Bridging data, research, and real-world problem solving through experimentation, analytical rigor, and practical implementation across domains.
My approach to applied research lies at the intersection of theory and practice. I focus on understanding real-world problems deeply, formulating research-driven questions, and translating analytical outcomes into actionable insights.
I strongly believe that meaningful innovation emerges when empirical data, methodological rigor, and contextual understanding are combined to support decision-making, policy formulation, and system-level improvements.
A curated selection of completed projects demonstrating applied research, analytical modeling, and real-world problem-solving across AI, machine learning, computer vision, NLP, and geospatial analytics.
Objective: To eliminate repetitive manual data downloading processes by designing a scalable, real-time automation system aligned with dynamic client requirements.
Methodology: Python-based automation architecture, command-controlled execution, real-time processing, structured logging, automated error detection, and self-recovery mechanisms.
Outcome: Delivered a stable, scalable, and low-maintenance data automation framework that reduced manual effort, minimized failures, and ensured accurate, timely client data delivery.
Objective: To automate structured data extraction from unstructured, multi-language PDF documents for analysis-ready outputs.
Methodology: Tesseract OCR integration, multi-language text recognition (Hindi, Marathi, English), NLP-driven keyword extraction, data preprocessing, and automated Excel/CSV export pipeline.
Outcome: Transformed complex PDFs into clean, structured datasets, significantly reducing manual document review effort while improving extraction accuracy and reliability.
Objective: To demonstrate accelerated neural network experimentation using modern ML frameworks and hardware-optimized computation.
Methodology: TensorFlow-based ANN implementation, GPU-enabled training, iterative model experimentation, and performance documentation across configurations.
Outcome: Showcased rapid model training and optimization, highlighting efficiency gains from hardware acceleration and bridging theoretical AI concepts with practical implementation.
Objective: To develop an intelligent conversational system that automates routine customer interactions and improves response efficiency.
Methodology: NLP-based intent recognition, supervised ML training on conversational datasets, platform integration (web/social), fallback logic design, and iterative model refinement.
Outcome: Achieved ~90% response accuracy, reduced response time by ~50%, and delivered a scalable, reliable chatbot solution for high-volume user engagement.
Objective: To identify key factors influencing accident severity and develop predictive models for early risk identification.
Methodology: Feature engineering, supervised ML models, cross-validation, and performance evaluation.
Outcome: Improved identification of high-risk scenarios supporting targeted safety interventions.
Objective: To support data-driven decision-making by analyzing behavioral and transactional patterns.
Methodology: Statistical analysis, predictive modeling, and scenario-based evaluation.
Outcome: Enhanced risk identification and optimization of operational strategies.
Objective: To build an automated AI-powered privacy protection solution capable of detecting and anonymizing sensitive visual identifiers — vehicle number plates and human faces — from large volumes of accident investigation and traffic imagery without manual intervention.
Methodology: YOLO-based object detection models were trained and integrated to identify number plates and facial regions across diverse image conditions. OpenCV-driven blur processing pipelines were applied post-detection to ensure consistent anonymization. The final system was packaged into a standalone desktop executable using PyInstaller with a Tkinter-based GUI, enabling non-technical operations staff to process bulk image batches independently.
Outcome: Delivered a production-ready desktop privacy tool that significantly accelerated image processing workflows, ensured regulatory compliance for sensitive data handling, and eliminated manual blurring effort across thousands of accident investigation images.
Objective: To develop an intelligent, computer vision-based traffic analytics platform capable of estimating vehicle speeds, classifying vehicle types, performing traffic counts, and flagging overspeed instances from road surveillance video feeds.
Methodology: YOLO-based vehicle detection and multi-object tracking algorithms were implemented to identify and follow vehicle trajectories across video frames. Calibration parameters were applied to compute real-world speed values from pixel displacement. Vehicle classification logic segmented traffic by type, and automated Excel-based reporting pipelines exported analytics outputs for engineering review.
Outcome: Produced a validated traffic analytics platform delivering accurate speed estimates, classified traffic counts, and overspeed detection reports — enabling data-driven traffic management decisions from existing surveillance infrastructure.
