Hi, I'm
Vaatsalya
Srivastava.

Self-supervised cross-modal retrieval system using Vision Transformers with 2D-ALiBi/X-ALiBi positional biasing for SAR-to-optical satellite image retrieval.

• Achieved 0.5885 Cross-Modal F1@10 for SAR-to-optical satellite image retrieval.
• Optimized a multi-objective pipeline (JEPA predictive loss, DeCUR alignment, SIGReg regularization) over 100 epochs, converging to 0.1604 overall loss while preventing representation collapse.

Agentic pipeline that writes, executes, and iteratively refines Python scripts inside isolated Docker containers, with a real-time observability dashboard.

• Resolves tasks in a single pass with 2.2s end-to-end latency.
• Real-time dashboard tracks per-task latency, token usage (133 tokens/task avg.), and session cost for full pipeline traceability.

Cross-platform desktop automation toolkit driven by a TensorFlow model trained on MediaPipe hand-landmark features for OS-level gesture control.

• 91% gesture recognition accuracy across all gestures using 63-dimensional MediaPipe hand-landmark features.
• Collected 8-10 custom gestures across 500 samples, validated via stratified train-test splits.

AI-driven platform that ingests and indexes public repositories, orchestrating a vector, relational, and graph database in parallel to answer natural-language questions about a codebase.

• Achieved a 90% latency reduction (1.2s -> 120ms) through concurrent multi-layer orchestration across Qdrant vector, Postgres relational, and Neo4j graph databases.
• Engineered a multi-agent RAG pipeline via LangGraph, using orchestrator agents to classify developer intent and query decentralized context paths across graph and vector stores.

Flask-based clinical decision support system for rural and low-resource settings, pairing a local LLM with offline speech-to-text for multilingual symptom triage and basic image diagnostics.

• Integrated TinyLlama-1.1B with Vosk speech-to-text to achieve under 3s latency for symptom triage and X-ray/skin-lesion analysis in fully offline environments.
• Built automated referral PDF generation (jsPDF) and a JSON-based inventory system with real-time stock depletion alerts, cutting manual data-entry overhead by 60%.

Dual-path transformer model that isolates individual speakers from multi-speaker audio by fusing audio features with lip-motion cues extracted from video.

• Achieved 8.5+ dB SI-SDR on the multi-speaker LRS2 dataset.
• Optimized the model to 2.0-3.5M parameters using mixed precision and gradient checkpointing for constrained GPU memory.

Chrome extension for local compression and decompression of text, image, audio, and video files, using hybrid lossless/lossy algorithms with real-time savings analytics.

• Reached up to 76% size savings on WAV-to-MP3 audio (4.16:1 ratio) and 70% on plain text via GZIP-based DEFLATE.
• Verifies output integrity with SHA-256 hash comparison after every compression pass.

3D-printed, 12-DOF quadruped robot driven by an Arduino Nano that solves inverse kinematics in real time to translate target foot positions into coordinated joint angles.

• Runs a real-time IK solver at 50Hz via timer interrupts, non-blocking across all 12 servos on a 20ms refresh cycle.
• Supports forward/backward walking and turning via tripod-style gait sequences, plus standing, waving, and dancing behaviors.

Supervised browser automation agent that fills sensitive web forms on a user's behalf while keeping control over exactly what data reaches a remote LLM.

• Masks filled controls and uncertain screen regions with opaque pixels before any screenshot is sent for remote interpretation, gated behind explicit human approval.
• Stores profile data and documents in a local SQLite vault encrypted with AES-GCM, with voice input via Whisper requiring explicit consent.