Three years of NLP and LLM work in production. These days that means LangGraph agents that call retrieval tools when they need them, plus the tracing and eval loop that tells me a change made things worse before a user does.
My habits: evals before features · a trace on every run · structured outputs instead of prompt warnings · cost and latency on the dashboard from day one.
flowchart LR
U([User query]) --> A{Agent<br/>LangGraph}
A -- needs context --> R[(Hybrid retrieval<br/>BM25 + dense)]
A -- needs fresh facts --> W[Web search tool]
R --> A
W --> A
A --> S[Structured output<br/>schema-validated]
S --> G{Eval gate<br/>groundedness · tool-call accuracy}
G -- pass --> O([Answer + citations])
G -- fail --> A
A -.-> T[[Trace · cost · latency]]
More projects
- Marketing Incrementality Platform: causal lift measurement across ~99K customers. Propensity modeling plus IPTW showed the naive campaign lift was overstated by ~2.7×.
- Human Activity Recognition: LRCN-based activity classification from video.
- BPE Tokenizer · Training Loop · Loss Harness: from-scratch building blocks for LLM training.
Instructional Technologist, AI/ML Infrastructure · GWU Data Science · Aug 2026 – Present
- Building an advising assistant with hybrid dense + sparse retrieval grounded in program documentation, so students get cited answers instead of generic ones.
- Run the AWS ML environments (EC2, S3, IAM, cost controls) for graduate capstone teams; containerized workspaces took setup from days to hours.
- Scope and de-risk applied AI capstones with faculty and industry sponsors.
Research Assistant · LAiSER, GWU · Feb 2025 – May 2026
- FAISS-based similarity search over 10,000+ job descriptions, mapped to the ESCO and O*NET skill taxonomies with a Neo4j graph for cross-taxonomy alignment.
- Shipped it as a multi-tenant API on AWS (FastAPI, Lambda, S3, EC2) with Docker and GitHub Actions CI/CD: 3× research throughput, and time-to-demo from weeks to a live public product.
- Presented the architecture to 150+ attendees at Badge Summit 2025.
Machine Learning Engineer · Ideapoke (now Chesz.ai) · May 2022 – Nov 2023
- Rebuilt the NER pipeline behind Signalz, which turns news, patents and papers into market and technology signals: +40% extraction accuracy, −60% processing time.
- Replaced brittle rule-based extraction with RAG (LangChain + GPT) producing structured output from large-scale unstructured web data.
- Built a contextual search engine that retrieved on query intent rather than keyword match, across 30+ industries.
| Agents & retrieval | LangGraph · LangChain · MCP · Hybrid retrieval (BM25 + dense) · Ollama · Neo4j · FAISS · pgvector |
| Evals & LLMOps | LangSmith · Langfuse · RAGAS · LLM-as-Judge · Pytest |
| ML & NLP | PyTorch · Hugging Face Transformers · scikit-learn · XGBoost · causal inference (IPTW) |
- 🎤 Badge Summit 2025: presented "Skills Are the Common Language" to 150+ attendees.
- 🛠 DevNetwork AI+ML Hackathon 2025: GearGuide, graph-powered RAG for automotive troubleshooting, in 48 hours.
- 🌍 UN Open Source Hackathon 2025 · 🎮 HoyaHacks 2025 (Microsoft track) · 📊 World Bank DataDive 2024 with DataKind DC
- 🌟 Global Leaders Scholarship, GWU (2024–25)
- 🎓 M.S. Data Science, George Washington University (2026, GPA 3.91) · B.E. Computer Science, VTU Bengaluru