Objective: To develop an intelligent, real-time traffic monitoring system capable of automatically detecting helmet violations, triple riding, and other road rule violations from surveillance video, with automated evidence capture for enforcement use.
Methodology: YOLO-based deep learning models were trained and optimized to detect multiple violation types simultaneously across variable lighting and traffic conditions. OpenCV-powered frame analysis pipelines processed video streams continuously, triggering automated screenshot capture upon violation detection. Structured evidence reporting workflows organized captured frames with timestamps, location metadata, and violation classifications.
Outcome: Delivered a scalable AI-driven violation detection system that enabled automated, continuous monitoring of traffic rule compliance, significantly reducing manual observation requirements while providing structured, camera-captured evidence for enforcement operations.
Objective: To develop a real-time, computer vision-based driver monitoring solution capable of detecting distracted driving behaviors — including mobile phone usage, drowsiness, and unsafe driving patterns — to support road safety research and fleet safety programs.
Methodology: YOLO-based detection models were trained on annotated driver behavior datasets to identify key risk indicators across variable in-cabin lighting and camera angles. TensorFlow and OpenCV-powered real-time monitoring workflows processed video streams continuously, applying multi-class behavior classification logic. Automated alert generation systems triggered notifications upon confirmed high-risk behavior detection, and validation testing quantified detection accuracy across diverse driver and vehicle conditions.
Outcome: Developed and fully validated a real-time driver behavior monitoring system that demonstrated strong detection accuracy across target behavior classes, establishing a technically complete and research-ready platform. (Project subsequently discontinued at organizational level post-development.)
Objective: To build a computer vision solution capable of automatically identifying road surface defects — including potholes, cracks, and structural damage — from video and image datasets, supporting systematic road condition monitoring at scale.
Methodology: YOLO-based object detection and CNN-based image classification models were trained on annotated road surface datasets encompassing diverse defect types, lighting conditions, and pavement materials. OpenCV preprocessing pipelines standardized input imagery for consistent model inference. GIS-based visualization modules mapped detected defect locations geospatially, enabling spatial analysis of road deterioration patterns. Automated defect reporting workflows generated structured condition assessment outputs.
Outcome: Delivered a scalable, automated road condition monitoring system that replaced manual visual inspection workflows with AI-driven defect detection, enabling faster, more comprehensive road surface assessments and providing structured geospatial defect data to inform maintenance prioritization.
Objective: To develop an intelligent vehicle trajectory analytics system capable of analyzing movement patterns, lane-changing behavior, and turning movements from traffic video datasets for use in traffic engineering studies and safety assessments.
Methodology: YOLO-based vehicle detection was combined with OpenCV multi-object tracking algorithms to maintain consistent vehicle identities across video frames. Trajectory computation logic derived movement paths, lane positions, speed profiles, and turning movement counts from tracked coordinate sequences. GIS tools integrated trajectory outputs onto map-referenced visualizations, and automated traffic behavior reporting modules generated engineering-standard outputs for intersection and corridor analysis.
Outcome: Produced a validated vehicle trajectory analytics platform that delivered detailed movement behavior data from standard traffic video inputs, enabling engineers to conduct turning movement studies, lane behavior assessments, and conflict zone analyses without manual video review.
Objective: To develop a deep learning image classification system capable of automatically categorizing accident images by collision type and severity level, enabling rapid, scalable visual evidence analysis across large investigation image libraries.
Methodology: CNN-based deep learning models were designed and trained using TensorFlow and Keras on annotated accident image datasets spanning diverse collision types, vehicle categories, and damage severity levels. OpenCV-based image preprocessing pipelines standardized input dimensions, color spaces, and quality parameters to optimize model inference consistency. Classification performance was iteratively improved through augmentation strategies, transfer learning, and hyperparameter optimization.
Outcome: Delivered a validated deep learning image classification system that accurately categorized accident images at scale, reducing manual visual review requirements and enabling structured analytical tagging of large investigation image datasets for downstream research and reporting use.
Objective: To develop a video analytics platform capable of automatically extracting key frames, vehicle information, incident timestamps, and structured evidence metadata from large surveillance video libraries for crash investigation and legal documentation purposes.
