vaishanth.Let’s talk

AI systems / full-stack development

VaishanthPranavC S

Turning complexity into clarity.

I’m a Computer Science undergraduate at VIT Chennai, building machine learning projects and full-stack systems — from respiratory sound analysis to emergency coordination.

Explore my workDownload résuméGitHub
FIG. 01 / NEURAL LAB2 · 8 · 8 · 1

Teach a tiny network.

Make a prediction. Change the data. Watch it learn.

56 points · Circle
Class B51.3% model confidence · untrained

Hover or tap to inspect; no points are added.

Focus the canvas and use arrows to move the crosshair. Shift moves farther. Enter probes, adds a point in Add mode, or selects a nearby point in Edit mode. Select a point from the list to move it with arrows or remove it with Delete. Class A is a circle, Class B a diamond.

Epoch0
Train accuracy—
Train loss—
Recent loss — TrainLower is better
Train to see loss
0 epochs

How to read the lab
Regions & boundary
Lime predicts A; blue predicts B. Stronger tint means greater model confidence. The line is where both classes have a 50% prediction.
Epoch & loss
One epoch is a full pass through the training points. Loss measures prediction error; lower is better. Gaps in the chart mark changes to the dataset.
2 · 8 · 8 · 1
Two inputs (x and y), two hidden layers of 8 neurons, and one output: the probability of Class B. Confidence is the model’s estimate, not a guarantee.
Train vs test
Training points teach the model. Optional test points only evaluate it. A gap between the two accuracies shows how well it handles examples it has not learned from.
Restart vs reload
Restart training resets the weights and chart, keeping your data. Reload sample points restores the last chosen preset and restarts the model.
Advanced controls

56 training · 0 test points. Toggling restarts the model to keep evaluation honest. Test assignments stay fixed during edits; new points join training.

Scroll to exploreComputer Science · AI & ML
VIT Chennai / 2024–2028
01 AI & Machine Learning02 Full-Stack Development03 Real-time systems04 Competitive Programming

01Selected projects

Built to solve. Designed to understand.

Machine learning, emergency coordination, and evidence-grounded reasoning.

01 / EchoAssist02 / ResQNet03 / Multi-agent reasoning

Respiratory sound analysis · Team project · 2026

01

EchoAssist

1st place · Parallax 2026 · Team win

Listening closer with machine learning.

A respiratory sound analysis project combining an attention-enhanced neural network with an interactive dashboard for exploring audio, predictions, and model explanations.

  • Python
  • PyTorch
  • NumPy
  • SciPy
  • JavaScript
View on GitLab
My contribution
  • Adapted ResNet18 with CBAM attention and dual prediction heads to classify normal, crackle, wheeze, and combined respiratory sounds.
  • Built 16 kHz audio preprocessing and Log-Mel spectrograms; trained and compared models with focal loss and weighted sampling.
  • Evaluated 2,756 ICBHI test cycles from 49 patients: 60.56% accuracy and 44.32% macro F1, with confusion-matrix and prediction-error analysis.
  • Built audio upload, waveform, spectrogram, prediction, and heatmap displays; tested individual modules and the complete dashboard workflow.
EchoAssist / audio intelligenceProject overview

Illustrative waveform

01
Prepare the signal

16 kHz audio · Log-Mel spectrograms

02
Learn with attention

ResNet18 + CBAM · Dual prediction heads

NormalCrackleWheezeCombined

Emergency coordination · Team project · 2026

02

ResQNet

1st place · Find-a-thon 2026 · Team win

Connecting incidents to responders.

An emergency coordination platform bringing incident intake, priority scoring, responder matching, and live location updates into one connected workflow.

  • Node.js
  • Express
  • PostgreSQL
  • Prisma
  • Socket.IO
  • MapLibre GL
View on GitHub
My contribution
  • Developed APIs, PostgreSQL schemas, and Prisma queries with authentication and role permissions; implemented multilingual extraction, incident classification, and priority scoring.
  • Built responder matching and transactional assignments to handle duplicate dispatch attempts.
  • Integrated MapLibre maps, geolocation, routing, and live incident and responder updates through Socket.IO.
ResQNet / coordinationExplore the workflow

Understand the incident

Multilingual extraction turns incoming reports into structured incidents. Classification and priority scoring help organize the response.

APIs · Incident classification · Priority scoring

Project architecture · select a stage to explore

LLM systems · Evidence-grounded reasoning

03

Multi-Agent Misinformation Detection

Putting claims through a better debate.

A multi-agent LLM pipeline that extracts claims, retrieves web evidence, and uses Pro, Skeptic, and Judge agents to examine competing perspectives.

  • Python
  • Ollama
  • Streamlit
  • Sentence-Transformers
  • FAISS
  • scikit-learn
My contribution
  • Built claim extraction, web evidence retrieval, and Pro, Skeptic, and Judge workflows, ranking evidence by semantic similarity and source credibility.
  • Implemented FAISS claim memory and DBSCAN campaign detection with temporal analysis.
  • Delivered Streamlit dashboards and batch CSV workflows.
Claim → Evidence → DebateSystem architecture
Extracted claim
Grounded evidenceSemantic retrieval · FAISS memory

Examines evidence in support of the extracted claim.

OutputCalibrated verdict

Conceptual pipeline · select an agent to explore its role

02About me

Curiosity, put into practice.

VIT Chennai / 2024–2028

B.Tech · Computer Science & Engineering

Artificial Intelligence & Machine Learning
8.70 / 10CGPA

I’m Vaishanth Pranav C S, a Computer Science undergraduate specializing in AI and ML at VIT Chennai.

I work across machine learning and full-stack development. My projects explore how models interpret respiratory sounds, how responders coordinate during emergencies, and how multiple agents reason about a claim.

Team projects and coding competitions are a big part of how I learn. Beyond building software, I lead outreach for the Sports Club VITC, taking on student engagement and club leadership responsibilities.

Current focusMachine learning & full-stack systems
CommunityOutreach Lead · Sports Club VITC

03Technical skills

The tools behind the work.

A toolkit spanning model development, interactive interfaces, and reliable backend systems.

From audio preprocessing and attention mechanisms to semantic retrieval and multi-agent workflows.

  • 01PyTorch
  • 02ResNet18
  • 03CBAM
  • 04Grad-CAM
  • 05scikit-learn
  • 06Sentence-Transformers
  • 07FAISS

04Experience & education

Building beyond the classroom.

Sports Club VITC
VIT Chennai

Outreach Lead

Jul 2026–Present

Student outreach & club leadership

Progressed from Crew Member to Outreach Lead, taking on student outreach and club leadership responsibilities.

Crew Member

Jul 2025–Jun 2026
Education

Vellore Institute of Technology, Chennai

B.Tech in Computer Science and Engineering
Specialization in Artificial Intelligence and Machine Learning

2024–20288.70/ 10 CGPA

05Achievements

Good teams. Challenging problems.

Two team hackathon wins in 2026, alongside results in coding competitions and hackathons.

1ST

Parallax 2026

Team win · EchoAssist

1ST

Find-a-thon 2026

Team win · ResQNet

1ST

Code in the Dark

Coding competition

3RD

Code and Conquer

Coding competition

6TH

Hack-a-Throne

Hackathon

7TH

Neural-DAO 2024

Hackathon

06Get in touch

Have an interesting
problem in mind?

Let’s talk about AI systems, thoughtful software, or your next project.

Start a conversationDownload résumé
Say hello
Phone+91 87780 78699