Available for opportunities

Sarthak
Mahire

AI/ML Engineer/Software Engineer
Boston, MA, USA
LinkedInORCIDEmail
Patent FiledIEEE PublishedMS in AI @ NEU
Scroll

About

A bit about who I am and what I do

Master of Science in AI student with a curious mindset and a strong research background. Semifinalist (top 10 out of 180+ teams) in the $1M Alzheimer's Insights AI Prize for developing an agentic AI framework. Published patent holder and proven innovator with a published paper in NLP and text mining (IEEE ICERCS 2024). I excel at evidence-based decision making and problem solving — highly stress-resilient, team-oriented, and passionate about tackling complex challenges.

Top 10of 180+ teams — $1M Alzheimer's AI Prize
PatentDeepfake Audio Detection — IP India (2025)
IEEEPublished — ICERCS 2024

Experience

Where I've worked and what I've built

Data Scientist

Dassault Systèmes

June 2026 — PresentBoston, MA, USA

Dassault Systèmes is a global leader in digital transformation for engineering and manufacturing. As a Data Scientist, I develop and deploy AI/ML-powered solutions across various domains.

  • Developed and deployed a machine learning pipeline for marketing lead scoring, now in production across all Dassault Systèmes branches worldwide, making scores 60% more reliable and boosting the F1 score by 0.26.
  • Built an end-to-end AWS system to automate data cleaning, loading, scoring, score routing, and periodic model retraining, creating a self-sustaining MLOps workflow.
  • Designed professional Vue.js dashboards to visualize data history, pipeline processes, and interactions, enabling deeper data analysis across the marketing domain.

Graduate Research Assistant

VOxx Lab — Northeastern University

Sept 2025 — PresentBoston, MA

VOxx Lab pioneers health acoustics, using voice as a non-invasive biomarker for neurological and physiological health. Research integrates acoustic signal processing and AI-driven analysis to uncover vocal biomarkers that precede clinical symptoms.

  • Designed and maintained the VOxx Unified Dataset, harmonizing 500,000+ audio biomarker samples into a secure, HIPAA-compliant cloud platform for real-time ML; built unit/integration tests for reliability.
  • Built and monitored agentic AI pipelines using QWEN3, Kubernetes clusters with NVIDIA A100 GPUs, Docker microservices, Redis queuing, and MLflow for the Doppel Data and DiscoADRD projects.
  • Managed and optimized HPC resource allocation across various concurrent research workloads. Automated job scheduling and resource monitoring.

Fullstack Developer

Appit Software Solutions

August 2023 — September 2023Hyderabad, India

Global AI transformation company serving 15+ industries including Healthcare, Banking, and Retail. Appit specializes in custom software development and AI-driven business automation for enterprises.

  • Built and launched a lead generation platform (Workisy) with Node.js, React, and PostgreSQL, implementing a comprehensive CI/CD pipeline with automated testing, deployment, and monitoring using Jenkins, Docker, and AWS services.
  • Conducted a comprehensive UX case study analyzing user behavior across 500+ platform users. Optimized a WordPress event platform by identifying pain points and implementing UI/backend improvements that increased engagement by 40%.
  • Collaborated with cross-functional engineering teams of 8+ developers to architect technical solutions, participating in code reviews and pair programming sessions.

Projects

17 projects across AI/ML, data engineering, backend, and more

AI/MLFeaturedJune 2026

ML-Driven Lead Scoring System (Dassault Systèmes)

  • Diagnosed a failing lead scoring model where leads tagged “hot” converted well below expectation; scoped the problem under uncertainty by prototyping two approaches in parallel to isolate whether the issue was logic, data, or noise.
  • Built rule-based scoring (HubSpot-native, low-cost, explainable) with rules derived from SHAP analysis of ML models, and benchmarked predictive ML (XGBoost, LightGBM, CatBoost) on tabular data, weighing maintainability against accuracy.
  • Selected and fine-tuned XGBoost after it clearly outperformed; deployed to production across all branches, making scores 60% more reliable and improving the F1 score by 0.26.
PythonXGBoostLightGBMCatBoostSHAPHubSpot
AI/MLFeatured2026

