Nakulsai Adapala

Organized and dependable candidate successful at managing multiple priorities with a positive attitude. Willingness to take on added responsibilities to meet team goals.

NLP Engineer at Inrythm Solutions Pvt Ltd <> Stocksnips.ai (2022-Present)

• Build custom NER model to detect companies in the text (Accuracy-91%) i.e. transformer based model.
• Deployed and used CORE NLP model on GCP (stanford core nlp) later replaced with spacy custom ner
• Sentiment analysis on news using classification algorithms i.e. BERT
• Built Reinforcement Learning model on stocks with sentiment and stock value
• Calculated the weightages between sentiment and moving stocks. i.e adaptive weights
• Coreference resolution on large chunks of text
• Unix Commands like: SCP My notes on unix commands: url: https://www.notion.so/Unix-commands-3dc1011b7874488b8bf13cb447fe3d9b?pvs=4
• A good usage of numpy arrays rather lists and such kind of things to handle huge data in less amount of time url: https://www.notion.so/numpy-a375f55b118140ecb2f4e20750fab297?pvs=4
• Interests: Distributed Learning: https://www.notion.so/Distributed-learning-a46d6a36718647dc85f0ba63128e69bb?pvs=4, Scala: https://www.notion.so/scala-4ffb19ec97f644f0988ebb26d01ae838?pvs=4

Phrazor Product KB Lead at vPhrase (2017-2022)

• Lead the product team on developing a product which generate insights on the fly
• Made Forecasting algorithms which is used by clients like Abbott, Sun Pharma. Used smoothening techniques like Khalman filter and also written statistical rules for data shift and model shift in aws for MLOPS
• PARAPHRASER: based on T5 model using hyper parameters as Top-k sampling and Top-P (nucleus) sampling . Used BLEU score for accuracy
• Made algorithms to determine outliers , seasonality and also effectively using them in diagnostics.
• Named Entity Recognization :using BERT-model
• Document classification and summarisation: Extracted text from documents and classified them to multiple buckets and also assigned sentiment to the each bucket. Used techniques like chunking and chinking and custom NER. Used ROGUE score to calculate accuracy
• TechGig BigData Analytics Runnerup: fraud detection using simple Bayesian models
• Grammar Correction module : BERT based model as this has transfer learning we have fine tuned it for grammar correction of sentences
• Patent: https://uspto.report/patent/app/20210097062