PROJECT SHOWCASE
Vedant Andhale
AI-Native Generalist Engineer
OVERVIEW
One Thread, Three Projects
🌿
Crop Cure Bot
WhatsApp AI disease detection for farmers · 3 languages
💰
Salary Scraper
Reverse-engineer obfuscated JS · GitHub Actions pipeline
🚢
FreightSense
Two-layer AI decision engine · LLM + deterministic scoring
PROJECT 1
🌿 Crop Cure Bot
AI disease detection via WhatsApp
ResNet9
Hybrid Model
98%
Confidence Gate
3
Languages
CROP CURE · DEMO
WhatsApp In Action
Photo → Disease detection → Treatment advice in farmer's language
CROP CURE · HOW IT WORKS
Architecture
📱
WhatsApp
→
⚡
FastAPI
→
🧠
ResNet9 + Attention
→
🌍
Translate
→
💊
Treatment
Model
ResNet9 + ECA (Channel Attention) + Spatial Attention
Classes
Black Rot · Esca · Leaf Blight · Healthy
Confidence Gate
Below 98% → "Unclassified" — safety over accuracy
Deploy
Docker + uvicorn → Google Cloud Run
CROP CURE · DESIGN DECISIONS
Why These Choices
🎯 98% Threshold
Wrong diagnosis = wrong chemicals. Safety first.
📱 WhatsApp
Rural farmers already use it. Zero onboarding.
🌍 Multilingual
English · Marathi · Hindi from first interaction.
PROJECT 2
💰 Salary Scraper
Reverse-engineering obfuscated data at scale
__NUXT__
IIFE Parsing
10×
Parallel Jobs
TLS
Fingerprint Spoof
SALARY SCRAPER · OUTPUT
Dataset Profile
87,974
Total Salary Records
18
Feature Columns
0.02%
Missing Values (High Fidelity)
Scraped from AmbitionBox · Cleaned & merged via automated pipeline
SALARY SCRAPER · PIPELINE
End-to-End Flow
🔍
Discover Slugs
→
💸
Scrape Salaries
→
🧩
Parse __NUXT__
→
📊
Merge + Analyze
Challenge
AmbitionBox blurs salary data in HTML — real numbers hidden in IIFE
IIFE Parser
Brace-counting extraction + variable cross-reference resolution
Anti-Detection
curl-cffi TLS spoofing · UA rotation · random delays · backoff
Scale
GitHub Actions matrix: 10 parallel jobs × 10 companies each
PROJECT 3
🚢 FreightSense
AI-powered shipment delay intervention engine
2-Layer
AI Decision
180k
Records
5
API Endpoints
FREIGHTSENSE · EVALUATION
Risk Scoring + LLM Reasoning
Form input → Deterministic risk score → LLM recommendation → Guardrail flags
FREIGHTSENSE · HUMAN OVERRIDE
Accept · Reject · Custom
Override with reason → Full accountability chain
FREIGHTSENSE · AUDIT LOG
Every Decision Tracked
Timestamps · Layer 1 vs Layer 2 · Override status · Persistent history
FREIGHTSENSE · ARCHITECTURE
Two-Layer Decision Engine
📥
Shipment Input
→
⚙️
Layer 1: Deterministic
→
🧠
Layer 2: LLM
→
✅
Decision + Audit
Layer 1
Risk score 0-100 from delay, exposure, benchmark lookup
Layer 2
LLaMA 3.3 70B via Groq — structured JSON → intervention + reasoning
Guardrails
When layers disagree → flag for human decision
Audit
Every decision + human override logged in aiosqlite
FREIGHTSENSE · KEY INSIGHT
Why Deterministic First?
🧮 Explainable
Risk score formula is auditable. No black box.
🛡️ Guardrails
LLM adds nuance, never overrides logic.
👤 Human Final Say
Override with reason. Full accountability chain.
THANK YOU
Vedant × Agntworks
I see ambiguity and move toward it.
3
Project
E2E
Ownership
AI-First
Mindset