Backend & AI Engineer • Agentic Systems • Full-Stack Development

Computer Science undergraduate at Presidency University, Bengaluru (Class of 2028).

I build AI systems that go beyond model inference — combining intelligent orchestration, reliable backend infrastructure, multimodal reasoning, and evidence-driven architectures.

Previously worked on LLM/VLM-based defense intelligence systems at DRDO. Currently building clinical AI and healthcare infrastructure at PetOLife.

🏆 Two-Time Hackathon Runner-Up

• 2nd Place — Fusion X National Hackathon | AI/ML & Data Science Track | 150+ teams
• 2nd Place — ISACA Hackathon | AI-Powered Fraud Detection | 80+ teams

Agentic Security Investigation • Semantic Fracture Detection • Evidence-Based Verification

An autonomous artifact intelligence platform designed to investigate what digital files actually contain, how different software interprets them, and whether those differences introduce security risks.

• Semantic Fracture Detection: Developed a deterministic investigation engine that executes independent parsers against identical file bytes, compares structural interpretations, and identifies security-relevant discrepancies.
• Multi-Agent Investigation: Built an orchestration framework coordinating Evidence, Hypothesis, Experiment, and Validation agents through bounded investigation and evidence-refinement loops.
• AI-Powered Triage: Integrated a fine-tuned Laya model for risk assessment and adaptive routing between PASS, DEEP_SCAN, and PRISM_LAB workflows.
• Verified Experimentation: Implemented typed experiment schemas, allowlisted execution plans, replay protection, and deterministic validation to ensure AI-generated hypotheses are independently verified.
• Secure Containment: Developed real-time artifact monitoring and encrypted quarantine using AES-256-GCM with Windows DPAPI-protected keys.
• Explainable Investigation: Built a full-stack interface exposing parser comparisons, investigation traces, experiment histories, and verified findings.

Stack: Python, FastAPI, Next.js, TypeScript, Laya, Ollama, BGE-M3, RAG, PyMuPDF, AES-256-GCM, SSE.

Geospatial AI • Vision-Language Models • Multi-Model Orchestration

A multimodal Earth Observation intelligence platform that transforms natural-language questions and satellite imagery into traceable spatial evidence.

• Multimodal Intelligence: Architected analysis workflows supporting RGB, Sentinel-2 multispectral, Sentinel-1 SAR, and bi-temporal satellite imagery.
• Specialist Model Orchestration: Integrated GeoChat-7B, Grounding DINO, ChangeFormer V6, Chg2Cap, and optical–SAR representation encoders into a unified backend.
• Change Detection: Developed workflows combining learned change masks, semantic change descriptions, and radiometric fallback analysis.
• Evidence-Grounded Reasoning: Preserved model provenance, bounding boxes, segmentation masks, change regions, and execution traces to prevent unsupported spatial claims.
• GPU Worker Architecture: Designed isolated inference workers for computationally intensive vision-language models with configurable local and remote execution.
• Interactive Analysis Studio: Built a React-based geospatial workspace for satellite imagery, model outputs, spatial overlays, and project-scoped AI conversations.

Stack: Python, FastAPI, PyTorch, TorchGeo, CUDA, Rasterio, React, TypeScript, Supabase, RAG.

Repository currently private.

Agentic Software Engineering • Stateful Execution • Human-in-the-Loop Workflows

A multi-agent software engineering platform that transforms natural-language project requirements into structured research, architecture, implementation, and deployable workspaces.

• Stateful Agent Orchestration: Built cyclical LangGraph workflows coordinating Research, Feasibility, Architecture, and Review agents with shared state and iterative refinement.
• Human-in-the-Loop Execution: Implemented persistent interruption checkpoints allowing developers to inspect, approve, modify, and resume agent workflows.
• Real-Time Infrastructure: Engineered asynchronous WebSocket event streams for live token generation, agent execution status, and terminal output.
• Structured Code Generation: Developed schema-constrained LLM output pipelines with workspace patching, validation, and exportable project generation.
• Interactive Developer Workspace: Built a Next.js interface integrating React Flow, Monaco Editor, live execution feedback, and project visualization.

Stack: Python, FastAPI, LangGraph, LangChain, SQLAlchemy Async, WebSockets, Next.js 16, React 19, Zustand, Monaco Editor, React Flow.

2nd Place — ISACA Hackathon | 80+ Teams

An AI-powered fraud intelligence platform designed to detect UPI payment scams, phishing URLs, malicious QR codes, and fraudulent transaction claims.

• Hybrid Fraud Detection: Combined fine-tuned DistilBERT, Qwen LLM reasoning, and deterministic verification for fraud assessment.
• Phishing Intelligence: Fine-tuned a DistilBERT classifier on a 210,000-record phishing and benign URL dataset, achieving 99.34% test accuracy and 0.9303 phishing F1-score.
• Multimodal Analysis: Integrated QR-code parsing, OCR-based evidence extraction, and contextual reasoning to investigate suspicious payment requests.
• Evidence-Based Decisions: Designed fraud analysis workflows that distinguish verified technical signals from model-generated interpretations.

Stack: Python, FastAPI, PyTorch, Hugging Face Transformers, DistilBERT, Qwen, Scikit-learn.

2nd Place — Fusion X National Hackathon | 150+ Teams

A predictive crime intelligence platform combining spatial risk modeling, temporal forecasting, explainable AI, and conversational analytics.

• Hybrid Predictive Modeling: Built an XGBoost + LSTM ensemble for spatial crime severity analysis and temporal crime-volume forecasting across 20,000+ records.
• Dynamic Ensemble Architecture: Developed a Meta-Dynamic Weighting System that adjusts model contributions based on regional data volatility.
• Explainable AI: Integrated interpretable risk analysis to surface important prediction drivers and support evidence-based decision-making.
• Conversational Analytics: Built a RAG-powered assistant using Gemma 3 and ChromaDB for natural-language interrogation of crime datasets.
• Geospatial Dashboard: Developed interactive maps and regional intelligence visualizations using Leaflet.js and Recharts.

Stack: Python, FastAPI, XGBoost, LSTM, TensorFlow, Gemma 3, ChromaDB, Next.js 14, Leaflet.js.

Document AI • Hybrid Retrieval • Distributed Processing

A full-stack document intelligence platform for evidence-grounded question answering across uploaded documents and external knowledge sources.

• Hybrid Retrieval: Combined dense vector retrieval through ChromaDB with live web retrieval for context-aware document analysis.
• Distributed Ingestion: Built asynchronous PDF processing pipelines using Celery workers and Redis message brokering.
• Adaptive Model Routing: Integrated Gemini Flash and Gemini Pro with query-complexity-based routing.
• Interactive Citations: Developed a PDF viewer with bidirectional citation navigation linking AI responses directly to document evidence.
• Backend Infrastructure: Implemented PostgreSQL-backed persistence and secure session management.

Stack: Python, FastAPI, Next.js, ChromaDB, Celery, Redis, PostgreSQL, Gemini.

🥈 Fusion X National Hackathon — 2nd Place (April 2026)
AI/ML & Data Science Track | 150+ Participating Teams
Built CrimeGrid, combining predictive modeling, explainable AI, and retrieval-augmented intelligence.

🥈 ISACA Hackathon — 2nd Place (September 2026)
80+ Participating Teams
Built SENTINEL, a neuro-symbolic fraud detection platform integrating transformer-based classification, LLM reasoning, and multimodal payment security analysis.

• Email: adithya.kri06@gmail.com

Building AI systems that don't just generate answers — they reason, verify, and turn intelligence into reliable software.