Hemashree V I

Hello, I'm

HemashreeV I

I build systems that sense the world, decide, and keep people safe. That ranges from a patent-pending smart walker that predicts falls to a disaster-response rover that maps hazards and guides people out.

Hiring for
See my projects Play the demos

Selected work

Hardware that senses, models that see, and backend systems that stay up. Hover the scenes to tilt them, and pick a role at the top to reorder this list.

fall risk · early warning
Patent filing in progressCo-inventorTeam project

IoT smart walker that predicts falls

Many older adults are never shown how to use a walker or walking stick properly, so a small loss of balance turns into a fall. This walker watches how it is being used and warns before balance is lost.

  • Arduino Uno with dual MPU6050 IMUs, dual HX711 load cells and dual HC-SR04 ultrasonics, running fully standalone with no cloud.
  • Designed a sensor-fusion Fall Risk Prediction engine and a Dynamic Stability Reserve metric, plus near-fall detection and a Gait Symmetry Index for rehab tracking.
  • Estimated hardware cost under USD 30.
Explore the walker blueprint Try the fall-risk simulator
victim found · safe path ready
2026Team projectI own the AI perception layer

Autonomous rover for disaster-response reconnaissance

A ground rover that sweeps a disaster area, builds a live map, finds people and hazards, and guides survivors out by voice.

  • LIDAR SLAM (BreezySLAM) on a Raspberry Pi 5 builds a real-time map overlaid with a hazard risk heatmap.
  • Modified YOLOv8n with a risk-score head and a loss reweighted toward people, alongside a camera YOLOv9t model and an MQ-2 gas sensor for fire, smoke and gas.
  • Modified A* finds the safest, shortest evacuation path; reinforcement learning adapts the sweep around new hazards; a Flask dashboard streams telemetry for human oversight.
mAP@50 · 38.7%
Jul – Sep 2025Computer vision

Dental malocclusion detection from X-rays

A YOLOv9 pipeline that finds malocclusion and analyses teeth alignment in dental X-rays, built to make orthodontic screening faster and easier to explain.

  • Detects 31 dental conditions at 38.7% mAP@50 (25.1% mAP@50-95) on a 1,380-image held-out test set.
  • Grad-CAM heatmaps show which regions drove each prediction, so clinicians can check the reasoning.
  • Interactive Gradio app for real-time upload, prediction and visualisation.
item 3 · at-risk → move to front
In developmentMy main project

Smart Grocer: produce spoilage detection and inventory

A multimodal system that spots produce going bad on the rack, item by item, so stores can act before it becomes waste.

  • YOLO pipeline classifies each item as fresh, ripening, at-risk or spoiled, with a rack-level anomaly baseline (rolling statistics, z-score).
  • Sensor-fusion freshness engine combines vision, gas/VOC and environment scores, and aims gas sampling at camera-detected items.
  • Planned barcode/OCR expiry tracking and a Firebase + Flutter app for live rack monitoring and waste analytics.
Try the freshness check
200 buyers, 1 seat · 1 winner
2026Backend · Concurrency

SeatLock: thousands of buyers, one seat, zero double-bookings

A ticket-booking backend for the moment a big show goes on sale. When 200 requests race for the same seat, exactly one wins and the other 199 get 409 Conflict.

  • Three layers: a configurable lock (pessimistic, optimistic or Redis), a @Version fencing check on every write, and a partial unique index in PostgreSQL.
  • Proved with 200 simultaneous threads in Testcontainers and a k6 load test that fails if any seat is ever double-booked.
  • Payments as a Kafka saga with a transactional outbox, plus a Redis waiting room, Lua rate limiting and a live seat map over SSE.
Watch the deep dive (6 min)
likely root cause · payment v2
2026Reliability engineering

RootLens: finds why production broke, then fixes it

An incident-response lab of eight containers. When a bad release starts failing orders, it detects, investigates, decides, acts and verifies, but only within strict limits.

  • Detects with hysteresis (open at 30% failures, close below 10%) and scores every cause from the evidence; high confidence needs 0.75+ and a 0.3 lead.
  • A pure-function rulebook allows only rollback or restart, on allowed services, at most twice per incident, with an audit log of every decision.
  • Real run: payment v2 shipped, incident opened 62 s later, rolled back to v1, resolved in under 2 minutes. 118 unit tests in CI.
Watch the deep dive (6 min) Play the incident hunt
LLM
billing
technical
urgent
ticket routed in · real time
2026LLM systems

TriageIQ: an AI support inbox, fully on-device

Reads every support email, sorts it by category, priority and mood, and drafts a reply from tickets already solved. Nothing is sent until a human approves.

  • Emails become tickets on Kafka; a local Qwen 2.5 3B model classifies them at temperature 0 with a JSON schema, and falls back to a human with High priority if anything goes wrong.
  • RAG over solved tickets in Qdrant (top 3, ignored below 0.55 similarity) drafts a grounded reply that may only use facts from past answers.
  • A FastAPI console streams tickets over server-sent events; replies go out over SMTP in the same thread, with de-duplication, loop protection and rate limits.
Watch the deep dive (5 min)
github.com/you → a website
2026Web app · Live

GitLoom: your GitHub, woven into a website

Type a GitHub username and it reads the public repos, languages and activity, then lays them into a portfolio site you style in a live studio and download as one HTML file.

