Intelligent Portfolio Management with Agentic AI

A Quantrium Case Study

SNAPSHOT

Opportunity
Client
Challenges
The Quantrium Way Forward
Implementation
Impact
Technology Stack

Opportunity

Forward-thinking investment firms are now piloting agentic AI to improve risk management and portfolio decisions through real-time data analysis.

The outcome?

Optimized asset allocation, dynamic, data-driven decision-making, and enhanced portfolio performance.
Unparalleled agility to navigate evolving market conditions.

Discover how a renowned service provider harnessed these advantages with agentic AI to drive superior portfolio results.

Client

A leading licensed fundamentals-driven portfolio service provider in India, offering tailored investment management across equities, required AI agents to sharpen their investment decisions.

They partnered with Quantrium to develop a set of intelligent, autonomous source-aware micro-research agents within their decision framework by:

Scanning market data automatically.
Gathering, validating, and synthesizing vast amounts of financial and market data orchestrated around daily, quarterly, and event–driven triggers.
Expanding company and sector coverage.
Improving consistency and transparency of the research process – while keeping investment judgment firmly in human hands.
Delivering actionable insights to make timely decisions.
Reducing manual data work by 60% – 70%.

Challenges

Low-leverage work and data-heavy tasks defined the average workday, with teams juggling several challenges – tackling various reports in different formats, diverse accounting standards and languages, besides managing inconsistencies that needed multiple reviews and quality checks.

These included and were not limited to:

Manual extraction of data from financial disclosures and market data aggregators, company quarterly reports, investor presentations, analyst calls, and international broker reports.
Standardization of heterogeneous formats (different layouts for each company and even quarter-to-quarter variations for the same company).
Calculation of derived metrics (Quarter on Quarter (QoQ) / Year over Year (YoY) Growth, Compound Annual Growth Rate (CAGR), Price Earnings to Growth Ratio (PEG), Segment Margins, and more).

These, in turn, resulted in raising fresh issues like:

Misplaced Priorities

Highly skilled analysts were engaged in routine and repetitive activities, distracting them from strategic priorities.

Limited coverage

Only 2–3 companies could be deeply reviewed per day per analyst.

Outdated data risk

Daily variables (price, P/E) and quarterly variables (financials, segment information) were often dated.

Human error

Copy–paste mistakes, inconsistent formulas, missing updates.

The Quantrium Way Forward

As a Portfolio Manager, the Client’s key strength came from a deep understanding of the fundamentals.

So, the solution needed an AI-powered decision support system that combined analyst/manager competencies with an AI agent ecosystem handling large data collection and analysis.

Quantrium designed and implemented an agentic research solution that consisted of:

An Automated Equity Research Data Aggregation Tool

Automatically collected, validated, transformed, and populated all required metrics into the Client’s existing investment templates (one sheet per sector, companies as columns, metrics as rows).
Covered 170 metrics per sector across: price/market data, performance, valuations, revenue, profitability, returns, balance sheet, cash flows, holdings/management, sector-specific ratios, (Example- For the banking sector, it covered Current Account and Savings Account/ Net Interest Margin and Non-Performing Assets, and more) , forecasts, segment and geography breakdowns.
Refreshed and updated data with the right cadence:

Daily: price, market cap, basic valuations.
Quarterly: financials, segment data, geography, holdings, banking ratios.
Event-driven: analyst reports, corporate actions, new filings.

Preserved analyst control wherein the Client chose the weights and made final calls-with a system that was a research co-pilot, and not a black-box allocator.

The Agentic AI Research Fabric

Quantrium designed an agentic AI architecture that treated each metric (or a small coherent group of metrics) as an independent “micro-job” executed by a specialized agent, coordinated by a multi-agent orchestration layer.

Multi-Agent Design Principles

Compact and Well-defined Tasks Per Agent

Each agent handled a focused and precisely defined responsibility. Example – “The latest quarterly revenue and Quarter on Quarter/Year over Year Growth,” or the “Shareholding pattern and promoter pledge data.”
Short reasoning chains –Fewer hallucinations, easier guardrails, and greater reliability.

Source-aware Agents

Agents were configured with a list of sources per metric.
Example– Stock Exchange Filings, Company Reports, International Analyst Reports, Reliable Data Aggregators, and more.
Agents were aware and knew where to look first, what fallbacks were permissible, and how to cross-check when the sources disagreed.

Guardrailed, Deterministic Behaviour

Strong guardrails were deployed around:

Permissible domains & pages.
Expected units and ranges. (Example: A percentage between 0–100, holdings sum to 100%, Capital Adequacy Ratio above regulatory minimum).
No Fabricate Rules – which meant that when the data wasn’t found within the allowed sources, the agent had to return a “missing / needs manual review” flag rather than a guess.
Derived metric formulas- hard-coded or sourced from a data dictionary- not left to the model to improvise.

Metric –grouped Agents

Examples of Agent Grouping:

Price & Market Data Agent – Current Price, 52-week high/low, all-time high/low, market cap, basic ratios.
Price Performance Agent – 1M/3M/6M/1Y/3Y/5Y returns. (M-Month/Y-Year)
Valuation Metrics Agent – P/E, P/B, sector P/E, 5-year average P/E, PEG. (Price to Earnings and Price to Books Ratios, Price/Earnings to Growth).
Revenue & Profitability Agents – Quarterly and annual revenue, margins, growth metrics.
Banking Ratios Agent (For the Banking Sector) – Current Account/Savings Account, Net Interest Margin, Credit-Deposit Ratio, Provisions, Slippage, Tier-1/Tier-2 Capital, and more.
Key Product / Business / Service Analysis Agent – Discovering market share, changes to them Year on Year, or Quarter over Quarter, as well as their respective share of total revenue, revenue growth rate, and margin percentage.
Segments & Geography Agents – Segment revenue/margins, domestic vs. export, regional split, rural /urban mix.
Forecast & Analyst Consensus Agent – Analyst coverage, rating split, target prices, 3-year earnings & revenue projections.

