Applied AI Researcher & Systems Engineer
Autonomous Multi-Agent Systems • Multimodal Remote Sensing • Verifiable & Physics-Informed Machine Learning
West Bengal, India • arpanghosh984@gmail.com

I am an Applied AI Researcher working at the critical intersection of autonomous multi-agent orchestration, multimodal earth observation, and verifiable, physics-informed machine learning for agroecosystem health, voluntary carbon-credit accounting, and agricultural supply-chain risk modeling.

My academic perspective unites formal postgraduate research in Agribusiness Management and Digital Market Economics with undergraduate laboratory training in Biotechnology (cellular kinetics and molecular assays) and production systems engineering across 20+ deployed AI prototypes and 6 client systems.

Rather than relying on unconstrained generative heuristics, my research emphasizes:

1. Physics-Informed Constraints: Embedding closed-form biochemical and environmental equations (such as soil $Q_{10}$ kinetic equations) into deep learning networks.
2. Deterministic Safety & Auditability: Enforcing strict schema validation, dual-evidence regulatory citation layers, and cryptographic agent authorization (HMAC-SHA256).
3. Open-Source Reproducibility & Honesty: Explicitly categorizing algorithms into runnable baselines, simplified approximations, and open future research directions.

Doctoral Objective: Seeking Ph.D. positions and cross-disciplinary research partnerships in AI in Agriculture, Digital Agriculture, Intelligent Systems, and Environmental Informatics (2026/2027 academic intake).

Savitribai Phule Pune University (Lotus Business School, Pune, India) | 2018 – 2020

• Academic Evaluation: Final Grade A (63.33%) • Medium of Instruction: English (Full 2-Year Program).
• Postgraduate Dissertation: A Study on Digitalisation of Agri-Market in India
  • Investigated structural market distortions, electronic spot exchanges (NSEL, e-NAM, e-Choupal), smallholder information asymmetry, and credit-risk mitigation through digital linkages.

University of Burdwan (St. Xavier's College, Burdwan, India) | 2015 – 2018

• Academic Evaluation: Second Class Honours (53.67%) • Medium of Instruction: English (Full 3-Year Program).
• Laboratory Thesis: Quantitative molecular assays, cellular kinetics, and biochemical pathways.

My proposed doctoral inquiries address the structural challenges of agricultural monitoring and financial inclusion under severe data-scarcity, monsoon cloud occlusions, and infrastructure constraints. This research builds directly on three open-source, deployed systems:

A Multi-Agent Architecture Interfacing Remote Sensing with In-Situ Vision & Physics-Informed Kinetics

• The Challenge: Traditional soil organic carbon (SOC) verification relies on high-transaction-cost physical soil coring, while optical satellite imagery is blinded by weeks of monsoon cloud cover over fragmented, sub-acre smallholder plots.
• The Core Approach & Deployed Implementations:
  • Macro-Scale Earth Observation via AgriSentinel:
    • Constrained Non-Negative Least Squares (NNLS) spectral unmixing disaggregating intercropped micro-plots against vegetation endmembers.
    • A Pix2Pix U-Net Generative Adversarial Network translating Sentinel-1 Synthetic Aperture Radar (SAR) backscatter into synthetic optical surface parameters to bypass persistent cloud cover.
    • Physics-Informed Neural Networks (PINNs) embedding closed-form soil $Q_{10}$ turnover kinetic equations with remote-sensing temperature/moisture proxies to model SOC trajectories without recurring core sampling.
  • In-Situ Ecological Ground-Truthing via Leyloom:
    • Computer vision classifier evaluating quadrat field photography of species-rich agricultural leys.
    • Categorizes grass, legume, forb, and senescent fractions; continuously logs the Shannon-Wiener ecological diversity index ($H'$) per plot and timestamp with an open validation protocol calibrated against manual reference quadrats.
  • Telemetry Pipeline Robustness via Betulog:
    • Semantic anomaly detection engine for distributed agricultural IoT telemetry and soil sensor logs.
    • Ingests unstructured system telemetry into dense vector embeddings, identifying operational drift and sensor anomalies before corrupted inputs propagate into downstream agronomic models.

From Satellite-Based Farm Monitoring to Alternative-Data Credit Access

• The Challenge: Smallholder farmers and rural agri-MSMEs are systematically excluded from formal credit markets due to lack of collateral or audited accounts, while standard remote sensing fails on complex intercropped land.
• The Core Approach & Deployed Implementations:
  • Metacognitive Multi-Agent Orchestration: Dynamic, uncertainty-aware routing that balances compute latency, API costs, and sensor confidence to route queries across specialized sub-agents and degrade gracefully under low-confidence data.
  • Agronomic-Financial Data Fusion via UdyamVitals / Avsar:
    • Fusing rolling $z$-score crop stress anomalies (evapotranspiration and land surface temperature) with transactional cash-flow signals (GST, UPI, Account Aggregator).
    • Unlocks intent-based underwriting for credit-invisible agri-enterprises, reaching a 31% qualified-lead conversion rate in prototype trials versus a <3% industry baseline (IDBI Innovate 2026 / AWS).

