CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS • CAUTION DO NOT CROSS •
SUSPECT PROFILE
CLASSIFIED
654
SHARMA, AYUSH
SUBJECT INFORMATION
NAME:Ayush Sharma
ALIAS:AI Engineer
LAST SEEN:Boston, MA
CURRENT AFFILIATION:ReSupply (AI Engineer) since 06/2025
CLEARANCE:M.S. in Artificial Intelligence, Boston University
KNOWN FOR:Building end-to-end AI systems for real-world workflows like pricing, logistics, HR, voice/chat support, etc. to improve efficiency and business outcomes, with prior research experience in computer vision and LLM systems.
VERIFIED EMPLOYMENT RECORDS • PROFESSIONAL HISTORY LOGGED • VERIFIED EMPLOYMENT RECORDS • PROFESSIONAL HISTORY LOGGED • VERIFIED EMPLOYMENT RECORDS • PROFESSIONAL HISTORY LOGGED • VERIFIED EMPLOYMENT RECORDS • PROFESSIONAL HISTORY LOGGED • VERIFIED EMPLOYMENT RECORDS • PROFESSIONAL HISTORY LOGGED • VERIFIED EMPLOYMENT RECORDS • PROFESSIONAL HISTORY LOGGED •
EVIDENCE BOARD
EXPERIENCE FILES
Subject's Work Experience — Click any case file to examine further
#001
THE RESUPPLY FILE
AI Engineer @ ReSupply | Boston, MA | June 2025 - Present
ONGOING
THE RESUPPLY FILE
AI Engineer @ ReSupply | Boston, MA | June 2025 - Present
AI Voice Agents
$60K/mo Saved, AI Handles 20K+ Monthly Calls
Built and deployed multi-model voice agents with live safety guardrails to fully automate customer and driver support calls.
RAG + Eval Pipeline
Custom retrieval system powering voice agents' knowledge, with automated call quality grading via dual-model evaluation.
Intent Classification
7K+ Automatable Tickets, 1K+ Misroutes - 30% Workload Cut using LLM Intent Tagging and Clustering
Clustered 25K+ monthly support conversations to identify which tickets AI agents could handle, eliminating the need for additional hires.
Resume-Ranking Tool
Improved ReSupply’s HR screening process by building an internal resume-ranking tool with Breezy ATS automation, Playwright, and LLM-based evaluation pipeline for resume retrieval, location filters, and custom JD-based comparative candidate ranking.
Geospatial Optimization
Automated dispatch routing and zone recommendations for haulers using MapBox travel-time isochrones and ZIP code patterns.
Pricing Model
Improved company's pricing algorithm by combining ML models with economic modeling using demand elasticity, Producer Price Indexing, and rigorous A/B testing.
Built a semantic search database for item retrieval that powers recommendations, autocomplete, and query handling across user-facing features.
TOOLS AT THE SCENE:
Pipecat, Twilio, Deepgram, ElevenLabs, OpenAI API (GPT 4, o3), Llama Guard 3, Transformer Embeddings (GTE-large, E5-Large-v2), FAISS, LLM-as-a-judge Evaluation, Google Gemini API (2.5 family), Intent Classification, UMAP, HDBSCAN
#002
THE NSF RESEARCH FILE
AI Researcher @ National Science Foundation, Boston University | Boston, MA | Oct 2023 - May 2025
UNDER REVIEW
THE NSF RESEARCH FILE
AI Researcher @ National Science Foundation, Boston University | Boston, MA | Oct 2023 - May 2025
DNS Domain Names Prediction
ACM SIGCOMM 2026 Submission | Advisor: Mark Crovella
Fine-tuned LLMs and trained ML models to predict unknown DNS domain names using reinforcement learning with a structured reward schema optimizing for novelty and out-of-distribution generalization.
Surpassed the previous MiniARC benchmark (33.1%) on abstract reasoning by combining re-ranking, meta-learning, and multimodal approaches to advance human-level AI reasoning, achieving 41.81% accuracy.
TOOLS AT THE SCENE:
Open Source GPT models, Ollama, Reinforcement Learning, VERL framework, CUDA, Multiprocessing, Data Augmentation, Meta Learning, Multimodal Reasoning, AlphaCode, Prompting Techniques
#003
THE SCHNEIDER ELECTRIC FILE
ML Intern @ Schneider Electric | Bangalore | Jan 2023 - Jul 2023
CLOSED
THE SCHNEIDER ELECTRIC FILE
ML Intern @ Schneider Electric | Bangalore | Jan 2023 - Jul 2023
Global Data Housekeeping Project
Built automated data-cleaning and validation pipelines that cut cloud migration time by 80% across 11.25 million records.
Ticket Classification
Automated the classification of incoming complaint and request tickets using deep learning NLP models.
TOOLS AT THE SCENE:
NumPy, Pandas, SQL, Transformer Models (BERT and RoBERTa Embeddings), LSTM Models
#004
THE JP MORGAN FILE
SWE Intern @ JP Morgan | Bangalore | Jul 2022 - Aug 2022
CLOSED
THE JP MORGAN FILE
SWE Intern @ JP Morgan | Bangalore | Jul 2022 - Aug 2022
Stock Visualization
Built an interactive visualization page for analyzing stock market data across 50,000+ data points.
Prediction Model
Improved stock price prediction accuracy from 78.5% to 83.8% by applying recurrent neural network and LSTM models to processed market data.
