Portfolio

All Projects

Spanning Computer Vision, LLM/RAG, Deployment, Robotics, and Generative AI. Filter by category to explore.

Showing 8 projects

Computer Vision
Dec 2025

AeroEyes — AI-Powered Drones for Search & Rescue

Problem

Field SAR teams need AI that can generalize to unseen object classes mid-mission — victims, debris, equipment — using drone footage and only 3 labelled examples per class. Standard detection models require thousands of annotations; the challenge is building a system robust enough to rank on a national leaderboard while handling real-world aerial noise, scale variation, and truly novel categories.

Results

  • ›Ranked Top 2 nationally in ZaloAI Challenge 2025 — highest-ranked submission in the Computer Vision track.
  • ›Achieved 0.77 public score while generalizing to novel object classes with near-zero labelling effort (3 images per class).
Top 2Competition Rank
0.77Public Score
3 / classRef. Images
YOLOEPythonPyTorchFew-shot LearningMulti-scale Inference
Details
LLM/RAG
LLMOps
2025 – 2026

Medical RAG Chatbot Platform

Problem

A healthcare client needed two complementary AI tools: (1) a trustworthy medical Q&A chatbot where every answer is grounded in curated clinical documents — not hallucinated from model weights — with full citation traceability; (2) a secure internal platform letting medical staff fine-tune smaller LLMs (Qwen3, Phi-3) on proprietary datasets (genetic counseling, Fukushima case studies) without exposing data to external APIs. Both systems had to run on-premise or Azure with strict security: no credentials in Git, no exposed ports, and observable pipelines that could be improved continuously.

Results

  • ›Delivered two production AI systems — RAG chatbot and fine-tuning platform — both fully containerized and serving real clinical workflows.
  • ›RAG pipeline grounds every response in cited source documents, eliminating hallucination from model weights and enabling clinical teams to verify each answer against original PDFs.
5Docker Services
LoRA · vLLMFine-tune
LangfuseLLMOps
6+AI Providers
LibreChatLangChainpgvectorMongoDBMeilisearch+13
Details
Computer Vision
Deployment
2024 – 2025

Multi-Camera Person Re-ID & Counting System

Problem

Retail and venue operators need accurate headcounts across multi-camera zones — but naive per-camera counting double-counts anyone who moves between camera views, making the numbers meaningless. The core challenge: assign a single global identity to each person across N live camera streams, even when they disappear from all feeds for hundreds of frames and reappear with different lighting, scale, or pose. Simultaneously, the system must count each unique person crossing a virtual entry/exit line — without counting the same person twice across cameras.

Results

  • ›Eliminates cross-camera double-counting: each person receives one global ID regardless of how many camera zones they traverse, giving accurate unique-person headcounts.
  • ›Pose-quality gates prevent low-confidence detections from corrupting global IDs — improving identity precision on crowded or partially occluded frames.
512-dimReID Features
17 ptsPose Keypoints
3000 framesIdentity Memory
Gradio + WebSocketInterfaces
YOLOv8OSNet (TorchReID)BoTSORTByteTrackPyTorch+6
Details
Computer Vision
Deployment
Robotics
Jul 2024 – Sep 2025

Industrial Defect Detection & AMR Warehouse System

Problem

A manufacturer needed two AI systems delivered simultaneously: an automated vision inspector to replace inconsistent manual defect checking on production lines, and a fully autonomous mobile robot (AMR) for warehouse pick-and-deliver — both shipping to production on a fixed 14-month timeline with no margin for failure.

Results

  • ›Eliminated manual re-inspection bottleneck on production lines — improving quality consistency and freeing operator capacity for higher-value tasks.
  • ›Shipped containerized model releases with zero-downtime deployments via Docker + ONNX Runtime microservices.
+25%Throughput Increase
<3 minDelivery Time
PyTorchYOLOONNXONNX RuntimeDocker+3
Details
Computer Vision
Generative AI
Oct 2024

Watermark Removal with Pix2Pix GAN

Problem

Removing diverse watermarks from high-resolution images while preserving fine texture detail — with zero paired real-world training data available. Standard inference breaks on large images due to memory constraints, requiring a solution that handles both data scarcity and full-resolution output without artifacts.

Results

  • ›Removed diverse watermark types (text, logos, semi-transparent patterns) with no real-world labelled data required.
  • ›Preserved fine texture details via combined L1 + adversarial objective — outperforming a pure L1 baseline in perceptual quality.
L1 + AdversarialLoss
Tiled (overlap)Inference
PythonPyTorchPix2PixGANImage Tiling+1
Details
LLM/RAG
Oct 2023

Tweet Sentiment Extraction

Problem

Extract the precise text span from a tweet that supports its sentiment label — a token-level localization task where a single off-by-one token prediction tanks the F1 score. Tweets are noisy: URLs, HTML entities, inconsistent casing, and slang all degrade tokenization and model accuracy.

Results

  • ›Achieved F1 0.72311 on the Kaggle competition test set through multi-model ensembling.
  • ›Demonstrated that tweet normalization preprocessing directly lifts accuracy by cleaning tokenization noise before training.
0.72311F1 Score
PythonPyTorchTransformersBERTRoBERTa+1
Details
Computer Vision
Nov 2022

Real-Time Face Mask Detection

Problem

Detect mask/no-mask status from live webcam feeds with high accuracy and real-time latency — while remaining robust to varied face poses, lighting conditions, and partial occlusions that naive classifiers struggle with.

Results

  • ›Achieved 98% classification accuracy on the held-out test set — surpassing the baseline EfficientNet fine-tune without alignment preprocessing.
  • ›Sustained 30 FPS inference on a standard webcam, enabling real-time deployment without specialized hardware.
98%Accuracy
30 FPSInference Speed
PythonTensorFlowKerasMTCNNOpenCV+1
Details
Generative AI
Jan 2023

Banner Generation — ZaloAI 2023

Problem

Generate high-quality promotional banner images from raw Vietnamese marketing copy — requiring accurate translation to English, structured keyword extraction for visual cues, and effective prompt construction for diffusion models, all while preserving the original brand tone.

Results

  • ›Delivered a complete pipeline from raw Vietnamese marketing copy to production-ready banner image — no manual prompt engineering required.
  • ›Demonstrated that LLM keyword expansion + vi2en translation meaningfully improves Stable Diffusion output quality over direct Vietnamese prompting.
vi → enTranslation
Stable DiffusionGeneration
PythonHuggingFace TransformersStable DiffusionPyTorchQwen LLM+1
Details