Skills

Natural Language Processing, Deep Learning, Large Language Models, Applied Research, Git, GitHub, Python, C++, C#, PyTorch, Machine Learning, Linux, SQL, Azure, GCP, MLOps, Kubeflow, Embeddings, Reinforcement Learning, vLLMs, Multi-GPU training, Prompt Optimization, Langchain, Langsmith, Streamlit, OpenAI API, Huggingface, CrewAI, SLURM, ONNX

JPMC

• Designed a robust evaluation framework for a long-form News QA RAG pipeline, outperforming open-source baselines and LLM-as-a-judge methods.
• Achieved +16% human-alignment scores, with measurable gains in factual accuracy and coverage.
• Created ~20 meta-evaluation metrics to rigorously benchmark recall and other key performance indicators.

Microsoft (MSFT)

• Built LLM-powered auto-refresh pipelines to enhance triggering logic, improving answer relevance across Bing search.
• Designed and deployed a schema-agnostic events extraction pipeline, enabling event discovery from all webpages and significantly boosting coverage.
• Trained and optimized XLM-R models, cutting L1 latency from 200ms+ to 12ms at 20B+ webpage scale, generating 1M+ new events.
• Applied optimizations including ONNX conversion, model pruning, and Blingfire tokenization, deploying models to serve live production traffic.
• Published SegRank, a novel LLM grounding method, reducing hallucinations in Bing Chat by 15.7%.
• Honorable Mention, Executive Challenge, Microsoft Global Hackathon 2024.
• Award of Excellence for Innovation (Microsoft IDC) for events data pipeline.
• Fast-tracked promotion to L60 in 9 months for exceptional ownership and impactful project delivery.

ETH

• Developed CausalCite, a novel causal metric for quantifying a paper’s scholarly impact.
• Conducted large-scale experiments on a citation graph with 2B+ edges and 200M+ nodes.
• Demonstrated 30.14% higher correlation with test-of-time impact compared to traditional citation counts.
• Accepted and presented as first author at ACL Main Conference 2024, Bangkok.

Deloitte

• Modelled Probability of Success of pursued inbound opportunities, resulting in 76% better prioritization.
• Created an internal MLOps platform that allowed conversion of Jupyter notebooks to Kubeflow Pipelines.