Methodology: FFmpeg-based video processing pipelines handled bulk ingestion and format normalization across diverse surveillance footage sources. YOLO-powered object detection identified and classified vehicles within extracted frames, while OpenCV-based temporal analysis algorithms detected incident-relevant visual events for targeted key frame extraction. Automated metadata indexing pipelines organized extracted evidence by vehicle type, timestamp, and incident classification.
Outcome: Produced a scalable video evidence processing platform that transformed time-intensive manual footage review into an automated extraction workflow, delivering organized, metadata-rich evidence packages that accelerated crash investigation timelines and improved the completeness of surveillance-sourced documentation.
Objective: To develop an intelligent traffic analytics platform capable of automated vehicle type classification, traffic exposure estimation, and traffic composition analysis from road surveillance inputs — supporting engineering studies and road safety exposure assessments.
Methodology: YOLO-based deep learning models were trained to classify vehicles across multiple categories including two-wheelers, three-wheelers, light vehicles, heavy vehicles, and non-motorized transport. OpenCV-powered video processing pipelines computed directional traffic volumes and vehicle type distributions across observation periods. Exposure analysis algorithms integrated volume counts with road segment data to derive traffic exposure metrics.
Outcome: Delivered a validated AI-based traffic classification and exposure analytics platform that produced accurate, category-level traffic composition data from surveillance inputs — providing transport engineers and road safety researchers with reliable exposure metrics essential for crash rate analysis, risk modelling, and infrastructure safety assessments.
Objective: To automate the extraction, translation, and structural transformation of Hindi-language government and crash-related PDFs into analysis-ready English datasets, eliminating manual translation and data entry bottlenecks.
Methodology: OCR engines were integrated to extract Hindi text from scanned and digital PDFs. Extracted content was routed through Google and GROQ Translation APIs for accurate multilingual conversion. NLP-driven text cleaning and structuring modules parsed translated outputs into field-mapped Excel datasets with automated quality validation checks.
Outcome: Transformed a time-intensive manual translation and structuring workflow into a fully automated pipeline, delivering clean, structured bilingual datasets that accelerated downstream analysis and reduced document processing time substantially.
Objective: To develop a large-scale OCR automation system that intelligently extracts structured accident information from scanned PDFs and handwritten forms, drastically reducing manual data entry operations across national-level road safety datasets.
Methodology: Tesseract OCR was integrated with OpenCV-based image preprocessing to handle variable-quality scanned documents. Custom extraction and field-mapping pipelines converted unstructured form content into validated tabular datasets. Power BI dashboards were connected to extraction outputs for real-time analytical visibility. The system was packaged as a deployable executable for operational team use.
Outcome: Delivered a scalable OCR automation platform that processed thousands of accident records with high accuracy, reduced data entry effort by a significant margin, and enabled near-real-time analytical reporting from previously unstructured documentary sources.
Objective: To architect a fully integrated analytical ecosystem that seamlessly connects OCR extraction, multilingual translation, data validation, and Power BI dashboard generation into a single automated pipeline.
Methodology: Tesseract OCR and OpenCV preprocessing modules handled bulk document extraction across variable-quality inputs. Translation and NLP layers processed multilingual content into structured datasets. Automated validation pipelines ensured data quality before routing outputs into Power BI dashboard generation workflows. End-to-end orchestration logic managed pipeline sequencing, error handling, and output delivery.
Outcome: Delivered a production-grade, end-to-end automated analytical ecosystem that transformed raw scanned documents into live dashboards with minimal human intervention, establishing a scalable foundation for data-driven crash analytics operations.
Objective: To develop an intelligent OCR-based extraction platform capable of automatically parsing structured crash-relevant information from First Information Reports (FIRs) and accident investigation documents at scale.
Methodology: Tesseract OCR integrated with LayoutLM-based document understanding models was deployed to handle the structural complexity of FIR forms, including multi-column layouts, handwritten annotations, and varied document formats. OpenCV preprocessing pipelines improved input image quality for OCR accuracy. NLP libraries drove named entity recognition and field-level information extraction, with automated validation mechanisms verifying output completeness and logical consistency.
Outcome: Delivered an intelligent FIR extraction system that transformed unstructured legal and investigation documents into structured analytical datasets, dramatically reducing manual data entry timelines and enabling faster, more comprehensive integration of FIR-sourced intelligence into national crash databases.
