ALI NASIR

Project: WinSupply- Agentic QTO Automation platform

Impact: To automate Bill of Quantities (QTO) generation from construction drawings using agentic AI workflows, improving classification accuracy and reducing manual validation through scalable, multi-stage architecture.

• Took ownership of MVP architecture; implemented YOLO + MaskRCNN baseline achieving 67% accuracy and establishing the first measurable benchmark.
• Redesigned into a 2-stage pipeline (YOLO → Structured Preprocessing → Siamese Metric Learning), improving accuracy to ~80% and reducing false positives.
• Architected an agentic AI-driven QTO platform using LangGraph multi-agent orchestration for task decomposition and structured BoQ generation.
• Integrated LLM extraction, PyMuPDF parsing, and computer vision symbol detection into a unified inference workflow.
• Implemented Kafka-based real-time feedback enabling interrupt/resume workflows, distributed consumers, confidence scoring, and decoupled ML retraining.

Project: PartsSource- Azure based Intelligent Doc Parsing

Impact: To reduce manual review and improve QC accuracy by building hybrid AI pipelines for MEL ID extraction and enterprise document validation using deterministic and generative methods.

• Led end-to-end Azure-based document ingestion and MEL ID extraction using Azure Document Intelligence and GPT-based extraction.
• Built hybrid validation framework combining fuzzy matching, Excel-based deterministic checks, and RAG-driven contextual extraction.
• Developed automated QC pipelines across 11 KPIs, improving accuracy to ~95% and reducing false positives across 3 REMI/RSA metrics.
• Mentored 3 team members and drove weekly L1/L2 QC alignment calls with client, ensuring data-driven corrective action tracking and transparent confidence reporting.

Project: Smart Doubt Recommendations

Impact: To reduce work of SMEs by providing a funnel to answer doubts asked by the students in any video lectures.

• Developed clusters of doubts based on contextual meaning for 200 chapters using Ada-003 embeddings
• Integrated a funnel of doubt classification models built to achieve an overall F-1 score of 90% and delivered API.
• NCERT data is chunked and inserted into Vector Store (AstraDB) using Langchain and Cassandra.
• Integrated AstraDB by DataStax for our RAG application and built the appropriate prompt which considers subject, exam and chapter metadata to extract the top 5 relevant document from the RAG using “mmr”.
• Extracted documents are fed to LLM (Azure OpenAI GPT 3.5) as context and the result is obtained.

Project: Personalized Test-series using Adaptive Learning: User-Engagement through ‘Infinite Practice’

Impact: To give personalized test questions to every user in ratios of Easy, medium and hard category depending upon his/her calculated knowledge using statistical modelling and engagement.

• Conceptualized state-of-art and innovative solutions, item response theory (IRT) to estimate the ability level of a person based on their performance on a test using difficulty, discrimination, speed and guessing parameters.
• Formulated an algorithm from scratch for the 3Parameter Logistic model (a statistical model) to fit parameters
• Created 2 APIs using FastAPI to return questions & optimized query to reduce latency from 40 secs to 530 ms.

Project- Topper Identification Using Academic Score (PATENT PROJECT)

Impact: Personalized Engagement Scores nurturing Potential Toppers using their test and level of engagement.

• Performed Queries to generate 41 KPIs from 60 raw parameters and trained Regression model for around 37K users of JEE With an RMSE: 0.11 and 20k users of NEET exam with an RMSE 0.09.

Project: VGyan (Doubt Solving through Ai)

Impact: Built an automatic doubt solving chatbot for Ai Live sessions to reduce assistance of teachers.

• Automated the process of classifying academic doubts & providing relevant answers for each subject and grade.
• Created sentence embeddings using distilroberta and BERT-transformer, built a classifier with 92% accuracy.
• Built a regressor with r2 score 0.978 to predict the number of doubts asked in a respective session.

Project: Post Session Comments

Impact: Developed an ETL pipeline to get dispositions &sentiments of students’ comments to automate work of Students’ Account Manager in designating the students’ issues received after a session on platform

• Performed multi-class classification on the text-comments with an accuracy of 81%.
• Created and deployed a cronjob-script from scratch in production using Apache-airflow

• Developed a route optimization solution integrated with land suitability analysis for Pune district using geospatial data modeling.
• Evaluated multiple ML models for supervised LULC classification (achieving 93% accuracy with XGBoost) and designed a CNN architecture outperforming baseline ImageNet-based classifiers.