NAVAM

Client Overview

Meras Water Solutions is a California-based agricultural water treatment company specializing in chemical treatments, equipment maintenance and system monitoring for farms across California and Nevada. Their field teams serve thousands of grower accounts and farm sites, managing water quality, treatment schedules and agronomic recommendations across diverse crops and growing conditions.

Technologies Used

Salesforce (LWC, Apex, Custom Metadata)
Claude Sonnet (AI recommendations)
Claude Haiku (text correction)
OpenAI text-embedding-3-small
PostgreSQL + pgvector
Python (migration pipeline)
Microsoft Graph API / SharePoint
Open-Meteo Weather API
pdf-lib.js

Challenge

Meras field staff submitted paper-based lab requests to Precision Agri-Lab (PAL), received PDF results by email and manually transcribed values into spreadsheets. There was no structured history of lab results, no automated matching of results to farm sites and no pathway to data-driven agronomic recommendations. The workflow was error-prone, slow and produced no analyzable data foundation for AI-driven decision support.

Over 2,200 historical water report PDFs sat in SharePoint with no structured data extracted, representing years of valuable agronomic intelligence that was inaccessible to field teams.

Solution

NAVAM Digital designed and delivered a four-phase, Salesforce-native platform digitizing the full lifecycle of agricultural water lab testing — from field submission through results processing into a predictive AI recommendation layer.

Phase 1 — Sample Submission: A mobile-friendly Lightning Web Component form captures sample data, signatures and site details, generates a signed pre-filled PDF lab request using pdf-lib.js and creates structured Salesforce records. Quick Actions on Account, Site and Lead pages launch the form, with unique tracking prefixes stamped on every submission for downstream matching.
Phase 2 — Automated Lab Results Processing: An inbound email service ingests forwarded PAL result notifications, downloads and parses CSV files via a metadata-driven parser and runs each row through a 7-priority matching cascade to link results to the correct submission, site and account. A Match Review dashboard lets admins resolve low-confidence matches manually.
Historical Migration: A Python pipeline extracted data from 2,222 legacy water report PDFs in SharePoint via Microsoft Graph API. A two-pass Claude API extraction parsed each PDF, matched results to existing accounts and sites and populated Salesforce with the full historical dataset.
Phase 3 — Data Sync and Vector Database: A PostgreSQL and pgvector schema synced from Salesforce — 2,303 accounts, 5,901 sites, 5,132 site components, 180,923 service reports and approximately 274,000 weather records spanning 10 years across 75 California and Nevada counties from Open-Meteo. OpenAI embeddings power semantic similarity search across all records.
Phase 4 — MERA AI Assistant: A conversational Salesforce LWC backed by Claude Sonnet. MERA (Meras Expert Resource Assistant) combines real-time Salesforce entity data with pgvector semantic search across service reports and lab results. Features include 71 categorized questions across 10 categories, 5-turn conversation history, weather-aware recommendations, a 21-product chemical knowledge base extracted from Meras product PDFs, voice input via Web Speech API and AI text correction via Claude Haiku that preserves agricultural terminology.

Results

All four phases delivered end-to-end — from field submission through AI recommendations in a single Salesforce-native platform
2,222 historical PDFs migrated from SharePoint into structured Salesforce records via automated Claude API extraction
180,923 service report records loaded into the AI data foundation alongside 5,901 sites, 2,303 accounts and 5,132 site components
Approximately 274,000 weather records spanning 10 years across 75 counties pre-populated for AI context
21 Meras products with full AI knowledge base — composition, dosing and applications — injected into every recommendation call
71 AI-ready questions across 10 categories deployed covering water quality, crops and soil, treatment, trends, weather and service history
Eliminated manual data entry from the legacy paper-based submission and PDF transcription workflow entirely
Architected for future fine-tuned self-hosted LLM (Llama 3 / Mistral) — provider switch requires only a Custom Metadata record change with no code deployment

Architecture Highlights

Provider-agnostic LLM design: Custom Metadata enable switching between Claude, OpenAI or a self-hosted model with no code deployment
Hybrid retrieval: Combines real-time Salesforce entity lookups with pgvector semantic search for contextually accurate, site-specific recommendations
Weather-aware AI: Every recommendation call includes current and forecast weather data for the relevant county, enabling timing-aware agronomic guidance
Voice-enabled interface: Web Speech API integration with Claude Haiku text correction preserving agricultural terms (EC, SAR, PPM, ORP, pH, GPM, TDS)