System & Artifact	Research Domain & Core Innovation	Architecture & Verification	Code & Deployment
AgriSentinel	Autonomous Remote-Sensing Fleet Converts heterogeneous satellite/sensor streams into verifiable agronomic & carbon indicators.	5-agent Gemini orchestration • Pix2Pix U-Net SAR-to-Optical • PINN $Q_{10}$ soil carbon turnover • NNLS spectral unmixing.
Leyloom	In-Situ Quadrat Botany CV Classifier Estimates species diversity and legume canopy ratios from in-field vegetation photography.	In-situ quadrat CV classification • Shannon-Wiener ecological diversity ($H'$) • Open bias calibration protocol.
Betulog	Semantic Anomaly Engine for AgTech IoT Ingests unstructured system and telemetry logs to detect operational drift before dashboard failure.	Dense semantic embedding projections • Hierarchical log parsing • Distributed telemetry discrepancy flagging.
Pehra	Deterministic Statutory Law Auditor Contract auditing engine under a zero-hallucination guarantee against an immutable statutory corpus.	Server-side schema constraints • Zero data retention • 72 automated integration tests • WCAG 2.1 AA certified.
ClauseTrace	Dual-Evidence Regulatory & AML Copilot Enterprise financial crime intelligence binding every output to raw data and statutory clauses.	Snowflake native execution • Dual-evidence chain-of-custody • Independent interceptor Guardrail Agent.
BuddyPass	Cryptographic Agentic Authorization Protocol HMAC-SHA256 protocol enabling humans to delegate mathematically bounded spending limits.	Cryptographic signed credentials • Independent service verification • Autonomous agent boundary control.
StadiumPulse	Safety-Critical Crowd Operations Copilot Real-time egress congestion and density threshold monitoring system for high-throughput venues.	Deterministic safety rules engine • Generative 6-language directive synthesis • 23 unit tests • WCAG AA.
Drishti & Avsar	Contagion Graphs & Alternative Underwriting Supply-chain risk propagation networks and cash-flow credit scoring for credit-invisible enterprises.	GST / e-way bill transaction graph • 31% qualified-lead conversion versus <3% baseline (IDBI Innovate 2026).

                                  ARPAN GHOSH: APPLIED AI RESEARCH LAB
  ┌──────────────────────────────┬──────────────────────────────┬──────────────────────────────┐
  │     MULTI-AGENT SYSTEMS      │    EARTH OBSERVATION & CV    │  PHYSICS-INFORMED & SAFE AI  │
  ├──────────────────────────────┼──────────────────────────────┼──────────────────────────────┤
  │ • Metacognitive Orchestration│ • SAR-to-Optical Pix2Pix     │ • PINNs (Soil Q10 Kinetics)  │
  │ • Uncertainty-Aware Routing  │ • NNLS Spectral Unmixing     │ • Deterministic Rule Engines │
  │ • Multi-Agent Consensus      │ • In-Situ Quadrat Botany     │ • Zero-Hallucination Schemas │
  │ • Tool Calling & Protocol    │ • Shannon Diversity Index H' │ • Dual-Evidence Provenance   │
  │ • Gemini, Claude, Vertex AI  │ • ET / LST Z-Score Detection │ • HMAC-SHA256 Authorization  │
  └──────────────────────────────┴──────────────────────────────┴──────────────────────────────┘

• Autonomous Agent Architectures: Metacognitive orchestration, dynamic dispatch under latency/cost bounds, tool execution pipelines, Model Context Protocol (MCP), structured JSON schemas.
• Remote Sensing & Spatial Data Science: Synthetic Aperture Radar (Sentinel-1 SAR), multispectral imagery (Sentinel-2 Optical), non-negative least-squares spectral unmixing, NDVI/LST time-series anomaly detection.
• Physics-Informed & Formal Verification: Physics-Informed Neural Networks (PINNs), $Q_{10}$ soil organic matter kinetics, server-side statutory citation validation, HMAC-SHA256 signed delegation credentials, privacy-preserving federated meshes.
• Full-Stack Engineering & Cloud: Python, TypeScript, PyTorch, Scikit-learn, OpenCV, Google Cloud Platform (Vertex AI, Cloud Run), AWS, Snowflake, ClickHouse, Docker, CI/CD automated test suites (70+ unit/integration tests), WCAG 2.1 AA accessibility standards.

• Google Cloud GenAI Academy — APAC Cohort 3 | Google Cloud & Hack2skill (2026)
  Verification ID: 2026H2S09GCGENAIAPACC3-P01382
  Enterprise foundation model orchestration, multi-agent frameworks, and Vertex AI production deployments.
• PromptWars Virtual Challenge 4 — GenAI Engineering | Hack2skill (2026)
  Verification ID: 2026H2S07PWVCHL4-AT00046
  Constraint satisfaction, structured output guarantees, and complex multi-step reasoning architectures.
• 2nd NextGen Hackathon 2026 | Soft Computing Research Society (SCRS) & ACM Fremont Chapter
  Evaluated on soft-computing paradigms and machine learning applied to socio-technical systems.
• Unstop National Benchmark Assessment & Prompt Engineering | Unstop & Alison (2026)
  Verification ID: 70b1ea68-c08a-4cd2-b0a6-805c33ff6bbb
  Standardized evaluation in algorithmic problem solving and generative AI systems engineering.

I welcome correspondence regarding Ph.D. research assistantships, lab rotations, and applied AI collaborations:

• 🌐 Interactive Portfolio Dossier: korvanta-ai.vercel.app/work.html
• 💼 LinkedIn Profile: linkedin.com/in/ai-expert-arpan-ghosh
• 🐙 GitHub Profile: github.com/Arpanthebaap
• ✉️ Primary Email: arpanghosh984@gmail.com
• 📍 Residence: West Bengal, India (International applicant requiring F-1 / doctoral visa sponsorship)
• 🗣️ Languages: English (CEFR C1/C2 equivalence; medium of instruction for all university degrees) • Bengali (Native) • Hindi (Professional)