Hi, I'mYash Sanghvi
Software Engineer specializing in AI/ML, Full-Stack Development, and building production systems with containerized ML pipelines.
Who I Am
Software Engineer passionate about AI infrastructure, machine learning, and robotics systems. Currently building production AI solutions at Quantea and TEAMCAL AI while pursuing my degree at Santa Clara University. I specialize in cloud architecture, containerized ML deployments, and embedded control systems.
Education
Santa Clara University
B.S. Computer Science and Engineering
Expected May 2027
Stony Brook University
Computer Science
Aug 2024 - Jun 2025
Silicon Valley Career Technical Education Center
Mechatronics, Robotics, and Automation Engineering
Aug 2023 - Jun 2024
Experience
Work Experience
My professional journey and contributions to various organizations.
Software Engineering Consultant
New Venture Visions
- •Developing a Python-driven tax automation engine using Supabase and Prisma to manage multi-dimensional data categorization with 98% accuracy for real estate investors.
- •Architected a scalable PWA (iOS/Web) featuring a low-latency data ingestion pipeline that reduces relationship-logging friction via automated priority logic.
- •Optimized cloud database schemas and serverless functions to ensure seamless data synchronization and high system reliability across distributed environments.
Cloud Infrastructure Engineer Intern
Quantea, Inc.
- •Orchestrated containerized microservices using Kubernetes and Docker, with Python scripts driving automated infrastructure orchestration and reducing deployment latency.
- •Executed rigorous inference benchmarking and performance tuning across NVIDIA H100/A100 clusters to ensure high availability and optimized throughput.
- •Engineered technical documentation for multi-node clusters, reducing infrastructure support tickets by 20%.
Undergraduate Research: NLP & Transformers
Santa Clara University - Trustworthy Computing Lab
- •Conducting directed research under Prof. Yuhong Liu at the Trustworthy Computing Lab.
- •Focusing on Deep Learning concepts including Transformers and Natural Language Processing (NLP).
- •Investigating robustness and security within AI/ML architectures.
AI/ML Software Engineer Intern
TeamCal.ai
- •Designed and authored Technical Integration Guides for Microsoft Graph API and OAuth protocols, adopted by the engineering team to standardize multi-tenant calendar management.
- •Developed real-time communication cascades using TypeScript and Electron, implementing CI/CD pipelines to ensure reliable, automated deployments.
- •Resolved complex concurrency conflicts and race conditions within distributed scheduling agents via rigorous state-management validation.
Competitions
Awards & Competitions
Competitive achievements and recognitions.
ToolForge — Self-Building Agentic Tool System
Llama Lounge Agentic Hackathon — Cerebral Valley x Snowflake
- •Built ToolForge with Abraham Bhatti and Rishit Shiramshetti in under 24 hours—an agent that doesn't just use tools but builds production-grade ones on the fly. A Planner picks the strategy for any task (reuse an existing tool, tap a Composio integration, or generate something new), and a cascaded CrewAI Builder writes Python tool code, tests it in a sandbox, and iterates until it works.
- •Engineered a tool memory layer in Snowflake so every validated tool is stored and reused near-instantly on the next task—after Skyfire verifies API usability. The agent accumulates capability over time: the more it's used, the more it can do, with no one manually wiring integrations.
- •Tracked full reasoning traces, execution history, and usage metrics end-to-end across the agentic loop. Built on Snowflake Cortex, CrewAI, Composio, and Skyfire, competing against an extremely competitive field.
PaperToProtein — Paper to Protein Binder Design Pipeline
Bio x AI Hackathon — Tamarind Bio (W24) x BioRender (W18)
- •Built PaperToProtein with Achyut Chebiyam in a single day at Y Combinator HQ—a Design Space Explorer that turns any research paper into an end-to-end protein binder design pipeline. Claude parses uploaded PDFs to extract target proteins, UniProt IDs, PDB structures, disease context, and binding site hints, then UniProt and AlphaFold pull real sequences and predicted 3D structures through Tamarind Bio's compute infrastructure.
- •Engineered the full de novo binder design pipeline: RFdiffusion generates binder backbones, ProteinMPNN designs amino acid sequences, and Boltz-2 predicts the target-binder complex structure and binding affinity—running on Modal GPUs with an NVIDIA NIM fallback for resilience. Not retrieval—real generative design end-to-end.