Objective: To develop an AI-powered, generative summarization engine capable of automatically producing daily accident summaries, district-wise crash reports, trend narratives, and scene-level analytical insights directly from raw accident datasets and reports.
Methodology: LLM integration pipelines were built using OpenAI, Gemini, and GROQ APIs, orchestrated through LangChain-based workflow automation. Prompt engineering strategies were designed to produce contextually accurate, domain-specific accident summaries with controlled output structure. Python and Pandas preprocessing modules prepared structured and semi-structured datasets for LLM consumption. Automated report generation pipelines scheduled and delivered formatted analytical summaries to operational and reporting stakeholders.
Outcome: Delivered a production-ready LLM-powered reporting engine that reduced manual crash report authoring effort significantly, improved analytical reporting consistency, and enabled daily automated generation of multi-level summaries — from scene-level narratives to district-wide trend analyses — across national road safety programs.
Objective: To develop an NLP-powered text classification engine capable of automatically categorizing accident reports by severity level, crash type, road type, and collision pattern — replacing manual tagging with an intelligent, scalable automated system.
Methodology: Accident report text data was preprocessed through NLP pipelines involving tokenization, stopword removal, and domain-specific feature extraction using SpaCy. Transformer-based and Scikit-learn classification models were trained and evaluated on labeled crash report datasets across multiple classification dimensions. Automated structured tagging pipelines applied model outputs to assign validated category labels, and model performance was assessed through precision, recall, and F1 metrics across all target classes.
Outcome: Delivered a production-ready NLP classification engine that automated multi-dimensional accident report categorization at scale, significantly reducing manual tagging workloads while improving classification consistency and enabling faster downstream analytics on structured crash report attributes.
Objective: To develop an intelligent duplicate detection platform capable of identifying redundant crash records across national databases using text similarity analysis, location matching, and metadata comparison — preserving data integrity at scale.
Methodology: Multi-layer deduplication logic was engineered combining NLP-based fuzzy text matching, GPS coordinate proximity comparison, and structured metadata field alignment. Fuzzy matching algorithms computed similarity scores across free-text crash descriptions, while SQL-based pipelines cross-referenced structured attributes including date, time, location, and vehicle involvement. Automated validation workflows flagged high-confidence duplicates for review and generated audit trails of detected redundancies.
Outcome: Deployed a robust duplicate detection system that identified and flagged previously undetected redundant records across large-scale crash databases, improving dataset uniqueness and analytical reliability — directly enhancing the quality of national road safety intelligence outputs.
Objective: To build an AI-powered conversational interface capable of answering natural language queries about crash datasets, generating on-demand summaries, and retrieving targeted analytical insights from large-scale accident databases.
Methodology: A LangChain-orchestrated chatbot architecture was developed integrating OpenAI API language models with SQL-backed data retrieval pipelines. NLP-driven query interpretation modules translated conversational user inputs into structured database queries. Conversational analytics workflows enabled multi-turn dialogue for iterative data exploration, and response generation logic synthesized retrieved data into natural language summaries and analytical narratives tailored to query intent.
Outcome: Successfully developed and validated a fully functional AI chatbot delivering accurate, context-aware crash data responses through a conversational interface — demonstrating practical LLM-to-database integration for road safety intelligence. (Project subsequently discontinued at organizational level post-development.)
Objective: To develop a generative AI platform capable of automatically producing comprehensive road safety audit reports, crash summaries, and analytical narratives — replacing manual report authoring with intelligent, prompt-engineered LLM workflows.
Methodology: OpenAI and Gemini API-powered generative pipelines were orchestrated through LangChain to ingest structured crash datasets and produce contextually accurate, domain-specific report content. Advanced prompt engineering strategies were designed to control output structure, analytical depth, and domain terminology consistency across report types. Automated pipeline workflows managed data ingestion, LLM API calls, output formatting, and final document assembly.
Outcome: Delivered a production-ready generative AI reporting system that automated the end-to-end production of road safety audit documents and analytical narratives, substantially reducing report authoring time, improving consistency across deliverables, and enabling analytical teams to scale reporting output without proportional increases in manual effort.
