"What I cannot create,
I do not understand."

Teaching assistant for the graduate NLP course (50 students): hold office hours, grade assignments, and advise students on final NLP projects.

The course covers language modeling, tokenization, part-of-speech tagging, information extraction, topic modeling, sentiment analysis, and supervised text classification, taught to social scientists rather than CS students.

Built the question-answering agent for RelGraph, the firm's relationship graph (LangGraph, LLM tool use): plans retrieval across 8 graph, SQL, and vector tools, finds warm-intro paths up to 4 hops out, and cites every claim to a database record, redacting uncited names. Wrote the eval harness gating every prompt and retrieval change in CI on retrieval recall and citation integrity.

Shipped the deal intake workflow on the firm's internal AI platform (React, Express, PostgreSQL, Docker, Azure): a subagent reads the shared inbox, extracts deal fields from emails and PDFs, and writes them into the CRM (~6,300 companies) after an associate confirms.

Shipped an NDA redlining agent through 7 releases of associate and VP testing on live deals, plus a meeting-note pipeline in daily use syncing 45 interactions per week into the firm's CRM through its REST API, behind a human-confirmation gate.

Ran 1:1s with the full deal team to map where time goes in sourcing, screening, and document review, turned the pain points into a prioritized build plan, and wrote the docs and prompting patterns for every tool shipped. Every deployed tool sits behind governance controls: human-confirmation gates on all CRM writes, queued mutations with audit logging, and zero autonomous external communication.

Built the document-to-knowledge pipeline behind the lab's hierarchical sparse autoencoder (SAE) interpretability work: 3,000+ pages of technical PDFs (MinerU), 17,000+ embeddings (Qwen3-Embedding), and 493 topics surfaced with BERTopic/HDBSCAN at 0.9791 mean cosine similarity.

Mapped prerequisite knowledge chains with Gemma topic labeling and hierarchical clustering. Separately studied how representations emerge across 24 models from 5 families on the Gutenberg corpus.

Trained YOLOv7/v8 defect detection models for manufacturing QC. Deployed on NVIDIA Jetson with DeepStream SDK and TensorRT (FP16/INT8) achieving 3x inference speedup and 25% reduced detection latency.

Designed Flask REST APIs for real-time inference across 3 production lines. Built multi-threaded Docker backend with AWS/Azure data pipelines, CI/CD, and 85% test coverage.

Built K-Means clustering to identify 5 customer personas. Designed Grafana dashboards tracking 15+ KPIs: revenue, churn, CAC, and operational efficiency.

Conducted EDA on 100K+ transactions using Python and SQL. Automated reporting pipelines, reducing manual work by 40% and enhancing operational visibility by 30%.