Work

Projects

A collection of AI/ML projects spanning computer vision, multimodal learning, generative AI, and full-stack deployment.

LoRA Fine Tuning

A QLoRA fine-tune that turns messy booking requests into schema-valid JSON, shipped with a baseline comparison and a rank sweep, because a falling loss curve doesn't prove anything.

LoRAQLoRAFine-TuningModel Evaluation

GitHub ↗

Content Guard

A desktop player that scans a video once for explicit content and skips past it during playback, three complementary detectors, fully offline, no GPU, runs on your own hardware for full privacy.

Computer VisionONNX RuntimeEdge InferenceModel Ensembles

GitHub ↗

Semantic Caching Layer

A semantic cache in front of Gemini that matches queries by meaning, with an evaluation harness that sweeps the similarity threshold to keep the operation cost efficient.

Semantic CachingVector SearchFAISSEmbeddings

GitHub ↗

Conductor

Four specialized agents coordinated by a state graph, running as durable background jobs behind a REST API, with state checkpointed after every step so a crashed worker resumes instead of starting over.

Multi-Agent SystemsAgent OrchestrationLangGraphDistributed Systems

GitHub ↗

NBA Halfcourt

An ML model predicts the second half of an NBA game from halftime stats; an LLM writes the broadcast commentary based on the model's SHAP scores and is never allowed to contradict it.

Machine LearningLLM IntegrationModel ExplainabilitySHAP

GitHub ↗

AffectSync

Generative models learn what a joke looks like, never whether whether it made anyone laugh. AffectSync aims to bridge that gap by exporting emotion-aware training data.

Computer VisionEmotion RecognitionMultimodal DataDataset Engineering

GitHub ↗