Objective: To build a predictive analytics platform capable of proactively identifying accident-prone road locations by integrating traffic volume, vehicle speed, road geometry, weather patterns, and historical crash data into machine learning risk models.
Methodology: Multi-source datasets were engineered into consolidated feature matrices combining spatial, temporal, environmental, and traffic parameters. XGBoost and Scikit-learn-based classification and regression models were trained, validated, and optimized to predict black spot risk scores at road segment level. GIS analytics modules visualized predicted risk zones on interactive maps, and automated reporting workflows generated prioritized intervention recommendations.
Outcome: Delivered a predictive black spot intelligence system that moved road safety planning from reactive to proactive, enabling authorities to allocate infrastructure improvement resources toward highest-risk locations before further crashes occurred.
Objective: To develop a machine learning-based predictive system capable of automatically estimating accident severity levels from crash parameters — including vehicle type, speed, road conditions, and weather — enabling faster prioritization of emergency response and road safety interventions.
Methodology: Multi-variable crash datasets were preprocessed and feature-engineered to extract the most predictive attributes influencing severity outcomes. Scikit-learn and XGBoost classification models were trained, cross-validated, and hyperparameter-tuned to optimize prediction accuracy across severity classes. Automated severity classification workflows were integrated with reporting pipelines that generated structured analytical outputs, including feature importance visualizations and model performance documentation.
Outcome: Delivered a validated severity prediction engine that accurately classified crash outcomes from input parameters, providing road safety teams with an intelligent triage tool that improved response planning, resource allocation, and post-crash analytical reporting efficiency.
Objective: To develop an AI-driven traffic forecasting platform capable of predicting future congestion levels and traffic flow patterns from historical datasets, supporting proactive traffic management and infrastructure planning.
Methodology: Historical traffic volume, speed, and temporal datasets were engineered into time-series feature matrices. LSTM-based deep learning models were implemented using TensorFlow to capture long-range temporal dependencies in traffic flow patterns. Forecasting modules generated short- and medium-term traffic predictions with confidence intervals. Power BI-integrated dashboards visualized predicted flow trends alongside historical benchmarks.
Outcome: Delivered a validated traffic flow prediction platform that produced accurate, horizon-based congestion forecasts, enabling traffic management teams to shift from reactive to predictive operational planning with data-driven evidence for infrastructure decision-making.
Objective: To develop an AI-powered recommendation system capable of intelligently suggesting missing accident attributes and probable crash parameters from historical patterns, reducing incomplete record rates in national crash databases.
Methodology: Historical accident datasets were analyzed to identify statistical relationships and co-occurrence patterns across crash attributes. Machine learning models — including collaborative filtering and pattern-based classification pipelines — were trained to predict probable values for missing fields based on known contextual attributes. Intelligent suggestion pipelines integrated with data entry workflows to surface real-time recommendations to operators, with confidence scoring applied to prioritize high-reliability suggestions.
Outcome: Delivered an intelligent data completion assistant that reduced missing attribute rates in crash records, improved overall dataset completeness for analytical purposes, and provided operators with evidence-based recommendations that accelerated quality data entry without sacrificing accuracy.
Objective: To develop a machine learning system capable of identifying recurring accident patterns and emerging crash trends from large-scale historical datasets, enabling proactive, evidence-based road safety policy and infrastructure interventions.
Methodology: Large-scale historical crash datasets were prepared through multi-stage feature engineering pipelines extracting spatial, temporal, vehicle, environmental, and road geometry dimensions. Clustering algorithms including K-Means and DBSCAN were applied to surface recurring accident groupings and spatial risk clusters. Anomaly detection pipelines identified statistically unusual crash concentrations warranting targeted investigation.
Outcome: Delivered a scalable crash pattern intelligence platform that revealed statistically significant recurring accident patterns and emerging risk trends across national datasets — providing road safety planners with actionable analytical evidence to guide targeted interventions, policy design, and infrastructure investment prioritization.
Objective: To develop an automated GIS-based analytical platform capable of identifying high-risk crash locations from GPS accident coordinates using intelligent hotspot clustering, supporting evidence-based road safety interventions.
