The Problem

Small and medium enterprises form the backbone of India's economy, yet accessing institutional credit remains painfully slow and opaque. When it comes to advising SME Clients, Cas have to put lot of efforts as

1. The data provided by SME Companies are manual, disorganized
2. There are chances of missing data, files and transactions
3. Each task is to be manually performed by Cas and it took lot of time to prepare financials, compute ratios and then drafting of advisory for each client.
4. Also, SME owners who doesn’t understand financials, it is difficult to explain them the business conditions which is time consuming.
5. There are high chances of human errors

There was no affordable, intelligent tool that could ingest raw, unstructured financial Excel files from SMEs — often messy, multi-sheet, and in varying formats — and produce a structured, explainable, lender-grade credit assessment in minutes.

The Solution — FinPulse AI

FinPulse AI is a full-stack, production-grade financial intelligence platform that accepts raw SME financial documents (XLSX/XLS/XLSM/CSV/PDF), automatically extracts every critical figure using a fine-tuned large language model, computes six industry-standard financial ratios, scores credit stress on a 0–100 weighted scale, generates a plain-English credit narrative, and delivers five prioritised, actionable recommendations — all in under 3 minutes.

It eliminates the need for manual data entry, formula spreadsheets, and templated Word reports. A credit analyst, CA, or relationship manager uploads files and receives a complete, bank-quality appraisal output they can act on immediately.

AI Components and Technical Architecture

FinPulse AI is built on a dual-model inference architecture, with two 7-billion parameter large language models running simultaneously across two GPUs, each assigned a distinct cognitive role.

The first model, Qwen2.5-7B-Instruct, handles all structured data extraction and narrative generation. It reads raw sheet text from P&L statements, balance sheets, cash flow statements, and loan repayment schedules and returns a precise JSON object containing over 38 financial fields — revenue from operations, EBITDA components, borrowings, trade receivables, net cash flows, principal repaid, and more. It is prompted with a domain-specific extraction schema that enforces strict field semantics: parenthesised values are treated as negative, subtotal rows are preferred over sub-item sums, and multi-column sheets are resolved to the most recent period. The same model then generates the executive summary, risk flags, and credit recommendation in plain English.

The second model, Mistral-7B-Instruct-v0.3, runs independently on a separate GPU and is dedicated entirely to generating the five strategic improvement recommendations. Separating these concerns across two models ensures that recommendation quality is not bottlenecked by extraction context and that both inference passes run in parallel where possible.

Beyond the LLMs, the platform includes a robust deterministic engine for sheet classification (profit & loss, balance sheet, cash flow, loan schedule, notes), unit detection (Lacs, Crores, Millions, Billions), period extraction via regex, JSON repair for malformed LLM outputs, and a priority-based field resolution system that always prefers the most authoritative source for each figure — for example, principal repaid from a loan schedule takes precedence over a cash flow statement figure, which in turn takes precedence over a balance sheet movement.

Financial Ratios Computed

The platform computes six ratios that collectively cover the four dimensions of SME credit health — debt serviceability, liquidity, profitability, and leverage.

Debt Service Coverage Ratio (DSCR) with a weight of 30% in the stress score. EBITDA is computed from PBT, depreciation, and finance costs, and total debt service is sourced in priority order from loan repayment schedules, STB movement in cash flow, and financing cash flow as a fallback.

Current Ratio (20%) measuring short-term liquidity, with a guard against mis-extracted figures where current assets approximate total assets.

EBITDA Margin (20%) as a measure of operating efficiency relative to total revenue.

TOL/TNW — Total Outside Liabilities to Tangible Net Worth (15%) — assessing leverage and creditor exposure.

Quick Ratio (10%) using cash, trade receivables, and short-term loans as the liquid asset base.

Debtor Days (5%) quantifying the receivables collection cycle relative to annual revenue.

Each ratio is assigned a zone (Green, Yellow, Amber, Red) with calibrated thresholds and a detailed plain-English interpretation sentence.

