Sai Teja Vaidya

AI Brain Lab (Contract) · Bangkok, Thailand (Hybrid)

• Collaborated with AI research and engineering teams on computer vision tasks and applied machine learning algorithms for production platforms.
• Designed intelligent platforms transforming complex operational workflows into data-driven insights, improving customer service outcomes.
• Contributed to AI-driven workflow automation solutions, enhancing operational efficiency and end-user experience.

Asian Institute of Technology (AIT) · Khlong Luang, Thailand

• Led “AI Agent-Orchestrated Instance Segmentation”: detection and canopy size estimation of oil palm trees using UAV imagery across 8 resolution levels (0.03–0.20m GSD).
• Deployed best-performing model weights as an AI Teacher Agent for automatic annotation, reducing manual labelling time by over 80%.
• Benchmarked 6 models (YOLOv8, YOLOv11, Mask R-CNN, Hybrid-SAM) achieving mean IoU of 0.77 and F1 of 0.789 at optimal 0.04m GSD.
• Supervised by Dr. Sarawut Ninsawat, Prof. Nitin Tripathi, and Dr. Sanit Arunplod; published Q1 journal paper and successfully defended thesis.

Asian Institute of Technology (Contract) · Bangkok, Thailand

• Planned and executed internal/external events for an international graduate community, managing hybrid Zoom setups for technical operations.
• Produced multimedia content and website updates, aligning messaging with program objectives and stakeholder needs.

Asian Institute of Technology (Part-time) · Pathum Thani, Thailand

• Coordinated academic programs and multicultural seminars/technical talks, boosting student engagement across the AIT_SDP program.
• Designed information-flow systems that improved communication with a diverse international student body.

Asian Institute of Technology · Pathum Thani, Thailand

• Conducted a literature review of AI-assisted COVID-19 trend analysis and spatial hotspot mapping methods (2019–2022).
• Synthesized four forecasting and four spatial hotspot-detection methodologies into a decision-intelligence comparison of predictive performance.

Tauri, React, Whisper (local inference)

• Built a real-time teleprompter and rehearsal coach ingesting scripts/slides via OCR with local speech transcription for instant pacing and filler-word feedback.
• Designed a latency-budgeted feedback engine: fast local pass for immediate flags, slower deep pass for structural/tone analysis.

NetworkX, Vector Store

• Built a self-organizing memory graph (episodes + abstractions as nodes) that reduces agent run cost over time instead of scaling context tokens.
• Implemented a reflection agent merging recurring patterns into reusable abstractions; benchmarked cost-per-task, retrieval hit rate, and accuracy delta across a 100–200 task stream.

LangGraph, Qwen2.5-3B, API Fronting

• Built a LangGraph-based agentic router that sends each support ticket to the cheapest capable model tier while holding accuracy steady.
• Added early-exit and guardrails with a “return-to-control” path, benchmarking cost-per-ticket reduction and p95 latency vs. an always-expensive baseline.

Multi-Agent Orchestration, LLM Guardrails

• Designed a 9-agent research writing system (interrogator, writer, reviewer, citation checker, and more) enforcing 6 immutable writing laws to eliminate hallucination.
• Built a 3-layer architecture and 90% confidence gate blocking drafting until research foundations are solid; used to support Q1 paper acceptance and thesis defense.

ONNX, Edge Inference, XR2 Gen 2

• Reproduced the LPCVC 2026 image-to-text retrieval baseline for XR2 Gen 2 (proxy edge device): ONNX export, model compilation/profiling, dataset upload.
• Achieved a validated baseline of Recall@10 = 0.7299 with 29.3ms image encoder and 5.4ms text encoder latency; resolved key Python/SSL/dataset reproducibility blockers.

YOLOv8/v11, Mask R-CNN, SAM

• Delivered accurate oil palm crown delineation and canopy size estimation from UAV RGB imagery across 8 GSD levels (0.03–0.20m) for the $67B global oil palm industry.
• Compared 6 segmentation models on precision, recall, F1, IoU, and crown geometry error; agent-in-the-loop annotation cut labelling time by 80%.

Asian Institute of Technology · Thailand · GPA 3.0/4

• Thesis: “Detection and Canopy Size Estimation of Oil Palm Trees Using YOLOv8, YOLOv11, and SAM on Multi-GSD UAV Imagery”

JNTUH College of Engineering, Hyderabad · GPA 8.8/10

• Built “College Buddy App” — a peer-to-peer campus marketplace for sustainable textbook exchange; Student Volunteer & Coordinator, QUEST Technical Fest.