Update-frequency-aware Scheduling

Orchestrator with a Metric-to-Frequency Map: Daily, quarterly, annual, event-driven.
Access to relevant agents when:

Daily price refresh runs were required.
Regulatory Boards announced quarterly results for a company.
A new analyst report was detected for a symbol.

The Agentic Architecture

The Trigger Layer

Schedulers (cron /workflow engine) and event listeners (Example: New exchange filings or RSS feeds) to fire jobs at the right cadence.

The Orchestration Layer (Multi-Agent Controller)

To take the trigger (Example: “Q3 results out for XXX Bank”) and:

Map it to a sector (Say, Private Banks).
Identify which metric groups need refresh (revenue, profitability, banking ratios, segment info, geography, holdings).
Dispatch sub-tasks to relevant agents for each company in that sector.
Track state and dependencies (Example: Ratios that depend on previously extracted base numbers).

The Agent Layer (Specialist Agents)

Each agent:

Receives a micro-task like “For XXX Bank, Q2 FY26, retrieve: revenue, QoQ/YoY change, 3Y and 5Y sales growth from company filings or market data aggregators and validate against at least two sources.”

Uses a constrained toolset:

Web connectors to permissible domains
PDF parsers for reports & investor decks
eXtensible Business Reporting Language (XBRL) Parsers and HTML Table Extractors

Applies validation & sanity checks (range checks, add-to-100 checks, cross-source comparisons).

Returns structured JSON with:

Metric values
Source provenance
Confidence flags/anomalies

Financial Metrics Engine

To convert raw extracted values into derived metrics using deterministic logic:

QoQ/YoY, CAGRs, EPS, PEG, RoE/RoA/RoCE, debt-equity, interest coverage, segment share, rural/urban mix, and more.

Apply sector templates (e.g., special banking metrics vs. FMCG vs. infra).

Ranking & Scoring Module

To compute within-sector ranks for each metric (Rank 1 = best performer, direction depending on metric).
To apply analyst-defined weights to produce weighted scores and an overall composite score per company, while allowing multiple “weight profiles” (e.g., growth-tilted vs yield-tilted).

Excel Output & UI Layer

Feed the output into an investment template identical to the Client’s existing template:
- One sheet per sector
- Rows = metrics; columns = company value + rank
- Preserve formulas for weighted ranks and overall scores

Optional thin UI for:
- Adjusting weights
- Viewing data-quality flags
- Triggering manual re-runs for individual stocks or sectors
- Implementation

Implementation

Phase 1 – Minimum Viable Product (Across 4 Sectors with a Single-Agent per category)

Sectors- Private Banks, Public Sector Banks, FMCGs, and Pharma Companies.

Implemented agents around market data aggregators + basic filings to:

- Fetch core price/market data, financial statements, and basic ratios.
- Populate:120+ metrics per sector.
Delivered output that matched the template, including ranking logic.
Established baseline guardrails and anomaly detection.

Outcome

Market aggregators and company reports + stock filings were adequate to start the process.
Data quality from regulated sources was effectively validated.

Phase 2 – Enhanced Source and Form Ingestion (PDFs, Segments, Geography, 10 Sectors)

Introduced agents for:
- PDF parsing of quarterly reports & investor presentations.
- Segment-level and geography-level extraction (revenue share, margins, growth).
Expanded coverage to 10 sectors and increased metric count per sector towards 170.
Added historical snapshots (at least 4 quarters) for trend analysis.

Outcome

Increased efficiency for Analysts:
- Enabled comparison of 10–12 companies per sector side-by-side, with clear segment/geography context.
- Reduced per-sector review from 2–3 hours to 30–40 minutes, providing more time for effective data interpretation.

Phase 3 – Complete Coverage (Analyst Reports, Forecasts, Alerts & Full Multi-Agent Orchestration)

Specialized agents built for:
- Parsing broker/analyst PDFs from Institutional Investors.
- Extracting 3-year revenue/EPS forecasts, rating distributions, and target prices.
Enhanced Orchestrator with:
- Event-driven triggers for new filings and reports.
- Alerts for major changes in ranking, forecast, or critical ratios (Example: a spike in NPAs)

Outcome

A dynamic research fabric system and real-time auto-refreshing of data based on the market and results calendar.
Proactive notifications to Analysts on the changes and the location of these indicators.

Impact

Extensive coverage

Options to cover up to 5+ sectors/day- reviewable at a comparable or deeper level with live views of 10-12 companies per sector.

More time for strategic priorities

Lesser time spent on routine tasks and increased focus on tasks like deeper thesis testing & scuttle criteria, cross-sector pattern recognition, and robust risk scenario analysis.

Standardized metrics and ranking logic across sectors

Reduced subjective inconsistencies while still leaving room for analyst-specific weight profiles.

Transparency

Visibility into data provenance and anomalies for improved trust.

Operational robustness

An agentic architecture made of small, guarded tasks.

Human Oversight

Agents tackled data grunt work, not judgment, and Analysts controlled interpretation and capital allocation decisions.

Increased Reliability

Short-horizon agents, explicit formulas, strong guardrails, and strict source whitelists for prioritization of accuracy and traceability over free-form generative reasoning.

Expertise

AI&ML Implementation

Software Engineering

Cases

Deep Dive

Quantrium Flux