LALKRISHNA H. TRIVEDI

Dynamic Data Scientist with a proven track record at Sagility India, specializing in advanced analytics and machine learning to enhance safety and decision-making. Achieved 89% accuracy in real-time surveillance analytics using ResNet50, demonstrating expertise in Python and leadership. Instrumental in driving advancements in BI applications and data-driven strategies through innovative AI solutions. Aiming to leverage analytical skills to further impact organizational performance.

• Spirituality
• Meditation
• Yoga
• Table Tennis
• Swimming
• Acting/Drama
• Writer

• Got promoted twice in 3 year during tenure at Sagility through stellar performance and innovations.
• Led KM portal initiative at Entomo.
• Had been part of core team for ISMS/ISO9001 certification at Entomo.
• Have been Class Representative during M.Tech
• Have been Placement coordinator during B.Tech.

• Mastering Git & GitHub Program
• Complete Data Science Boot camp 2024 - Udemy
• AWS cloud practitioner

SensAI - NLP Rule Based Audit Analytics Solution (2.5 years, At Sagility)

• Domain: NLP, Web App Development
• Client: Blue Shield California (3 LOBs), Talispoint
• Tools & Tech: Spacy, Pandas, Fuzzywuzzy, NER, Django, Python, HTML, CSS, MySQL, Redis, regular expression, pyparser, Dify, LLMs (Llama)
• Description: SensAI is a multi purpose application offers wide varieties of features for data scientists and clients. From client perspective it offers an one stop audit page for auditing healthcare insurance customer care calls (agent-customer interaction) data to improve customer experience based on various client specific parameters such as whether agent showed empathy during call or whether agent greeted the customer, reason for calling, authorization etc. On the data scientist side, SensAI offers admin pages such as measures to define and maintain client specific NLP rules with unbounded hierarchy and associations, NLP Planner, ETL, Scheduler, User/Role management, Tenant management, Data Source, Data Engine, etc. Different types of NLP rules (rule based,ML/DL based, LLM based) can be created configured through measures page. NLP core engine is built on fuzzy-wuzzy with token sort ratio to match patterns defined by client specific rules and multi valued logic to resolve inter-dependency of rules. This NLP engine works on custom built vector representation of call transcription data for all ngram combinations possible. Batch processing and parallelization was used to compute large data in limited timeframe. NLP core engine generates predictions based on defined parameter rules with evidences (matched keywords/phrases) along with location in time in call transcription data which then becomes visible in Audit page. Recent work on creating dify pipelines for client specific rules with prompt engineering in Llama 3.2 8B model for SensAI 2.0 microservice architecture. We achieved stable accuracy above 80% for most of the client rules.

CRM IDB Information Validation through NER, RE, POS (6 Months, At Sagility)

• Domain: NLP, Web App Development
• Client: Talispoint
• Tools & Tech: , Spacy NER, usaddress, re, Spacy PoS, MySQL, MSSQL Connector
• Description: From call transcription data different client specific details such as provider names, group names, customer name, address, phone number, zip code, city etc was extracted with combination of regular expressions, Spacy PoS and NER which was then validated against queried data stored on remote client CRM microsoft Dynamics365 database. Achieved 70% accuracy for NER IDB extraction data.

Dynamic Word Cloud Generation (6 Months, At Sagility)

• Domain: NLP
• Client: Humana, LabCorp, Blue Shield California
• Tools & Tech: Spacy, Pandas, NLTK, TFIDF, Clustering, Sentiment Analysis
• Description: Based on daily client call volume transcription data, one month call transcription data used for training discarding previous days data on daily basis. Based on certain threshold 30 top most important words extracted for each positive, neutral, negative sentiment using TFIDF vectorization scores.

Automated Keyword Generation (3 Months, At Saglity)

• Domain: NLP
• Tools & Tech: Spacy, Pandas, NLTK, TFIDF, Clustering, BERT Topic modelling, LDA, ELBOW method
• Description: This was a self initiated project to solve and reduce contextual dependency of obtaining keywords (churning) from SMEs for client specific AQEP audit parameters. This was implemented taking one month worth of call transcription data (~90000) and performed statistical analysis related to average n-gram for keywords, min-max range of keywords for agents, customer its corelation with call length and number of segments. Then Glove, word2vec embedding/vectorization techniques were used for K - Means clustering. DBScan was used to store vectorized data for processing. using elbow method obtained optimal number of clusters mainly 6 group of keywords. However, objective was to obtain as many as non redundant cluster of keywords possible and assign tags as per AQEP parameters to which they were close. In the end TFIDF was used to extract Top-N meaningful keywords per cluster. Alternatively LDA / BERT topic modelling approach was also tried out.

Multi Level Multi Class Classification of Survey Verbatim Data (3 Months)

• Domain: Machine Learning
• Client: Humana (PPI)
• Tools & Tech: GridSearchCV, Decision Tree, Logistic Regression, SVC, Clustering, BERT, roBERTa, DistiBERT, Llama
• Description: The object of this project was to automate manual process of classifying Humana PPI customer survey verbatim data into 5 hierarchical target variables. Data consisted attributes such as date, call id, UUI, agent id, survey question from 1 to 5 namely Q1 to Q5 etc., where Q5 was free text field containing comments from customers. Besides 5 target classes were perspective (mainly sentiment of text), NPS reason 1 to 4 as hierarchy. But data was imbalanced with limited data points specially for last two hierarchical target variables NPS reason 3 and 4. So solution was created for only initial 3 target variables. With Greed Search CV tried various classification models such as logistic regression, decision tree, SVC among which Logistic regression was best fit with 75% accuracy for perspective class, 69% for NPS reason 1 and 55% NPS reason 2. After then, DL approach was also implemented trained fine tuned BERT model and its variant and llama. With BERT, accuracy was improved upto 89% for perspective class , 77% for NPS reason 1, and 75% for NPS reason 2.