  • Ten templates, any accent colour, nine fonts, drag-to-reorder sections and hand-picked repos, with undo, redo and saved designs.
  • The generated site has phone layouts, a Ctrl+K quick menu, project search, a PDF résumé and share links that rebuild the design from the URL.
  • No backend, no sign-up, no build step: vanilla JavaScript and the GitHub REST API, with caching and an optional token.
Try it live Watch the demo (1 min)
session risk · step-up OTP
Nov 2025 – Jan 2026Security

Intrusion detection and threat modeling

A behavioural intrusion detection system for a simulated online bank. It scores risk across logins, access and transactions instead of relying on fixed rules alone.

  • Threat classification with automated responses to anomalous activity.
  • Detailed security logging for audit trails and post-incident analysis.
"open VS Code" · done
Personal projectRuns offline

Jarvis: a local voice assistant

A privacy-first voice assistant that runs entirely on my laptop. Nothing is sent to the cloud.

  • Voice commands open apps and run system controls, including shutdown and restart.
  • Reports battery and system status, with basic stock-price notifications.

Patent

A walker that understands how it is being used. Tap the numbered parts in the blueprint to see what each sensor does.

Filing in progress

IoT-Enabled Smart Walker with Sensor-Fusion Fall Risk Prediction and Adaptive Mobility Assistance

Co-inventor, with Ponn Oviyaa S. Faculty guide: Dr. Subitha D. Inventive disclosure filed with IIPRD (Khurana & Khurana).

Many elderly people never learn to use their walking aid properly, and that is when falls happen. This walker reads tilt, weight balance and nearby obstacles together, and flags failing stability before a fall.

  • Bilateral fall detection and real-time weight-asymmetry monitoring.
  • Fall Risk Prediction engine, Dynamic Stability Reserve, near-fall detection and Gait Symmetry Index.
  • Standalone on an Arduino Uno, under USD 30 in hardware.

Anatomy of the walker

Front view · smart walkernot to scale IMUIMU LOADLOAD ULTRASONICULTRASONIC ARDUINO UNOfusion · risk score 1 1 2 2 3 4 5 5

Schematic for illustration. Component positions are not the filed design.

Play while you're here

Four small games built from my projects. They run in your browser and are simplified for fun, so their numbers are illustrations, not project results.

Keep them standing

Change what the walker's sensors feel and watch the stability score react before a fall would happen.

tilt + load + distance → stability
Dynamic Stability Reserve
—
Stable
Left load 50%
Right load 50%

A toy model of the idea. It is not the patented method, and the numbers are not measured data.

Skills and tools

What I build with, grouped by the kind of work. Every tool here is one I have used in a project above.

About me

Final-year B.Tech CSE student specialising in Cyber Physical Systems at VIT Chennai.

HV

I like problems where software meets the physical world: a sensor reading has to become a decision, and the decision has to be safe. That led me to co-invent a fall-predicting walker, build the perception layer of a disaster-response rover, and train vision models that explain what they see.

Lately I'm also building the backend side properly, with microservices, monitoring, Kafka and local LLMs. Outside code, I lead a student team at CodeYGen Club and design for the Fitness Club, so I care about how things look as much as how they work.

தமிழ்
Tamil
Native
English
English
Fluent
हिन्दी
Hindi
Intermediate
日本語
Japanese
Basic, course completed

Journey

  1. Aug 2023
    Started B.Tech CSE (Cyber Physical Systems)
    Vellore Institute of Technology, Chennai
  2. May – Jul 2025
    Research intern, VIT Chennai
    Human pose estimation under Prof. Pavithra: MMPose heatmap models, keypoints validated against Kinovea motion analysis.
  3. Jun 2025
    NUS Introduction to Industry 4.0 Programme
    National University of Singapore (SCALE and College of Design and Engineering). Team project on digital transformation with the Business Model Canvas.
    View certificate
  4. Jun 2025
    SuperAI Conference 2025, Singapore
    Asia's largest AI conference at Marina Bay Sands, with 7,000+ innovators and 1,000+ AI companies.
  5. Jun 2025
    IBM Generative AI certification
    Gen AI Using IBM Watsonx, IBM Career Education Program via IBM Developer Skills Network.
    View certificate
  6. Jul – Sep 2025
    Dental malocclusion detection
    YOLOv9 with Grad-CAM, 38.7% mAP@50 across 31 conditions.
  7. 2025 – 2026
    Smart walker patent
    Inventive disclosure filed with IIPRD; filing in progress.
  8. 2026
    Disaster-response rover and backend platforms
    Rover perception layer, then SeatLock, RootLens, TriageIQ and GitLoom.

Education

Vellore Institute of Technology, Chennai
B.Tech, Computer Science Engineering, Cyber Physical Systems
Aug 2023 – present
Maharishi International Residential School, Kanchipuram
Class 10 and Class 12, CBSE

Leadership

Team Leader, CodeYGen Club
Led a student team through coding and design projects, coordinating development and presentations.
Design Lead, Fitness Club
Designed posters, promotional creatives and digital content for events and member engagement.