- •Shipped an interactive D3 force-directed graph where nodes are binders and edges are structural similarity, with a Three.js viewer that renders predicted complexes on click and Claude narrating clusters, the Pareto frontier, outliers, and top candidates. Built both a cached demo mode and a live mode wired to real computational biology infrastructure—sponsored by Anthropic, OpenAI, and Modal, competing against teams that flew in internationally.
Agent Watch — Real-Time AI Agent Observability
AWS x Anthropic x Datadog Hackathon
- •Built Agent Watch solo at the AWS Startup Loft in SF—a real-time observability and reliability platform that gives full visibility into everything an AI agent does before any tool executes. Built on AWS Bedrock with Anthropic's Claude, it runs three checks on every call: behavior, security, and cost/performance.
- •Engineered a Neo4j policy graph that enforces exactly which tools an agent can use and with what parameters (because the smarter agents get, the better they get at finding loopholes), plus a behavior layer that cross-references input intent with output to catch hallucinations, misinterpretation, and drift.
- •Stress-tested against prompt injections, data exfiltration, social engineering, and cost spike attacks—Agent Watch caught every one. Token usage, latency, and compute tracked in real time across 15+ metrics on a Datadog dashboard.
AI Sports Broadcast — Autonomous Cinematography for Youth Sports
Gemini 3 SuperHack — Cerebral Valley
- •Built an autonomous sports broadcast system that lets parents keep their phone in their pocket while AI handles the camera crew—turning raw sideline footage into a broadcast-style highlight reel for any field or court sport. The pitch: stop watching the game through a screen; let AI be the camera crew.
- •Engineered a YOLO11n-based 'star tracking' pipeline with a 7-signal scoring algorithm (acceleration, verticality, pose dynamism, etc.) to identify the player driving each play, then layered EMA smoothing and dead-zone logic on top to deliver a cinematic 9:16 vertical crop without the usual shaky-cam jitter that plagues automated framing.
- •Wired Gemini 3 in as the 'director'—watching the footage and generating context-aware descriptions—then piped those into a high-energy script engine and OpenAI TTS to mux a professional play-by-play voiceover onto the reel. Stack: YOLO11, OpenCV, Supervision, Google GenAI, OpenAI, FastAPI, Next.js 15, React 19.
Local SF Tourism App — 7 Models Co-Scheduled on DGX Spark
NVIDIA Spark Hack Series — NVIDIA x Dell x Arm x Antler
- •Built a fully local SF tourism and culture app with Sanjay Sai, Ago Lajko, Akhil Devarasetty, and Sai Vasanth Kattamuri—using the NVIDIA DGX Spark as a pure on-device backend. No cloud, no APIs, just raw compute. Most 'AI apps' today are clever UI glued together by API calls; we wanted to do the opposite.
- •Pushed the DGX Spark with a massive inference co-scheduling stress test: 7 models running simultaneously on the box—Nemotron-3 8B (text) + Qwen 2.5 3B (logic) for generation, Nemotron 0.6B + Pocket-TTS on ARM CPU for speech, a custom diffusion inference path for vision, and NeMo 0.6B Embeddings + cuVS for retrieval.
- •Hit the ARM ecosystem gap head-on—vLLM Omni doesn't support the DGX series for non-standard paths like our diffusion model, so we built a custom inference pipeline from scratch. Sustained local inference performance once the gaps were bridged was incredible. Hands-on experience running a data-center-grade stack at the edge.
Sentinel SDK — Runtime Security for AI CLIs
Google DeepMind Continual Learning Hackathon
- •Built Sentinel SDK with Abraham Bhatti after watching an AI CLI almost nuke a codebase—a runtime security layer that sits in front of OpenCode, Claude Code, Aider, and other agentic CLIs to intercept, evaluate, and safely execute every tool call. AI CLIs silently delete files, leak secrets, and misread 'clean up the project' as rm -rf /; teams either accept the risk or abandon the tooling. Sentinel does neither.
