive-state service rather than replacing the language model stack

Governance Plan

maintain version lineage of models and datasets
keep policy gates before live strategy deployment
preserve auditable prompt, context, and output records
keep high-risk actions approval-gated

Primary Use Cases for Vedaniq ALM

grounded analyst notes for Indian equities
valuation interpretation for stocks and sectors
OHLC and price-action explanation
peer comparison and watchlist commentary
portfolio review and concentration commentary
options strategy explanation and hedging communication
macro interpretation for symbols or sectors
finance learning, certification preparation, and interview-prep assistance
internal research copilots inside PredictRAM

Talk To The Team

If you want a Vedaniq walkthrough, deployment discussion, or product access conversation, share your details and the team will receive your request in the Vedaniq admin workspace.

Request A Vedaniq Callback

Share your name, email, and phone number. Your response will be received in the Vedaniq admin dashboard for follow-up.

Share your details and our team will get back to you.