Portfolio 路 lteresa@umd.edu 路 LinkedIn
Retrieval, agents, and distributed data, with a habit of building the measurement before trusting the system. Most of what I learn ends up here as something runnable.
** Hot take - Data cleaning is 80% of the job and 100% of what nobody puts in the job description.**
Reinforcement learing and agent reward modeling 路 PyTorch, HuggingFace, SmolLM2
Working through the post-training stack by implementing it from scratch rather than calling
trainer.train(): supervised fine-tuning, reward modeling, PPO, DPO and GRPO on a 135M model, plus a
small GPT-2 and a BPE tokenizer. The part I find most interesting is the reward-hacking probe, which
optimizes against a learned proxy reward while scoring on a held-out gold metric and watches the two come
apart: gold quality collapses to 0.03 under a weak KL penalty, against 0.98 under DPO.
biomed-rag 路 Python, Pinecone, LlamaIndex, HuggingFace Retrieval over biomedical literature with cited sources. Hybrid dense and BM25 retrieval, reranking, and a hand-labeled 56-query benchmark split easy and hard, because a single average would have hidden that Recall@10 falls from 1.00 to 0.50 on the queries that actually matter. Grounding and citation checks run offline in CI, so a regression blocks a merge instead of reaching a user.
bayes-execution-engine 路 Python, LangGraph, MCP, pgmpy Multi-agent orchestration on a deterministic plan-and-execute schedule, agents calling tools over MCP. Bayesian graphical models resolve conflicting agent outputs into one answer carrying an uncertainty estimate, instead of retrying until the agents happen to agree.
big-data-analytics 路 Spark, Airflow, Dask, PostgreSQL, MongoDB, Neo4j, Redis Batch and stream processing with DAG orchestration, running comparable workloads across relational, document, graph and key-value stores to see where each access pattern actually pays.
peekaboo-webapp 路 Vanilla JS, Tesseract.js, Ollama Learning app for preschoolers with OCR and an AI tutor running Gemma3 locally through Ollama. No framework, hand-rolled state-driven rendering, six-tier adaptive progression.
Python 路 Java 路 SQL 路 PyTorch 路 scikit-learn 路 pandas 路 NumPy 路 HuggingFace 路 LangGraph 路 MCP 路 Pinecone 路 Spark 路 Airflow 路 Dask 路 PostgreSQL 路 MongoDB 路 Redis 路 Docker 路 Terraform 路 AWS 路 Azure 路 Git