Skip to content

ALGO.md

Latest commit

 

History

History
52 lines (43 loc) · 3.36 KB

ALGO.md

File metadata and controls

52 lines (43 loc) · 3.36 KB
Raw

🧠 Gramin Saathi - Technical Algorithms & Logic

This document details the core technical implementations and mathematical models used in Gramin Saathi.

1. 🤖 AI Context Window & NLP

Condensed Context Strategy

To maintain high performance and stay within token limits of Gemini 2.5 Flash, we use a "Condensed Profile Context" instead of sending the full database.

  • Context Injection: User metadata (state, crop, land size) is minimized into a single-line string injected into the system instruction.
  • Sliding History: Only the last 5 messages are sent to the API, preventing "context drift" and reducing cost per request.
  • Intent Detection: Before calling the LLM, a local regex/keyword-based NLP engine (detectIntent) classifies the query to decide if it can be handled by a faster, deterministic Local Algorithm (88% accuracy).

2. ⚖️ Matching Algorithms

A. Scheme Eligibility (Scoring Model)

We use a Weighted Multi-Factor Scoring System (Max 100 points):

  • State Match (25 pts): Binary match or "National" status.
  • Landholding (25 pts): Uses a Proximity Bonus formula. If the user is in the "sweet spot" (mid-range), they get +5 bonus points.
  • Income (25 pts): Logarithmic priority scoring for marginal farmers.
  • Crop Fit (20 pts): Calculates matching ratio.
  • Category (10 pts): Handles specific demographics (Young vs Senior).

B. Loan Affordability Engine

Uses a Debt-to-Income (DTI) Ratio model:

  • EMI Formula: $P \times r \times \frac{(1+r)^n}{(1+r)^n - 1}$
  • Logic: Any loan where the EMI exceeds 60% of monthly income is automatically disqualified to prevent debt traps.
  • Ranking: Sorts by $(Match Score \times 0.6) + (Interest Rate Inverse \times 0.4)$.

3. 🔐 Security & Privacy

End-to-End Encryption (E2EE)

Financial data in the Khata is encrypted BEFORE it hits the disk or cloud.

  • Algorithm: AES-GCM 256-bit (authenticated encryption).
  • Key Derivation: PBKDF2 with 100,000 iterations using a per-user salt.
  • Zero-Knowledge: The server never sees the raw transaction descriptions or amounts; only the client holds the derivation key.

Banking-Grade IDs

Transaction IDs are generated using a cryptographically secure random string: GS-XXXX-XXXX-XXXX. This ensures non-sequential, non-guessable IDs for accounting integrity.

4. 📈 Yield Prediction Model

A multi-variant deterministic model simulating agro-climatic conditions: $$Yield = B \times (S_{factor} \times W_{factor} \times F_{factor})$$

  • Soil Factor ($S$): Weighting based on pH stability and moisture retention.
  • Weather Factor ($W$): Rainfall deviation from the crop's ideal mean.
  • Fertilizer Factor ($F$): Efficiency gain based on soil pH compatibility.

5. 🔄 Offline State Machine

  • Storage: IndexedDB via Dexie.js.
  • Sync Strategy: "Outbox" pattern. Transactions are written to a local pending_sync table and a Service Worker attempts to flush them when navigator.onLine is true.
  • Conflict Resolution: Last-Write-Wins (LWW) based on high-resolution timestamps.

6. 🗣️ Voice Processor

Uses the Web Speech API for Synthesis (TTS) and Recognition (STT).

  • Cleaning Logic: TTS input is pre-processed to remove Markdown symbols and extra spaces to ensure natural cadence.
  • Navigation Maps: Voice commands are mapped to internal React routes using fuzzy matching.