Methodology: GPS coordinate data from accident records was ingested and processed using Python-based geospatial pipelines. Hotspot clustering algorithms — including proximity-based greedy clustering — grouped spatially related accidents to identify black spot zones. Folium and Plotly-driven HTML visualization outputs rendered interactive maps with severity overlays, crash densities, and location markers for field and audit use.
Outcome: Produced an operational black spot identification platform that translated raw coordinate datasets into actionable geospatial intelligence, enabling road safety teams to prioritize intervention sites based on data-driven risk rankings.
Objective: To build an interactive, web-based crash visualization platform for plotting accident locations and black spots on dynamic maps, providing road safety auditors with a spatial investigation and reporting tool accessible via browser.
Methodology: A Flask-based backend served crash coordinate data through RESTful APIs to a Leaflet.js-powered frontend. Interactive map layers were developed to display individual accident markers, black spot clusters, road category differentiation, and severity filters. HTML, CSS, and JavaScript components delivered a responsive interface optimized for both desktop and field tablet use across audit deployments.
Outcome: Delivered a fully functional web mapping platform that replaced static report-based crash reviews with dynamic spatial investigation workflows, enabling road safety audit teams to explore, filter, and interpret accident geographies in real time.
Objective: To develop a GIS-enabled analytical platform capable of computing and assigning evidence-based road risk scores to individual road segments using crash density, vehicle speeds, geometric parameters, and environmental conditions.
Methodology: Multi-source spatial datasets encompassing historical crash records, traffic volume data, road geometry attributes, and environmental factors were integrated into a unified GIS analytical framework. Risk scoring algorithms weighted and combined input variables to produce segment-level risk indices. Folium and Plotly-powered interactive dashboards rendered scored road networks as color-graduated risk maps.
Outcome: Delivered a comprehensive road risk scoring platform that translated diverse data inputs into standardized, actionable risk intelligence — equipping road safety planners with an evidence-based spatial tool for systematic identification and prioritization of high-risk road sections.
Objective: To build an automated heatmap analytics platform capable of generating dynamic traffic density and congestion visualizations from traffic camera datasets, providing spatial traffic intelligence for road network management and planning.
Methodology: OpenCV-based video frame analysis pipelines computed vehicle density metrics across defined spatial zones within camera coverage areas. GIS mapping frameworks ingested density outputs and rendered georeferenced heatmap layers using Folium and Plotly visualization libraries. Temporal aggregation logic built time-series density profiles, enabling peak hour and corridor-level congestion pattern analysis.
Outcome: Delivered a dynamic traffic density heatmap platform that converted raw camera footage into spatial congestion intelligence, enabling traffic management teams to visualize network-wide density patterns, identify chronic congestion hotspots, and make evidence-based decisions on signal timing and road capacity management.
Objective: To develop a fully automated geospatial platform capable of dynamically mapping accident locations, generating analytical GIS visualizations, and delivering interactive map-based outputs that support road safety investigations and audit workflows.
Methodology: GPS coordinate data from multi-source accident datasets was ingested through automated Python ETL pipelines. Folium and Plotly-based GIS visualization modules rendered dynamic, layer-controlled interactive maps with accident markers, severity overlays, road network context, and time-filter capabilities. Automated map generation workflows processed new coordinate inputs on schedule, refreshing geospatial outputs without manual intervention.
Outcome: Delivered a dynamic, automated geospatial mapping platform that transformed raw coordinate datasets into rich, interactive accident location intelligence — enabling road safety teams to conduct spatial investigations, track temporal accident trends, and generate map-based audit evidence efficiently and at scale.
Objective: To perform comprehensive large-scale crash data analysis and develop interactive Power BI dashboards delivering trend analysis, geospatial maps, severity distributions, and KPI insights for national road safety programs.
Methodology: SQL and Python-based data engineering workflows cleaned, structured, and aggregated crash datasets spanning multiple states and years. Analytical frameworks were designed to surface time-series trends, severity distributions, road-type breakdowns, and geographic hotspot patterns. Power BI dashboards were developed with dynamic filtering, drill-through capabilities, and map-based visualizations for multi-level stakeholder review.
Outcome: Delivered a comprehensive analytical reporting system that transformed raw crash records into decision-ready intelligence, enabling program managers and research teams to identify priority intervention areas and track road safety KPIs effectively.
