Kiran Sheoran

Built an autonomous AML investigation system using LangChain AgentExecutor with tool-calling across 5 datasets (3,000 customers, 56,000+ transaction records) to investigate fraud alerts step-by-step. Engineered 6 LangChain @tool functions covering KYC profiling, account analysis, suspicious transaction extraction, and SAR deduplication logic to reduce false positives. Designed a 7-section sequential SAR generation pipeline grounded in an evidence JSON schema, preventing LLM hallucination by restricting each section to verified tool outputs only. Integrated Langfuse observability (self-hosted via Docker) for full agent trace capture and built a 3-panel Streamlit UI for live trace inspection and analyst review of SAR drafts before filing.

• LangChain
• Groq
• Langfuse
• Streamlit
• Python

Won 1st place nationally at Smart India Hackathon 2025 (ISRO) — built an offline-first Android app classifying 12 fish species using YOLOv8 + TensorFlow Lite, deployed for fishermen in low-connectivity coastal regions. Achieved 97% accuracy in species detection and freshness analysis entirely on-device without internet connectivity. Implemented coin/ArUco marker-based volume estimation using OpenCV with biological weight-length formulas and dual-model inference pipeline balancing real-time preview speed with capture accuracy. Designed offline-first SQLite + Firebase Firestore sync using Android WorkManager for automatic background upload of geo-tagged catch data.

• Android
• Kotlin
• TensorFlow Lite
• OpenCV
• YOLOv8
• SQLite
• Firebase Firestore
• Android WorkManager