- •Engineered Bastion Guard, a three-stage pipeline that doesn't just block risky commands—it learns and finds safer paths forward. Triggers detect what the CLI is about to touch (files, directories, APIs, secrets), Checks consult an LLM-bootstrapped rule set plus a persistent JSON memory of past decisions, and Enforcement chooses one of four outcomes: execute, kill, reroute to the LLM for a safer rewrite, or suggest an alternative—each decision logged back into memory.
- •Wired in Akash Network for distributed LLM inference so the guard scales with workload (not laptop compute), Composio for real-time logging to GitHub, and You.com Search for live context on flagged commands so the LLM cross-checks against current docs—not stale training data. Speed without gambling.
Robotics, Urban Search and Rescue
SkillsUSA California
- •Achieved 1st place in the regional SkillsUSA Urban Search and Rescue Competition and top 10 in State.
- •Spearheaded the design and development of a sophisticated robotic solution capable of navigating obstacles and retrieving objects, such as a cube from a simulated mailbox environment.
- •Invested over 120 hours in the meticulous engineering process, ensuring the robot's competitiveness and readiness for high-stakes competition.
Skills
Technical Skills
From the math behind backprop to GPU clusters in production.
Agentic AI & LLM Systems
Building autonomous AI agents with memory, tool-use, and multi-step reasoning
Production ML Infrastructure
Deploying and maintaining GPU clusters for ML model serving at scale
Languages
Machine Learning Foundations
Deep Learning & Computer Vision
Cloud & Infrastructure
Full-Stack Development
Hardware & Embedded
Projects
Featured Projects
A selection of my recent work and personal projects.
SF Quest (Golden Gate Quest)
A gamified mobile web app enabling users to discover San Francisco through personalized photo-based treasure hunts. Features AI voice guide powered by Pipecat, historic photo comparisons, and personalized itineraries based on user preferences. Uses RAG with NeMo Embeddings and CuVS Vector Search.
Autonomous Scheduling Assistant
An intelligent voice-enabled scheduling assistant powered by LangGraph and GPT-4o-mini. Features natural language processing, Google Calendar two-way sync, smart conflict detection, context-aware memory, and multi-platform support for Zoom, Teams, and Google Meet.
Highlight Generator
Auto-generate basketball highlight reels with AI narration. Uses YOLO11n tracking with 7-signal scoring (motion, acceleration, jumps, size, centrality, persistence, pose) to identify key players, auto-crops to 9:16 vertical format, and adds AI-generated sports commentary via Gemini and OpenAI TTS.
SafePath
Full-stack safety navigation analyzing real-time SF crime & 311 incident data to recommend the safest walking routes. Features time-decay risk scoring algorithm (72hr for high-risk, 24hr for low-risk incidents), 200m safety buffer detection, and offline route optimization with waypoint injection.
PortPlateAI
Comprehensive data analytics dashboard for California's top agricultural commodities. Features interactive Recharts visualizations, AI-powered natural language query interface (LangGraph-ready), and spoilage simulation with temperature/transportation delay modeling and economic loss estimation.
Robotic Motion Control System
Designed and 3D-printed 4+ robotic subsystem prototypes integrating servo/motor assemblies for autonomous actuation. Programmed C++ control logic to automate motion, sensor polling, and test workflows, reducing manual validation time by ~40%. Performed electromechanical stress testing improving response consistency by 25-35%.
Certifications
Licenses & Certifications
Professional certifications and credentials I've earned.
Deep Learning Specialization
DeepLearning.AI (Coursera)
Issued 2026
Machine Learning Specialization
DeepLearning.AI (Coursera)
Issued 2026
UAV PART 107
Federal Aviation Administration
Issued Jun 2024 · Expires Jun 2029
Devtools Pro: Beginner to Expert w/ Chrome Developer Tools
Udemy
Issued Dec 2025
Credential ID: UC-222fe103-e042-4cb3-afdc-023a58cd622a
CodePath Intermediate Technical Interview Prep (TIP102)
CodePath
Issued Spring 2025
Credential ID: 114914
Generative AI for Java Developers with Azure OpenAI ChatGPT
Udemy
Issued Feb 2025
Credential ID: UC-f74b562b-f236-4d35-a970-5f118e8e7012
Contact
Get In Touch
I'm currently open to new opportunities. Whether you have a question or just want to say hi, feel free to reach out!
Contact Information
I typically respond within 24-48 hours. For urgent matters, please reach out via LinkedIn or email directly.