yashvarma
Selected Work9 projects

Selected Work

Systems engineered under real constraints: hackathon clocks, conference reviews, or production pipelines operating under strict SLAs.

01
AI & Cybersecurity

Kavach.ai

Code
Problem

Security analysts need to know what a suspicious binary actually does before detonating it in production, and standard tools only flag static hashes.

Solution

A desktop cybersecurity threat triage suite that detonates suspicious binaries inside an isolated, disposable Docker sandbox, fuses static file entropy analysis with dynamic syscall flagging, and uses an XGBoost classifier trained on ~184 GB of malware bytecode.

Architecture

Electron desktop app frontend communicating with a Python/Flask engine. Combines static entropy analysis, dynamic syscall logging in ephemeral Docker sandboxes, an XGBoost classifier with pseudo-soft labeling, and a fine-tuned LLaMA 3.2 model that translates call-stack telemetry into plain-language security verdicts. Validated live against GajShield malware samples during judging.

Electron.jsPythonFlaskDockerXGBoostLLaMA 3.2PyTorch
1st Place (Winner) · GajShield Hack 8 (2,000+ Participants)
02
AI & Multilingual Fintech

LoanSaathi

Code
Problem

Over 40% of loan applications in India get rejected on paperwork errors, and predatory hidden terms exploit rural applicants who cannot read complex fine print.

Solution

A 24-hour multimodal conversational assistant that extracts loan documents via PaddleOCR, detects predatory hidden clauses using Mistral Saba 24B and fine-tuned FinLLaMa, and guides eligibility and EMI calculations across 8 Indian languages.

Architecture

Next.js frontend and FastAPI backend; PaddleOCR for document OCR extraction, Mistral Saba 24B and domain-fine-tuned FinLLaMa for clause risk scoring, Sarvam AI for multilingual voice and speech synthesis, with Supabase managing ephemeral sessions where documents are automatically deleted after processing.

FastAPINext.jsMongoDBSupabaseMistral Saba 24BFinLLaMaPaddleOCRSarvam AI
1st Place (Winner) · The Great Bengaluru Hackathon (~5,000 Participants)
03
AI & Maritime Safety

SeaGuard

Code
Problem

Traditional fishermen navigate open waters with no reliable way to predict rich fishing zones or detect sudden treacherous maritime weather and international boundary incursions.

Solution

A mobile assistant that predicts Potential Fishing Zones (PFZ) at ~97% accuracy from satellite and oceanic weather telemetry using TensorFlow, paired with SOS emergency broadcasting, GPS maritime boundary warnings, catch logging, and multilingual voice access.

Architecture

React Native mobile client with FastAPI backend. TensorFlow convolutional models trained on satellite telemetry and sea-surface temperature data predict PFZ coordinates. Integrated with Supabase and Sarvam AI voice models. Exhibited to Dr. V. Narayanan, Chairman of ISRO in 2025.

React NativeFastAPITensorFlowPythonSupabaseSarvam AIGPS
Most Product Readiness · Hackcelestial 2.0 (Exhibited to ISRO Chairman)
04
Multi-Agent AI & Fine-Tuning

Intelligent Cooperative Q&A Agents

Problem

Evaluating and synthesizing deep mathematical and domain-specific knowledge requires cooperating agentic models that can self-verify and prevent hallucinations under strict compute constraints.

Solution

Two cooperating AI agents fine-tuned under a 24-hour AMD GPU constraint: a Question-Generator (Qwen-3 4B) trained on domain texts and OEIS sequences, and an Answer Solver (LLaMA 3.2 3B) using multi-pass verification.

Architecture

Built with PyTorch using Supervised Fine-Tuning (SFT) and synthetic data pipelines on AMD GPUs. Achieved ~10s question generation and 6-7s answer latency with multi-pass consensus verification.

Qwen-3 4BLLaMA 3.2 3BPyTorchSFTSynthetic DataAMD ROCm GPUs
2nd Runner Up · AMD AI Sprint @ IIT Bombay (~200 Participants)
05
AI & Consumer Fintech

FinSaathi

LiveCode
Problem

Everyday retail investors lack access to financial advisory tools, and thousands of eligible government welfare schemes go unclaimed due to discoverability barriers.

Solution

An AI-powered financial advisory platform combining algorithmic stock analysis, paper-trade simulation, financial document parsing, government scheme discovery, and real-time WhatsApp stock notifications.

Architecture

Next.js frontend, Flask/Python backend, PyTorch predictive models, and MongoDB data store. Twilio API integration delivers personalized market alerts and scheme notifications directly to users over WhatsApp.

Next.jsFlaskPyTorchMongoDBTwilio APIPython
1st Place (Winner) · Datathon 24, KJSCE (2,000+ Participants)
06
AI & Misinformation Detection

FactFinder

Code
Problem

The rapid spread of online disinformation poses grave threats to information integrity and public awareness.

Solution

A web-based platform assessing news article authenticity using an ensemble model combining Bidirectional Encoder Representations from Transformers (BERT) and classical machine learning classifiers, achieving ~95% accuracy on 72,000+ labeled articles.

Architecture

Pre-trained contextual BERT representations combined with Naive Bayes, Passive Aggressive, and Logistic Regression ensemble classifiers. Trained on the WELFake dataset (72,000+ articles) and deployed on Flask. Accepted and presented as a conference paper at Springer ICDSM 2024.

PythonFlaskBERTScikit-LearnHTMLCSS
1st Runner Up (2nd Position) · Hackanova 3.0 TCET · Published at Springer ICDSM
07
Computer Vision & Optical Measurement

Dimension Capture

Problem

Industrial 3D spatial measurement systems typically depend on expensive LiDAR hardware that is cost-prohibitive for lightweight field deployments.

Solution

An ArUco marker computer vision measurement system measuring single and multi-object dimensions from 10 cm to 5 m within 2 mm accuracy at ~90% lower hardware cost than LiDAR.

Architecture

Custom computer vision pipeline developed in Python and OpenCV that performs camera calibration, perspective transformation, and sub-pixel ArUco fiducial marker corner refinement to compute true 3D spatial dimensions with millimeter precision.

PythonOpenCVComputer VisionCamera CalibrationArUco
Registered Software Copyright (Government of India, 2025)
08
SaaS & Web Engineering

Eivora Suite

Live
Problem

Businesses struggle with sub-optimal SEO, slow web vitals, and fragmented API integrations that directly degrade customer acquisition.

Solution

A suite of industry-grade web tools including eivora.app, apios.eivora.app, and webaudit.eivora.app providing automated performance benchmarking, 100% SEO audit engines, and unified API management.

Architecture

Built with Next.js, Node.js, and modern headless CMS and caching layers. Delivers sub-second audit reports, automated Lighthouse metrics scoring, and optimized metadata generation pipelines.

Next.jsTypeScriptTailwind CSSREST APIsNode.jsSEO Audits
Production SaaS (Live)
09
Web Platform

SwapIt

Code
Problem

University students overpay each semester for textbooks and study materials that are used briefly and then discarded.

Solution

A zero-cost study-material exchange platform for university students, using Google Sheets and Drive as a serverless data and storage layer.

Architecture

Next.js frontend authenticated via Google OAuth, leveraging Google Sheets API and Drive API as a lightweight database and media storage layer.

Next.jsGoogle APIsOAuthGoogle SheetsTypeScript
Shipped & Open Source
Chat

Start a conversation