Resume Tailor — LLM-Powered Resume Optimization App

  • Built a full-stack, self-hosted web app (FastAPI + SQLite backend, React + Vite + Tailwind frontend) that tailors a LaTeX resume to any job description by selecting the most relevant experiences, projects, and skills and rewriting the title, summary, and skills to match.
  • Engineered a provider-agnostic LLM layer supporting Claude (Anthropic) and any OpenAI-compatible provider (Groq, OpenAI, OpenRouter, Ollama), with API keys encrypted at rest using Fernet (AES-128-CBC + HMAC).
  • Implemented sandboxed latexmk PDF compilation fit to a page limit, an applications history dashboard, and single-container Docker deployment bundling full TeX Live; runs cross-platform on macOS, Windows, and Linux.
FastAPIReactSQLiteTailwindDockerLLMs
GitHub
AI/MLFeaturedSept 2025 — Dec 2025

Doppel Data — Alzheimer's Insights AI Prize Semifinalist

  • Advanced to top 10 semifinalists (from 180+ teams) in the $1M Alzheimer's Insights AI Prize with an innovative agentic AI framework.
  • Developed a multi-agent system addressing missingness in medical datasets by identifying digital twins across diverse cohorts, studies, and databases.
  • Engineered agents for missing data detection, digital twin matching, bias mitigation, privacy, temporal context, and real-time critique, ensuring robust and compliant data harmonization.
PythonMulti-Agent SystemsLLMsHIPAAMLflow
AI/MLFeaturedFeb 2026 — Mar 2026

Autonomous Market Gap Identifier (AMGI)

  • Built a multi-agent market research pipeline that mines complaint-heavy signals from YouTube and web articles, transcribes (Whisper fallback), cleans (spaCy), and indexes embeddings in ChromaDB for semantic retrieval.
  • Implemented complaint-density retrieval + per-gap isolated RAG verification (Tavily) and exported strict JSON/CSV outputs: unmet needs, competitors, reasoning, proposed solution, Greenlit/Redlit.
  • Benchmarked local LLMs on the Razorpay Fix My Itch dataset (50,000 datapoints); best run reached 87.5% accuracy (90% precision) with Llama3.3:70b.
PythonCrewAILangChainspaCyChromaDBWhisperOllama
GitHub
BackendFeaturedMar 2025 — May 2025

API Sentinel — Smart API Gateway & Analytics Proxy

  • Built a high-concurrency smart proxy in TypeScript/Express that intercepts, validates, and forwards requests to third-party APIs (OpenAI, Hugging Face, OpenWeatherMap) using asynchronous streams, centralizing auth injection and security policy enforcement.
  • Implemented Redis-backed Leaky Bucket rate limiting resolving quota decisions in under 2ms, and a PostgreSQL analytics engine logging per-request latency, status codes, and payload size with a real-time usage dashboard.
  • Enforced strict TypeScript interfaces as type-safe contracts over external API response schemas, eliminating runtime crashes from upstream JSON format changes.
TypeScriptNode.jsExpressRedisPostgreSQLDocker
Computer VisionFeaturedSept 2024 — Dec 2024

Open-Vocabulary 3D Scene Graph Generation from Monocular Video

  • Developed a research-grade perception pipeline reconstructing semantic 3D environments from a single monocular camera feed, eliminating the need for expensive LiDAR hardware in robotics.
  • Integrated DepthAnywhere-V2 for metric depth estimation and Grounded-SAM for open-vocabulary object detection, achieving zero-shot detection of novel objects through language-aligned point clouds.
  • Engineered dynamic 3D scene graphs with spatial relationship mapping (occlusion, proximity) for robotic path-planning, advancing autonomous navigation in complex environments.
PythonPyTorchDepthAnything-V2Grounded-SAMComputer Vision
AI/MLSept 2025 — Oct 2025

DiscoADRD: Automated Scientific Discovery

  • Developed an automated hypothesis generation and testing framework with a multi-agent pipeline (Hypothesis Generator, Data Preparation, Planner, Scientist, and Critic Agents) to test scientific hypotheses at scale.
  • Integrated QWEN/Qwen2.5-4B-Instruct LLM for agent reasoning with an iterative validation loop between Scientist and Critic Agents.
  • Tested on the NACC dataset — the system validated well-established hypotheses with citations, classified discoveries, and used RAG to ground claims in scientific literature.
PythonQWEN LLMMulti-AgentRAGNACC Dataset
ResearchFeaturedFeb 2024 — May 2024