Objective: To develop a data-driven traffic congestion analysis platform capable of measuring queue lengths, vehicle density, congestion levels, and Level of Service classifications to support traffic engineering assessments and capacity planning.
Methodology: OpenCV-based video processing algorithms were developed to track vehicle positions, measure queue formations, and compute density metrics across road segments. Automated congestion classification logic mapped measured parameters to standard LOS categories. Pandas-powered analytical pipelines aggregated time-series traffic metrics, and Matplotlib-generated reporting outputs delivered structured traffic engineering documentation.
Outcome: Produced a validated traffic congestion analytics platform that delivered objective, data-driven LOS assessments from video inputs, supporting traffic engineers with quantitative evidence for capacity improvement planning and intersection management decisions.
Objective: To develop a live monitoring dashboard capable of tracking incoming crash records in real time, generating automated alerts for critical incidents, and visualizing dynamic accident statistics for operational road safety management.
Methodology: Real-time data pipeline integrations connected live crash data feeds to a centralized monitoring backend via APIs and SQL-powered data streams. Python-based alert logic applied configurable thresholds across severity, location, and frequency parameters to trigger automated notifications for high-priority incidents. Power BI dashboard components delivered live-refreshing visualizations of crash volumes, geographic distributions, severity breakdowns, and time-series trends.
Outcome: Deployed a production-grade real-time monitoring platform that gave road safety operations teams live visibility into crash activity, automated critical incident alerting, and a dynamic analytical dashboard — enabling faster operational response and informed real-time decision-making across national road safety monitoring programs.
Objective: To automate the generation of accurate 2D and 3D accident scene reconstructions from GPS coordinate inputs, replacing time-consuming manual diagram creation in crash investigation workflows.
Methodology: Coordinate-based plotting logic was developed using Plotly and Matplotlib to render spatial crash scene layouts in both two-dimensional and three-dimensional formats. Automated export workflows generated investigation-ready outputs in image and PDF formats. Scene elements including vehicle positions, impact zones, and road geometry were represented with configurable visual parameters.
Outcome: Delivered a rapid, coordinate-driven crash reconstruction tool that standardized scene visualization outputs, reduced investigation preparation time, and improved the clarity and consistency of accident documentation across investigation teams.
Objective: To eliminate repetitive manual data downloading processes by engineering a robust, scheduled web automation bot capable of extracting and organizing lakhs of records overnight without human intervention.
Methodology: Selenium and BeautifulSoup-based scraping workflows were developed to navigate target portals, authenticate sessions, and systematically retrieve large data volumes. Scheduling automation frameworks managed overnight execution cycles. Intelligent retry mechanisms handled connection failures and dynamic page structures, while organized storage pipelines maintained data integrity across bulk downloads.
Outcome: Deployed a reliable, self-managing automation bot that executed overnight large-scale data collection runs without manual oversight, ensuring accurate and timely availability of critical datasets for downstream analytical operations.
Objective: To automate the validation and metadata correction of accident scenario files, replacing a manual, error-prone verification process with an intelligent, self-correcting automation workflow.
Methodology: OCR and OpenCV-based metadata extraction modules were developed to read embedded attributes from accident scenario files. Automated comparison logic identified mismatches between declared and actual metadata values. Correction routines applied targeted fixes and resaved updated files automatically, with Pandas-driven reporting capturing all changes for audit purposes.
Outcome: Delivered an intelligent auto-correction tool that resolved metadata inconsistencies at scale, significantly reduced manual QC effort, and ensured standardized, validated scenario files ready for operational and analytical use.
Objective: To develop purpose-built internal software applications that automate repetitive engineering computations — including speed calculations, distance analysis, and crash parameter assessments — used daily by the operations team.
Methodology: Modular Python calculation engines were designed around specific analytical formulas and operational workflows. NumPy-powered computation modules ensured numerical accuracy across edge cases. Tkinter-based desktop interfaces provided operators with intuitive, form-driven input experiences, and automated output generation eliminated manual spreadsheet-based computation steps.
Outcome: Delivered a suite of reliable internal calculation tools that standardized analytical computations, reduced manual calculation errors, and freed engineering time for higher-value analytical activities by automating routine operational workflows.
