Open to opportunities

Aansh User

@aanshuser1

Message

AI and Data Science student building deep learning and consent-aware intelligence systems.

India

Message

What I'm looking for

I’m looking for an AI/DS role where I can ship deep learning and speech systems end-to-end—prioritizing low-latency performance, robust evaluation, and production-grade workflows with auditability and reliability.

I’m a B.Tech. (AI&DS) student who turns research-grade ideas into working systems—deep learning models, low-latency mobile inference, and production-style AI pipelines.

At Samsung Research, I developed an attention-based deep learning model for gesture recognition and improved accuracy from 60% to 92%. I also built a Galaxy Watch companion app in Java with real-time sensor fusion and on-device inference via the TFLite Interpreter API, and engineered an AccessibilityService to map gestures to screen coordinates with sub-100 ms end-to-end latency.

On the voice intelligence side, I’m building a consent-aware AI voice processing platform (Meridian). I designed event-driven orchestration with Fastify, BullMQ, and PostgreSQL, including consent-first gating, auditability, speaker intelligence, and cross-call identity resolution using voice embeddings (pgvector + cosine similarity).

I also build evaluation and tooling that make models reliable in practice. For Truth on Trial, I adapted AASIST graph attention networks for cross-lingual deception detection and created an evaluation pipeline (ROC-AUC, Precision-Recall, confusion matrix, EER). For Workflow Clinic, I’ve built an LLM-backed multi-agent CLI to detect and patch portability gaps in open-source repos with rule engines, validation, and GitHub automation.

Experience

Work history, roles, and key accomplishments

Current

Voice Intelligence Infrastructure

Current

Meridian

May 2026 - Present (2 months)

Architected an event-driven, consent-aware AI voice processing platform with speaker intelligence and automated intake workflows. Built a distributed pipeline for call ingestion, consent verification, transcription, schema-constrained LLM extraction, identity resolution via voice embeddings, and confidence-scored human review with auditability.

TypeScript Fastify BullMQ Redis PostgreSQL Pgvector fastAPI Distributed Worker Pipelines Docker OpenAI API

Current

Samsung Research

Current

Samsung Research

May 2025 - Present (1 year 2 months)

Developed an attention-based gesture recognition model (60%→92% accuracy) using accelerometer and gyroscope data and contributed to planned integration into Galaxy Watch 8. Built a Galaxy Watch companion app with real-time sensor fusion and implemented an Android AccessibilityService to map gestures to screen coordinates and inject touch events with sub-100 ms end-to-end latency.

Accelerometer Gyroscopes scikit learn Java Android SDK TFLite Interpreter API On Device Inference Accessibility Services Sensor Fusion

Agentic Code Auditor & Auto-Fixer

Workflow Clinic

Mar 2026 - May 2026 (2 months)

Built an LLM-backed multi-agent CLI to detect and auto-fix portability gaps in open-source repositories with GitHub automation. Implemented a model-agnostic LLM client, a deterministic Pydantic schema-based rule engine with test coverage, selective LLM-as-judge validation, and end-to-end PR creation using GitHub REST APIs.

Python Pydantic Ollama OpenAI Anthropic Gemini Httpx GitHub API Pytest Ruff Mypy GitHub Actions LLM As Judge Deterministic Rules

Speech-Based Deception Detection

Truth On Trial

Jan 2026 - Apr 2026 (3 months)

Adapted the AASIST Graph Attention Network for cross-lingual speech deception detection by repurposing an anti-spoofing architecture for binary classification. Fine-tuned and evaluated on RLDD and RODeCAR and built an evaluation pipeline (ROC-AUC, precision-recall, confusion matrix, EER) with experiment tracking and visualization.