Answer Engine Optimization
Full-stack SaaSA distributed B2B platform that tracks and optimizes visibility across LLM answer engines, with real-time prompt ranking, competitive intel dashboards, and AI content recommendations.
PythonGoReact
I build AI agents for businesses at Meta, and the evaluation frameworks that prove they actually work. Data Scientist by training, ML Engineer by habit.
I am comfortable owning a system from the data layer to the deployed service, not just the model in the middle.
Backends in Python and Go, React frontends, distributed services, and ML pipelines on GCP and Vertex AI with Docker, Kubernetes, Airflow, and Terraform.
Production deep learning, computer vision, and NLP, plus LLM systems built with LangChain: RAG and retrieval over FAISS and Pinecone, fine-tuning, LlamaGuard-style guardrails, and agent tooling on PyTorch, HuggingFace, LLaMA, and GPT.
Evaluation frameworks, A/B and A/A testing, causal inference, and the metrics that turn model quality into decisions leadership can act on.
A distributed B2B platform that tracks and optimizes visibility across LLM answer engines, with real-time prompt ranking, competitive intel dashboards, and AI content recommendations.
An LLM-powered prompt-optimization platform on Llama with LlamaGuard guardrails, auto-enhancing free-form input through git-style, version-controlled iterative refinement.
Fine-tuned Vision Transformers for image classification, reaching 95% accuracy on the CIFAR-10 dataset.
A face-authentication system that flags counterfeits from StyleGAN, DCGAN, and PGGAN using a DenseNet classifier, deployed on GCP.
A music IR system reaching 80% MAP, using MAGENTA's MT3 (a T5 architecture) for transcription over a custom dataset built on GTZAN.
Models, notebooks, and experiments across NLP, computer vision, and LLM tooling, plus everything that did not fit above.
A running record of specific changes to tools I work with. Each entry links to the pull request and its review history.
Store reasoning tokens separately from visible output and expose the count in request history, stats, and the Control Plane.
Read PR #4745Import AI SDK schemas from generation metadata so the trace-to-Playground flow retains the output constraint.
Read PR #17310Add a default HTTP timeout at every Gemini client construction point, with regression tests, so a stalled request cannot hold the deriver indefinitely.
Read PR #903Label OpenRouter requests from cron jobs so repeated runs can be identified in usage analytics.
Read PR #68290The full stack I reach for, from raw modeling to the infra that ships and monitors it.
M.S., Data Science, Khoury College
B.Tech, Computer Science