DeepFake Audio Detection (Patent Filed)

  • Invented a deepfake audio detection system combining Bi-Directional GRU-RNN and WGAN-GP, achieving 94.2% detection accuracy on the ASVspoof2019 benchmark.
  • Achieved 25% reduction in computational complexity compared to Bi-LSTM approaches, enabling real-time audio authentication in telephony security and media forensics.
  • Filed patent with IP India (Application No. 202541122126, Dec 2025).
PythonPyTorchBi-GRUWGAN-GPASVspoof2019
Data EngineeringFeaturedJan 2024 — June 2024

Global Air Quality & Weather Analytics Pipeline

  • Engineered an end-to-end automated data pipeline ingesting real-time weather (OpenWeatherMap) and air quality data (OpenAQ), processing 10,000+ daily records across 50+ global cities.
  • Architected a cloud-native ELT pipeline using Apache Airflow orchestration, AWS S3 data lake, Snowflake data warehouse, and dbt transformations for scalable environmental analytics.
  • Built interactive Power BI dashboards visualizing global pollution trends and weather correlations, enabling data-driven environmental policy insights.
Apache AirflowSnowflakeAWS S3dbtPower BIPython
Computer VisionSept 2024 — Nov 2024

DeepFake Image Detection with CNN and Vision Transformers

  • Architected a novel hybrid deepfake detection model integrating CNN feature extraction with Vision Transformer classification, achieving 91% accuracy on image authenticity verification.
  • Designed an end-to-end pipeline with CNN spatial feature extraction, ViT patch embedding with positional encoding, and multi-head self-attention for complex contextual relationships.
  • Implemented an MLP classification head processing transformer encoder outputs, demonstrating effectiveness of the hybrid CNN-ViT approach.
PythonPyTorchCNNVision Transformer (ViT)MLP
AI/MLJan 2024 — Apr 2024

Movie Planner AI Agent

  • Developed an intelligent AI agent processing natural language invitations to automate movie/event planning, extracting key details (title, time, venue) from informal text messages.
  • Built FastAPI backend with n8n workflow orchestration, integrating OMDb API for metadata enrichment and OpenRouter LLM hosting for contextual response generation.
  • Implemented an end-to-end automation pipeline from message parsing to intelligent reply generation, enabling human-like conversational planning.
FastAPIn8nOpenRouterOMDb APIPython
BackendJune 2024 — Aug 2024

Cryptocurrency Price Alert System

  • Developed a real-time cryptocurrency price monitoring system using Django REST Framework, processing live Binance API data with Redis caching (1-minute TTL) for 100+ digital assets.
  • Implemented secure JWT-based authentication and custom threshold management, enabling users to set personalized price alerts with upper/lower bounds across multiple cryptocurrencies.
  • Engineered an asynchronous Celery task queue triggering instant email notifications when price thresholds are crossed, ensuring sub-minute alert delivery with PostgreSQL persistence.
Django RESTRedisCeleryPostgreSQLBinance APISimpleJWT
Computer VisionFeb 2024 — May 2024

DeepFake Detection with GAN-Augmented Data

  • Implemented advanced deepfake detection using GAN-generated synthetic data for training augmentation, improving classifier generalizability across diverse deepfake generation techniques.
  • Trained an adversarial GAN architecture with a Generator creating synthetic deepfakes from noise and a Discriminator learning on the Flickr-Faces-HQ dataset.
  • Enhanced the DF-40 dataset with GAN-generated samples for CNN training, achieving improved classification performance through increased dataset variety.
PythonTensorFlowGANsCNNScikit-learn
Data EngineeringSept 2023 — Dec 2023

Walmart Sales Analysis and Prediction

  • Analyzed historical Walmart sales data to forecast weekly revenue patterns, implementing an ensemble ML approach combining Linear Regression, Decision Tree, Random Forest, and XGBoost.
  • Conducted comprehensive EDA revealing sales trends across stores, seasonal patterns, and holiday impacts, optimizing feature engineering for improved demand prediction accuracy.
  • Evaluated models using R², MSE, and RMSE metrics to enhance inventory management and supply chain optimization.
PythonXGBoostScikit-learnPandasSeaborn
Data EngineeringMar 2023 — June 2023