Stress Scoring Engine

The six ratios feed into a weighted stress score from 0 to 100. A score of 0–35 is Green (financially healthy), 36–55 is Yellow (early warning), 56–75 is Amber (moderate stress requiring intervention within 60 days), and 76–100 is Red (critical stress requiring immediate escalation). The scoring logic is fully transparent and explainable — every sub-score, weight, and zone is returned in the API response so analysts can trace exactly how the final number was reached.

Features

The platform supports multi-file upload, allowing separate P&L, balance sheet, cash flow, and loan schedule files to be submitted together and cross-reconciled automatically. Sheet-level classification means the system correctly handles workbooks with multiple tabs of mixed types.

The compare endpoint accepts up to three analysis results and produces a structured period-over-period or entity-to-entity comparison across all ratios, stress scores, and key figures — enabling trend analysis for annual reviews or portfolio screening.

The Next.js and Tailwind CSS frontend provides a clean, responsive dashboard where analysts can upload files, view the stress gauge, inspect ratio cards with zone colour coding, read the AI-generated narrative, and export findings — all without touching a spreadsheet.

Technology Stack

Backend: Python

 AI inference: Hugging Face Transformers, BitsAndBytes 4-bit NF4 quantisation, PyTorch with multi-GPU device mapping.

 Models: Qwen2.5-7B-Instruct (extraction and narrative), Mistral-7B-Instruct-v0.3 (recommendations).

Data processing: pandas, numpy, openpyxl, xlrd.

Frontend: Next.js, TypeScript, Tailwind CSS.

Deployment: Kaggle dual-T4 GPU environment, ngrok public tunnel

Relevance and Impact

For banks and NBFCs, FinPulse AI compresses a 3–14 day manual appraisal cycle to under 3 minutes. A credit operations team processing 100 cases per month could reclaim over 200 analyst-hours monthly while improving consistency and reducing human error on ratio calculations.

For CAs and financial advisors, the platform functions as an instant health-check tool — upload a client's financials before a bank meeting and walk in with a structured, ratio-backed briefing.

For SME owners themselves, the plain-English narrative and prioritised recommendations translate complex financial data into concrete actions: restructure short-term debt, improve debtor collections, infuse promoter equity. These are not generic suggestions — every recommendation is grounded in the actual extracted figures from that specific company's statements.

Key performance indicators directly impacted include: credit appraisal turnaround time (from days to minutes), analyst throughput per FTE, inter-analyst recommendation consistency, DSCR calculation accuracy (eliminating the most common manual error — incorrect principal sourcing), and early identification of SMEs in the Yellow and Amber zones before they migrate to Red.

What Makes This Different

Most existing credit assessment tools require clean, structured data input through web forms or proprietary templates. FinPulse AI works with the actual documents SMEs produce — unformatted Excel files with inconsistent headers, merged cells, multi-tab structures, varying units, and mixed-language annotations. The Rule based LLM-based extraction layer handles this ambiguity natively, the same way an experienced analyst reads a messy spreadsheet.

The dual-model architecture separates extraction accuracy from recommendation quality, ensuring neither task compromises the other. The deterministic ratio engine on top of the LLM extraction layer means the system is both flexible (handling document variety) and auditable (every figure is traceable to a source sheet and cell region).

This is not a chatbot with a financial skin. It is a structured inference pipeline purpose-built for credit operations, with explicit handling of every edge case a real-world SME appraisal surfaces — loan repayment schedules, deferred tax assets, prior-year STB movement, intangible deductions from net worth, and CFS-to-P&L cross-reconciliation.

Built for the Future of Credit

FinPulse AI is a working demonstration that AI-native credit infrastructure is not a futuristic concept — it is buildable today, on accessible hardware, with open-weight models, and it outperforms manual processes on speed, consistency, and depth of analysis. The architecture is designed to scale: additional models can be added for sector-specific benchmarking, historical trend analysis, or fraud signal detection. The API-first design means it can be embedded into any existing loan origination system, CA workflow tool, or banking CRM with minimal integration effort.