Hemesh Chadalavada

Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo Photo

hey, i'm hemesh!

builder. engineer. filmmaker. computer engineering @ rice university.

see builds contact me

builds

Alpha Monitor

alpha monitor

spent 6 years building a wearable device for my grandma with alzheimer's.

hardware on-device ml
2018 - 2024
agent evan

agent evan

agi? the ultimate desktop agent that takes action - do anything, from anywhere.

agentic orchestration
sept. 2025
Summer AI

summer ai

a proactive ai agent on imessage that knows you better than you do (with 135+ tools).

multi-agent systems
oct. 2025
HootPlanner

hootplanner

1st place hackrice. the only rice course + club recommendation system.

etl pipeline lancedb
sept. 2024
Universis

universis

ai medical coding tool that outperformed human experts with 98.2% accuracy.

rag pipeline medical nlp
2023 - 2024
LockIn

lockin

a dynamic ai browser that blocks distractions automatically.

vector db semantic search
nov. 2024

In The Past:

"Achievements" that my Indian mom is proud of:

Let's hang out:

Reach out on my socials, or shoot me an email at hemesh@rice.edu.

Things in my mind:

attention and memory fourier transforms arduinos biryani films hard sci-fi books aperture SIGMA 18-50/2,8

Featured in:

old site (do not click)

let's chat!

i'd love to hear from you. drop me a line anytime.

LockIn

An AI Chrome extension that blocks distractions automatically. You're probably doom scrolling right now. Not good.

November 2024
sentence transformers cosine similarity agentic filtering
TL;DR

Built the first dynamic distraction blocker that actually understands context. Just type your goal and set a time. AI automatically blocks anything unrelated to your task.

the problem

Distractions are everywhere. You sit down to learn systems design and somehow end up watching videos about the Rizzler. Three hours gone.

Every current blocker uses hardcoded, static URL lists. You manually add sites to block, and the moment you need to use YouTube for research, you're stuck choosing between blocking everything or nothing.

We built the first dynamic system that actually understands context.

how it works

Just type your goal and set a time. Browse normally. LockIn's AI automatically blocks anything unrelated to your task without disrupting your workflow.

LockIn task setup

Select a time and define your task. Built-in to-do list and calendar included.

If your goal is "learn systems design," it will let you access YouTube videos, Google searches, and anything related to that topic. But it will restrict access to unrelated content.

LockIn blocking distractions

AI automatically blocks distractions unrelated to your goals. Blocked URLs are saved for later.

The magic happens through our three-layer AI filtering system:

Layer 1: The moment you start a task, a fine-tuned GPT-4o model generates a smart safelist of approved websites directly related to your goal.

Layer 2: For URLs not on the whitelist, we use sentence transformers (all-MiniLM-L12-v2) to compare your goal with tab titles and URLs through cosine similarity.

Layer 3: An AI agent makes the final decision on borderline cases, determining if that tab is helping or hurting your focus.

the build

The backend runs on Node.js and Express, managing the three-layer filtering smoothly. The Chrome extension uses Manifest V3, which came with its own set of challenges.

Dealing with Manifest V3 restrictions was tough, particularly accessing tab data securely. We also faced limitations with Chrome's local storage when implementing a timer that constantly updates without performance issues.

Building the cosine similarity system was interesting. Using all-MiniLM-L12-v2 to generate vectors and then building a cosine similarity function with proper threshold adjustments in 24 hours was a challenge.

We achieved very high response times for blocking. It operates smoothly without any delays or disruptions to a user's regular workflow.

LockIn productivity guardian

LockIn becomes your personal productivity guardian.

LockIn

Built with Rayaan Damani, Sandeep Chris Ramlochan, and Pradyun Chebolu.

Devpost GitHub

Universis

AI medical coding that outperformed human experts with 98.2% accuracy.

2024
rag pipeline semantic search medical nlp
Universis Dashboard
TL;DR

Built an AI medical coding system that achieved 98.2% accuracy in orthopedics, outperforming human experts. 69,000+ ICD-10-CM codes. 80,000+ CPT codes. All indexed and searchable. Received acquisition offers from top RCM companies and interviewed with Y Combinator W25.

the problem

Healthcare providers waste $19.7 billion per year due to incorrect medical claims, denials, and delays. Medical coders spend hours manually converting patient charts into ICD-10, CPT, and HCPCS codes. The cognitive load leads to errors that cost hospitals serious money.

Other companies have tried AI-powered coding assistance. They haven't been adopted because their accuracy sucks. Fine-tuned LLMs hallucinate codes that don't exist. Healthcare providers can't afford that risk.

Universis vs Human Coders

the solution

Universis converts patient charts into billing codes with higher accuracy than human professionals. Upload electronic health records, get the correct diagnosis and procedure codes back.

We built our own RAG pipeline and semantic search engine. No LLM hallucinations. The system achieved 98.2% accuracy and reduced manual coding time by 85% across pilot deployments.

69,000+ ICD-10-CM codes. 80,000+ CPT codes. All indexed and searchable.

Universis Dashboard

the traction

Worked with 40+ revenue cycle managers. Received acquisition offers from top RCM companies. Interviewed with Y Combinator W25.

Built with Siddharth Pilli.

HootPlanner

The only Rice course schedule + extracurricular recommendation system.

September 2024
1st place hackrice rag pipeline lancedb
TL;DR

Won First Place at HackRice 2024. Built a personalized course scheduling, research opportunity finder, and club recommendation system for Rice students using RAG and vector search.

the problem

Course discovery and scheduling is overwhelming. Students miss research opportunities they never knew existed. Too many choices in clubs and activities leads to decision paralysis.

the solution

HootPlanner is the one-stop shop for students preparing for their careers in college. Craft perfect class schedules in seconds. Discover research projects tailored to your passions. Find clubs and activities you'll actually love.

HootPlanner Course Scheduler

Course Scheduler: Find the best courses for you.

HootPlanner Research Opportunities

Research Opportunities Finder: Discover research you're truly passionate about.

HootPlanner Club Discovery

Club Discovery: Find clubs that suit you.

the build

Backend powered by FastAPI handling student onboarding, course recommendations, club suggestions, and research opportunities. We integrated LanceDB for vector search, using pre-trained language models to generate embeddings for semantic matching.

Frontend built with Next.js and Tailwind CSS. The RAG pipeline handles batch processing and cosine similarity calculations to provide personalized recommendations based on students' interests and academic profiles.

Goes beyond simple keyword matching to offer truly valuable recommendations.

HootPlanner

Built with Evan Tu and Justin Lee.

Devpost GitHub