Harshit Sharma Email & Phone Number

Boston, Ma, Us

Sunnyvale, Ca, Us

Sunnyvale, Ca, Us

Project: Conversational Interface Microservice / AB Testing of LLM Vendors• Developed synchronous and streaming REST endpoints in a Langchain and FastAPI based CI Microservice integrating third party LLM vendor including memory functionality and feedback mechanism. • Conducted statistical hypothesis testing for A/B testing of different LLM vendors and created a dashboard to present comparative results to business stakeholders.Project: Anomaly Detection from Logs• Implemented an ensemble of Masked Language Model (RoBerta) and Graph Convolution Network augmented LSTM to identify anomalies from logs. Trained models from scratch on custom data. • Deployed model with optimized inference (ONNX and quantization) using HuggingFace Inference Endpoints.• Hosted Streamlit and flask based application on EC2 to incorporate feedback from domain experts.• Performed log-parsing, ontology creation and RDFization to represent logs as Knowledge-Graph to visualize and aid in debugging logs using Neo4jProject: AIOps for Optics Telemetry Data• Developed XGBoost model to identify anomalous telemetry datapoints from optics. • Currently being integrated into the production pipeline consisting of Kafka, Apache Storm and Spark.

Sunnyvale, Ca, Us

Amsterdam, Noord-Holland, Nl

Boston, Massachusetts, Us

Course : Natural Language Processing (CS 4120) under Prof. Felix MuznyGrading assignments and aiding students in topics : Probabilistic language modeling, Text classification with language models, Text classification with single layer neural networks, Vector semantics & word embedding, Part-of-speech tagging, Viterbi algorithm, Interactive dialog agents, Machine translation

Boston, Massachusetts, Us

Course : Artificial Intelligence (CS 4100)

Mumbai, Maharashtra, In

Project: Geographical Address Matching• Created geographical address similarity algorithm using CNN-BiLSTM-CRF model.• Devised an improved textrepresentation technique for capturing semantic characteristics of addresses by combining context dependent and independent embeddings (Word2vec, FastText and Elmo).• Architected a data driven framework to profile millions of addresses using information theory and regularexpressions. The framework enabled automated generation of realistic training samples for powering supervisedsequence learning algorithms. • Currently utilised by some of the largest financial institutions of India for KYC &due-diligence purposes.Project: Named Entity Recognition from Legal documents• Implemented Attention based BiLSTM-CRF model for NER from legal documents. • Built an end-to-end pipeline forparsing data from unstructured text and preparation of training samples using AWS Groundtruth and rule-basedsyntactical manipulation techniques.• Designed a real-time feedback mechanism to enable out-of-core learning,eliminating need for manually tagged data.

Chennai, Tamil Nadu, In

• Used Multi-Label classification model for classifying 120,000 websites into 422 categories. • Applied BOW model andimproved traditional Tf-Idf approach by incorporating class-frequency.• Modified mathematical formulation ofComplement Naive Bayes to handle faults of algorithmic assumption. • Reduced impact of redundant features andapplied text-length normalizations, leading to a classification accuracy of 93%

Master'S Degree, Data Science

Graduate Course, Deep Learning

Bachelor'S Degree

High School