Objective: To design and deploy a large-scale, multi-client real-time automation platform handling scheduled data scraping, duplicate detection, automated dashboard refreshes, and report generation across multiple simultaneous operational deployments.
Methodology: Selenium and API-based data acquisition modules were developed per client configuration, with centralized scheduling automation orchestrating multi-client execution cycles. SQL-backed duplicate detection logic ensured data uniqueness across ingestion pipelines. Power BI integration automated dashboard refresh sequences, while Python-based report generation modules produced structured client deliverables on configurable schedules.
Outcome: Deployed a production-grade, scalable automation platform serving multiple clients concurrently, reducing manual data operations to near zero, ensuring consistent on-time delivery of clean datasets and analytical reports, and establishing a reliable operational backbone for data-driven client programs.
Objective: To develop a real-time data quality monitoring platform that validates user data entries instantly at the point of input, preventing errors from propagating into national-level crash datasets.
Methodology: A rule-based validation engine was designed to define field-level constraints, cross-field dependencies, and logical consistency checks across accident data attributes. Python and Pandas-powered monitoring scripts evaluated entries in real time, triggering immediate on-screen alerts for incorrect or inconsistent inputs. Tkinter-based interface components provided operators with clear, actionable error notifications.
Outcome: Built a robust live QC system that intercepted data quality issues at source, significantly reduced downstream correction workloads, and improved the overall integrity and reliability of national crash database records.
Objective: To manage enterprise-scale data quality operations across national road safety datasets and develop advanced Power BI reporting systems delivering multi-dimensional charts, maps, and trend analytics for program stakeholders.
Methodology: Python and SQL-based validation frameworks were deployed to perform systematic QC checks across high-volume crash datasets, flagging anomalies, duplicates, and structural inconsistencies. Automated validation reporting workflows generated exception logs for remediation. Power BI dashboards were built with complex DAX measures, geographic map visuals, time-series analytics, and executive KPI summaries tailored to program reporting requirements.
Outcome: Established a robust enterprise QC and reporting infrastructure that ensured data integrity at scale, accelerated reporting cycles, and delivered high-quality analytical dashboards that supported national road safety program governance and decision-making.
Objective: To develop an intelligent metadata extraction and automated validation system capable of surfacing and verifying hidden crash-data attributes from files, embedded forms, and structured datasets — reducing dependency on manual QC processes at scale.
Methodology: OCR and OpenCV-based extraction modules were engineered to identify and retrieve embedded metadata attributes from diverse file types including scanned forms, structured documents, and binary data containers. Python and Pandas-powered validation logic applied domain-specific rules to verify attribute completeness, format compliance, and logical consistency. Automated correction workflows resolved identified metadata anomalies.
Outcome: Deployed a robust automated metadata intelligence system that surfaced previously inaccessible data attributes and validated their integrity systematically, significantly reducing manual QC dependency and improving the completeness and traceability of crash database records across national road safety programs.
Objective: To design and deploy a secure, enterprise-grade web portal for managing confidential crash analytics, restricted investigation data, analytical dashboards, and role-based reporting access across authorized road safety program stakeholders.
Methodology: A Flask-based web architecture was engineered with multi-tier authentication and role-based access control systems enforcing data confidentiality at user, module, and dataset levels. Integrated analytical dashboards connected Power BI and Python-generated reports within the secured portal environment. SQL-backed data management workflows handled confidential crash record storage, retrieval, and update operations.
Outcome: Delivered a fully operational, secure enterprise analytics portal that centralized confidential crash intelligence, investigation data, and reporting workflows within a controlled, access-governed environment — providing authorized stakeholders with a professional, reliable platform for sensitive road safety program management and analytical oversight.
I am particularly interested in exploring how data-driven systems can support societal challenges, policy-level decision-making, and sustainable development initiatives.
My ongoing thinking revolves around integrating spatial-temporal analytics, interpretable machine learning, and domain knowledge to build systems that are not only accurate but also transparent, explainable, and practically usable.
Beyond completed projects, I continuously reflect on emerging research directions, real-world data limitations, and ethical considerations in applied analytics and AI-driven systems.