AUTHORIZED PERSONNEL ONLY • TECHNICAL INVENTORY • AUTHORIZED PERSONNEL ONLY • TECHNICAL INVENTORY • AUTHORIZED PERSONNEL ONLY • TECHNICAL INVENTORY • AUTHORIZED PERSONNEL ONLY • TECHNICAL INVENTORY • AUTHORIZED PERSONNEL ONLY • TECHNICAL INVENTORY • AUTHORIZED PERSONNEL ONLY • TECHNICAL INVENTORY •
WEAPONS INVENTORY
CONFISCATED
Subject's Weapons and Tools — On Record
AI SYSTEMS
LLM APIs and SDKs (OpenRouter, Anthropic Claude, OpenAI, Google Gemini)RAGAI Guardrails and SafetyTransformers and Embedding ModelsMultimodal SystemsFine-TuningReinforcement Learning TechniquesData SynthesisPrompt Engineering TechniquesDiffusion Models
FRAMEWORKS AND LIBRARIES
LangChain (certificate)LangGraphPipecatDeepgramElevenLabsOllamaPyTorchTensorFlowScikit-LearnHugging FaceAutoGen for Multi-Agent Systems (certificate)NumPyPandasMatplotlibFAISS
Built and deployed an AI assistant using GPT-4o-mini, FAISS, and Hugging Face all-mpnet-base-v2 embeddings to let users upload and query multiple research papers and arXiv links through conversational RAG, contextual memory, and LaTeX Math rendering.
Tech Stack: Python, SQL, PyMySQL, Pandas, NumPy, Matplotlib, Seaborn, Statistical Data Analysis, Data Visualization
Led a team project for WGBH, a public radio station in Boston, MA, analyzing Massachusetts debt-collection court data using SQL and Python to uncover trends across 177,248+ debt-collector court cases, 233,074 Capias warrant cases, and 2,912 wage-garnishment cases. Identified major debt-collection entities, litigation patterns, mortgage-servicing leaders, and state/country-wise debt-collector distributions through large-scale statistical analysis and visualization workflows.
Built a text-guided video-to-video generation pipeline using diffusion models to transform input videos into stylized, temporally consistent outputs. Extended the Rerender-A-Video framework with adaptive key-frame sampling, optical-flow/frame-difference analysis, CLIP-based prompt-image evaluation, cross-frame attention, FreeU, and Ebsynth-based frame propagation to reduce flickering and improve visual coherence across generated videos.
Built an autonomous driving perception system using YOLO-based object detection and computer vision techniques to identify vehicles and road objects in real time from driving footage, enabling accurate bounding-box localization and multi-object traffic scene understanding for self-driving applications.
Built and optimized GAN and BigGAN architectures in PyTorch to generate realistic dog images using the Stanford Dogs dataset, implementing techniques such as spectral normalization, self-attention, conditional batch normalization, and residual networks to reduce mode collapse and image artifacts. Improved image fidelity significantly across training iterations under constrained Kaggle GPU resources.
Tech Stack: Python, TensorFlow, Keras, VGG-19, Streamlit, NumPy, Pillow, Matplotlib, CNNs, Transfer Learning, Neural Style Transfer
Built a Neural Style Transfer model that takes in a content image and a style image, then generates a new artwork by preserving the content structure while transferring artistic style features using a VGG-19 CNN. Implemented content loss, style loss with Gram matrices, multi-layer style extraction, and iterative image optimization over 500 epochs to produce stylized generated images.
Built a jazz music generation system using an LSTM-based sequence model trained on a corpus of jazz music to learn note/chord patterns and generate new MIDI-style jazz compositions that sound like a full-band performance.
ACADEMIC RECORDS • VERIFIED TRAINING HISTORY • ACADEMIC RECORDS • VERIFIED TRAINING HISTORY • ACADEMIC RECORDS • VERIFIED TRAINING HISTORY • ACADEMIC RECORDS • VERIFIED TRAINING HISTORY • ACADEMIC RECORDS • VERIFIED TRAINING HISTORY • ACADEMIC RECORDS • VERIFIED TRAINING HISTORY •
TRAINING DOSSIER
VERIFIED
Where was the subject trained?
Boston University
M.S. Artificial Intelligence with a Master's Thesis | 2023 – 2025
Coursework: Machine Learning, Deep Learning, Natural Language Processing, Computer Vision, Artificial Intelligence, Data Science, Metrics and Evaluation in NLP and LLMs, Directed Reasearch Study
Shri Mata Vaishno Devi University
B.Tech Computer Science | 2019 – 2023
Coursework: Neural Networks and Fuzzy Sets, Artificial Intelligence, Machine Learning, Soft Computing, Digital Image Processing, Nature Inspired Algorithms, High Performance Computing, Engineering and Discrete Mathematics
ACTIVE COMMUNICATION CHANNELS • SUBJECT REMAINS AVAILABLE FOR CONTACT • ACTIVE COMMUNICATION CHANNELS • SUBJECT REMAINS AVAILABLE FOR CONTACT • ACTIVE COMMUNICATION CHANNELS • SUBJECT REMAINS AVAILABLE FOR CONTACT • ACTIVE COMMUNICATION CHANNELS • SUBJECT REMAINS AVAILABLE FOR CONTACT • ACTIVE COMMUNICATION CHANNELS • SUBJECT REMAINS AVAILABLE FOR CONTACT •
WANTED
AYUSH SHARMA
For building AI systems that make manual repetitive processes obsolete
END OF DOSSIER • SUBJECT STILL ACTIVE • END OF DOSSIER • SUBJECT STILL ACTIVE • END OF DOSSIER • SUBJECT STILL ACTIVE • END OF DOSSIER • SUBJECT STILL ACTIVE • END OF DOSSIER • SUBJECT STILL ACTIVE • END OF DOSSIER • SUBJECT STILL ACTIVE • END OF DOSSIER • SUBJECT STILL ACTIVE • END OF DOSSIER • SUBJECT STILL ACTIVE • END OF DOSSIER • SUBJECT STILL ACTIVE • END OF DOSSIER • SUBJECT STILL ACTIVE •
