- Reduced legacy service migration time by 98% (2 weeks to under 2 hours) by designing an adaptive AI discovery agent and deterministic code-generation workflows, authoring 20 agent skills to translate Java monoliths.
- Automated the generation of over 13,000 lines of production code to transform a 9,400-line warehouse scheduling monolith into a pluggable microservice architecture deployed across every fulfillment center.
- Re-architected the core decision engine into stateless solvers on AWS ECS, using Java CompletableFutures to orchestrate concurrent S3 data extractions, preventing execution bottlenecks during an 8-minute scheduling cron.
- Built a semantic-diff validation harness to replay production JSON payloads against generated endpoints, evaluating behavioral parity via a confidence metric that flags inconclusive noise across 42 execution paths.
// Computer Engineering + Econ minor @ Waterloo · GPA 3.95
Karan Gupta.
01 experience
- Implemented multi-step execution tracing and failure isolation for TRACE, an AI agent reliability engine, reducing root-cause diagnostic latency by 70% for complex RAG and tool-use workflows.
- Eliminated 95% of runtime non-determinism across 4 downstream production applications by implementing a canonical verifier interface and Pydantic schemas to enforce strict data contracts over LLM outputs.
- Deployed a high-throughput, 7-check deterministic execution engine that intercepts critical policy violations to enforce 100% compliance for financial-action workflows.
- Reduced system downtime risk by scripting a New Relic data exporter and architecting NRQL queries, enabling early detection of 3+ resource bottlenecks before system failure.
- Designed a Python analysis script and SQL telemetry pipeline to process clickstream data, identifying 20+ suspicious user anomalies by applying mathematical modeling and cross-referencing behavioral logs.
- Implemented 15+ Salesforce features to automate data entry during document uploads for specialized lending workflows, reducing manual input for high-value client processing using Apex and LWC.
- Improved real-time autonomous UAV signal detection accuracy by 13% by developing OpenCV computer vision algorithms, optimizing performance through iterative parameter tuning and extensive flight simulation validation.
- Built multi-process telemetry and command systems in Python using PyMAVLink to simulate flight communication, reducing message latency by 30% across distributed processes.
- Engineered a Python-based automation bot that auto-responded to Facebook Marketplace messages, improving response time by over 70% and automating 700+ customer interactions.
- Provided technical support in database management, hardware setup, and network troubleshooting.
- Designed and sold 35+ custom keyboards, generating $5,500+ in revenue, while building and maintaining an e-commerce platform for orders, client communication, and margin optimization.
- Planned and designed electrical engineering (EE) components and constrained the robot's physical build process.
- Collaborated with design and build sub-teams to define functional boundaries and integrate electronics efficiently.
02 projects
Won 1st place ($5,000) in Reactiv Track (shortlisted for Most Complex AI Hack) by building an Agentic AI system using Gemini vision models for product repairability assessment with 90%+ accuracy. Reduced search time by 70% and enabled access to 1,000+ real-time listings via Shopify Storefront API and SerpAPI.
→Raspberry Pi, Arduino, Python, TensorFlow, scikit-learn, OpenCV, HTML, CSS, Flask, C/C++DeliriumWatchBuilt a real-time Python monitoring pipeline with SQL-backed secure login and role-based access control for hospital staff, resulting in over 90% reduction in manual environmental monitoring. Implemented OpenCV continuous eye detection with a custom ML model (TensorFlow, scikit-learn) and live Flask dashboard to reduce nurse response time by 40%.
→Python, PRAW, MongoDBAI Admissions Similarity ToolAI-driven tool to scrape admissions data and compute similarity scores (GPA, tests, interests). Applicant benchmarking with reach, target, and safety school classification.
→Python, TensorFlow, scikit-learn, OpenCV, Node.js, React, MongoDBTailorAI (Silhouette)Built a full-stack AI platform using computer vision and the MERN stack, processing 200+ images to automatically extract 10+ body measurements per user. Implemented Python CV pipeline with landmark normalization, improving measurement consistency by 40%. Designed scalable backend APIs, MongoDB schema, and responsive React frontend.
→Raspberry Pi, Python 3, FastAPI, Spotipy, JSSpotify Pi ThingA standalone Raspberry Pi Spotify controller with a touchscreen UI acting like an in-car console. Uses FastAPI matching OAuth caching for independent boot-time kiosk playback.
→03 beyond code
- film — I log everything on Letterboxd.
- keyboards & PC building — 35+ custom boards built/sold, love the craft. Current rig: NZXT H6 Flow, 13600KF, 9070XT, Vengeance 7000MHz CL34.
- otherwise: manga, martial arts, motorsports, music (rage, EDM, Hyperpop).
- wanna know more? check out my resume ↗
04 stack
awards
- Valedictorian
- 1st Euclid Math Contest — School Medal
- 1st Amazon Robotics Hackathon
- Bronze Chess AI Bot — Waterloo Tech Week
- Gov. General's Academic Award
- 1st Hack Canada 2026 — Reactiv Track ($5,000)
- Shortlisted Most Complex AI Hack — Hack Canada 2026
05 see me in action
Find me on these platforms — code, music, film, and everything in between.
06 contact
Open to internships & collaboration.