Dev Patel

Applied AI/ML Engineer Intern, Julia AI

Cambridge, MA. Sep 2026 – Present.

• Building LLM agent workflows on LangChain and FastAPI that automate decision steps people handled by hand in 3 internal processes; the hardest, request triage, has to route messy, inconsistent requests without relying on a fixed input format.
• Shipped 4 LLM services on shared components; defined the API contracts for 3 and built their React/Next.js client.

Computer Vision and Machine Learning Engineer Intern, SPAN Inspection Systems

Gandhinagar, India. Jan 2025 – May 2025.

• Productionized an EfficientNet-B0 defect detector trained on 50,000+ images that catches broken tablets, missing pills, and damaged packaging, reaching 98.2% production accuracy and taking over 40.9% of manual inspection across 3 lines.
• Cut false positives 32.2% by teaching the model to ignore glare: reflections on foil and blister packs looked like real damage and were being flagged as defects, so added training augmentations that simulate that glare.
• Converted the model to TensorRT FP16 for NVIDIA Jetson: 42.8% lower latency, 27.3% more throughput, within 0.3 pts of FP32.
• Automated retraining and rollout with S3, Docker, and CI/CD, ending the 6 hours a week operators spent on manual updates.

Full-Stack Software Engineer Intern, Octopus Technologies

Remote. May 2024 – Jul 2024.

• Cut page load 38.3% for 5,000+ monthly users: the portal shipped as a single JavaScript bundle, so every page loaded the code for all 12 routes even when only one was opened; split it into route-level, lazy-loaded chunks.
• Rewrote slow queries behind compound indexes and a Redis cache, raising API throughput 35.6%.
• Added JWT auth for 8 user roles and moved deploys to GitHub Actions and AWS ECS (about 25 minutes down to 10).

Backend and Quantitative Systems Engineer Intern, WealthAffinity Softech

Ahmedabad, India. May 2023 – Nov 2023.

• Rebuilt the C++ order-matching core on lock-free queues and async I/O: 2.3 ms round trips, 1.4x capacity at 10,000+ orders.
• Some crashes in the new core only showed up sometimes, so a test could pass several times in a row and then fail; tracked them down, fixing 3 memory leaks with Valgrind along the way.
• Served an XGBoost risk model over REST that checks orders before they execute; it blocked 23% of erroneous trades, an estimated $50K a month in avoided write-offs.
• Wrote 150+ unit tests (6 modules at 95% coverage) and cut CPU use 28% after JMeter load testing.

Computer-Use Automation System

Browser agent that completes tasks on legacy web apps and turns each successful run into a replayable script.

Built with TypeScript, Node.js, Playwright, Gemini API, Zod, Vitest, Docker Compose.

SentiMind

Real-time sentiment and toxicity detection for Reddit threads and chat apps.

Built with Python, MCP, Gemini 2.0 Flash, WebSocket, Chrome Extension.

AI Task Assistant

Conversational task manager that turns plain-language requests into tasks.

Built with Cloudflare Workers AI, Llama 3.3, Durable Objects, Cloudflare Pages.

Sandbagging Model Organism

AI safety research: an LLM fine-tuned with LoRA adapters to deliberately underperform.

Built with Python, MLX, PyTorch (MPS), Hugging Face, scikit-learn.

FlightRoute

Multi-leg flight route planner between any two airports.

Built with Python, Flask, MongoDB, Pandas, NumPy.

Money Management App

Full-stack personal expense tracker.

Built with React, Node.js, Express.js, MongoDB, JWT, bcrypt.

Food Delivery Platform

Food ordering web app built with a team.

Built with React, Node.js, Express.js, MongoDB, JWT.

Real-Estate Listing Scraper

Collects project listings and specifications from Housing.com.

Built with Python, Selenium, Pandas.

StoryTelling

Web app that narrates a genre-based story in the user’s own cloned voice.

Built with Python, Flask, Play.ht API, pydub, Docker.