Electricity Production Forecasting

  • Developed a time-series forecasting model predicting monthly electricity production for a 5-year horizon using 33 years of historical data (1985–2018) with advanced statistical methods.
  • Applied Augmented Dickey-Fuller stationarity testing and log transformation to prepare non-stationary data, implementing SARIMAX model optimized for seasonal electricity patterns.
  • Generated accurate 60-month production forecasts outperforming ARIMA baseline, providing energy sector stakeholders with reliable capacity planning insights.
PythonStatsmodelsSARIMAXPandasMatplotlib
AI/MLJan 2023 — Mar 2023

Predicting Sales from Advertising Data

  • Built a predictive sales model analyzing advertising expenditure impact across TV, Radio, and Newspaper channels, identifying TV advertising as the strongest sales driver through correlation analysis.
  • Performed comprehensive EDA using Plotly interactive visualizations and statistical heatmaps, revealing key advertising-sales relationships for marketing budget optimization.
  • Implemented a linear regression model achieving robust performance metrics (R², MSE, MAE), providing actionable insights for advertising spend optimization.
PythonScikit-learnPlotlyPandasJupyter
EmbeddedFeb 2022 — May 2022

Ultrasonic Distance Measurement System

  • Designed an embedded distance measurement system using 8051 microcontroller and HC-SR04 ultrasonic sensor, programming real-time distance calculations in C with sub-centimeter accuracy.
  • Implemented a complete firmware solution with USB AVR + AT89Sxx ISP programmer for microcontroller flashing, integrating sensor control and echo time processing algorithms.
  • Developed a real-time display interface with a 16x2 LCD screen showing live distance measurements.
C8051 MCUHC-SR04LCDEmbedded Systems

Skills

Technologies and tools I work with

Languages

PythonTypeScriptJavaScriptJavaC++SQLBash / Shell

ML & AI

PyTorchTensorFlowScikit-learnTransformersLLMs / RAGCNNsComputer VisionNLPLSTM / GRUPrompt EngineeringMulti-Agent Systems

Backend Development

FastAPIDjango RESTNode.js / ExpressREST APIsRedisPostgreSQLCeleryDistributed Systems

Data Engineering

Apache AirflowSnowflakeETL / ELT PipelinesdbtPandasNumPyData CleaningLarge-scale Datasets

Cloud & Infrastructure

AWS (S3, EC2)DockerKubernetesCI/CDMLflowHIPAA ComplianceHPC / GPU Clusters

Dev Tools

GitLinuxAgile / ScrumTechnical DocumentationCross-functional Collaboration

Education

Academic background

Master of Science

Artificial Intelligence

Northeastern University

Boston, MA, USA

Class of 2027GPA 3.75 / 4.0

Bachelor of Technology

Computer Science Engineering

Vellore Institute of Technology

Vellore, India

Class of 2025GPA 3.7 / 4.0

Research

Publications and patents

publication

A Comprehensive Study on Advancements in Text Mining and Natural Language Processing

2024 International Conference on Emerging Research in Computational Science (ICERCS) — IEEE

2024

  • Provided a comprehensive review of the evolution of text mining and NLP, from early rule-based and statistical methods to modern deep learning and transformer models, highlighting key advancements and techniques.
  • Discussed real-world applications, major challenges (data bias, interpretability), and future directions for fair, explainable, and sustainable NLP systems across healthcare, finance, social media, and law.
View on ORCID
patent

Deepfake Audio Detection Using Bi-Directional GRU and WGAN-GP

IP India — Indian Patent Office

December 4, 2025

App. 202541122126
  • Invented a deepfake audio detection system combining Bi-Directional GRU-RNN and WGAN-GP, achieving 94.2% detection accuracy on the ASVspoof2019 benchmark and 25% reduction in computational complexity vs. Bi-LSTM.
  • Enabled real-time audio authentication with deployment-ready performance for telephony security, media authentication, and forensic analysis.
  • Evaluated using Accuracy, F1 score, ROC-AUC, Equal Error Rate (EER), and t-DCF.