Open to internships · Hyderabad, India

I build apps where AI does the work and code keeps it honest.

I'm Murthy Roshan S L V Tolety. Computer Science undergraduate at Geethanjali College of Engineering and Technology, Hyderabad (Class of 2028). I like taking a messy real-world problem, designing a system around it, and building something people can actually use.

See my work→ GitHub LinkedIn Email
kartiq · agent pipeline · illustrative demo

❯

AI agent projects shipped
3
automated tests across them
580+
hackathons entered
6
HackSeries 3 with KartIQ
Semi-final

01 · Selected work

Things I've built

murthyroshan.github.io/while-youre-out
Echo Show emergency screen reading step 1 of the family's allergy plan: give the EpiPen in the outer thigh
01

While You're Out

Amazon Build, Ship, Shape Hackathon · Alexa+ track

An Alexa+ add-on for whoever looks after your kids

The babysitter asks the kitchen Echo Show anything. It answers in the parents' own words, asks the parents when their notes don't cover it, and in an emergency walks the sitter through the family's own plan, step by step, even when the internet drops.

  • MCP server (spec 2025-11-25) with MCP Apps cards, Tasks that wait for a parent's answer, and elicitation
  • Safety rules are code, not a prompt: same answer every time, no AI, under 10 ms. Gemini's second look can only add caution
  • Honest eval: 102 of 120 (85%) on a test set another AI wrote without seeing the rules, first run frozen in the repo
  • Parents answer from their lock screen (Web Push); runs on today's Alexa through an Alexa-hosted skill
  • TypeScript
  • Node.js
  • MCP
  • Gemini
  • SQLite
  • Alexa Skills
Live site ↗ View code ↗
lastlift.vercel.app
LastLift landing page: Find the last lift, beside a cutaway model of a station with its lifts lit
02

LastLift

Graph Hacks · Agents That Act on Connected Data

Find the last lift before it fails

Turns Boston's MBTA station layouts into a flow network in FalkorDB, so for every platform it knows how many lift failures it takes to cut off riders who need step-free access, and which lift that is. 89 of 366 step-free platforms hang on a single lift.

  • Max-flow and minimum cut find each platform's single points of failure; every FalkorDB result is certified and cross-checked against NetworkX
  • Grounded agent over MCP: right lift and platform in 23 of 24 answers with the graph, 5 of 24 without
  • Operations agents turn a closure into a ticket that is filed only after a person approves
  • Answers are marked out of date the moment a lift they read changes state
  • Python
  • FastAPI
  • FalkorDB
  • MCP
  • React
  • three.js
Live site ↗ View code ↗
github.com/murthyroshan/autonomous-commerce-agent
KartIQ search results showing a recommended laptop with AI reasoning
03

KartIQ

🏆 HackSeries 3 · Semi-finalist

Autonomous shopping agent for Indian e-commerce

Type what you want in plain English. An agent pipeline searches live listings, ranks every option with a deterministic scoring engine, and an LLM explains the pick. It narrates the decision but can't change it. Confirmed purchases are recorded on the Algorand testnet.

  • Search → compare → decide agent pipeline, streamed live to the browser over SSE
  • Reproducible ranking: weighted min-max scoring with store trust tiers and anomaly flags
  • On-chain receipts: purchase note, escrow smart contract and NFT receipt on Algorand testnet
  • 94 automated tests
  • Next.js
  • TypeScript
  • FastAPI
  • Groq · Llama 3.3
  • Algorand
View code ↗

02 · Hackathons

Built under pressure

  1. HackSeries 3

    Semi-finalist

    KartIQ: Autonomous shopping agent with on-chain purchase receipts.

  2. Amazon Build, Ship, Shape · Oct 2026

    Participant

    While You're Out: Alexa+ add-on that walks a babysitter through the family's own emergency plan (Alexa+ track).

  3. Graph Hacks · Oct 2026

    Participant

    LastLift: Agents on a FalkorDB flow network that find the lift a platform depends on (Agents That Act on Connected Data).

  4. Smart India Hackathon 2026

    Participant

    Freight Intelligence: Freight-rate forecasting and vessel scheduling dashboard for SAIL (problem SIH26006).

  5. Microsoft Imagine Cup 2026

    Participant

    Student Insight System: Early-warning system that flags academic risk and suggests faculty-led support, on Azure AI.

  6. HackerRank Orchestrate · Sep 2026

    Participant

    Aureon: Decides whether a purchase is safe now, later or on a plan using a 90-day cash-flow forecast.

03 · Approach

How I work

01

AI narrates, code decides

In While You're Out the safety rules are code that gives the same answer every time; the AI can only add caution. In KartIQ the ranking is deterministic and the LLM only explains it.

02

Never trust a number unchecked

LastLift certifies every FalkorDB max-flow result and re-derives each headline number with NetworkX. While You're Out quotes its honest 85%, not just its own 100%.

03

Build fast, verify properly

I build with AI-assisted tools and I'm upfront about it. What I own: the problem, the system design, and the tests that prove it works.

Languages

  • Python
  • JavaScript
  • TypeScript
  • SQL

Frontend

  • React
  • Next.js
  • Tailwind CSS

Backend

  • Node.js
  • FastAPI
  • Express
  • SQLite
  • FalkorDB (Cypher)

AI

  • AI agents & tool calling
  • Model Context Protocol (MCP)
  • Gemini, Groq, Azure OpenAI
  • Alexa Skills

Other

  • Git & GitHub
  • Docker
  • Vercel

04 · Contact

Let's build something real.

Open to internships, hackathon teams and interesting problems. Email is the fastest way to reach me.

toletyroshan@gmail.com
GitHub LinkedIn