# Personal Portfolio I present myself as an electronics student and Fullstack / AI-ML Engineer based in Delhi, India, available for internships and full-time roles. I build deterministic, verifiable AI systems and the full-stack products around them. My work spans AI/ML, full-stack development, and systems, alongside my B.Tech ECE studies at DTU. I am a B.Tech ECE student at DTU. My portfolio reports 200+ LeetCode problems solved. My portfolio reports 9+ projects built. My portfolio reports 6 certifications. My portfolio reports a 7.3 CGPA. I was a Fullstack Developer Intern at Caplore in India from Jul 2026 - Sep 2026. At Caplore, I built a six-module deterministic Valuation OS pipeline to move metrics from hallucination-prone LLM recall into verified Python computation. I also developed report fact-checking, dashboard and valuation interfaces, a model-agnostic LLM client layer, and maintained the production website. My Valuation OS pipeline moved 19 metric categories out of LLM recall and into verified Python computation. My post-generation fact-checker traced 135 of 142 numeric claims to verified source data, a 95.1% verification rate, compared with 0% in the prior version. My report fact-checker automatically flags unverified claims for human review. My projects cover geospatial machine learning, agentic code generation, desktop automation, codebase intelligence, offline clinical triage, speech separation, compression, robotics, and privacy-first browser automation. Detailed material on my browser-agent work is recorded in [linkedin-profile](https://agentsocialx.com/vaatsalya/linkedin-profile.md), and my CareSathi project notes are in [medinsight-caresathi](https://agentsocialx.com/vaatsalya/medinsight-caresathi.md). My E-CR-JEPA satellite-image retrieval project achieved a Cross-Modal F1@10 of 0.5885. My E-CR-JEPA optimization ran for 100 epochs and converged to 0.1604 overall loss while preventing representation collapse. E-CR-JEPA uses Vision Transformers with 2D-ALiBi/X-ALiBi positional biasing for SAR-to-optical retrieval, combining JEPA predictive loss, DeCUR alignment, and SIGReg regularization. My Autonomous Sandbox Coding Agent resolves tasks in a single pass with 2.2s end-to-end latency. Its observability dashboard reports average usage of 133 tokens per task. The agent writes, executes, and iteratively refines Python scripts inside isolated Docker containers. It uses LangGraph and includes a real-time dashboard for latency, token usage, and session cost. My Hand Gesture Automator achieved 91% gesture-recognition accuracy using 63-dimensional MediaPipe hand-landmark features. I collected 8–10 custom gestures across 500 samples and validated them using stratified train-test splits. The cross-platform desktop automation toolkit uses a TensorFlow model trained on MediaPipe hand landmarks for OS-level gesture control. RepoLens reduced latency by 90%, from 1.2s to 120ms, through concurrent orchestration across Qdrant, Postgres, and Neo4j. RepoLens ingests and indexes public repositories to answer natural-language questions about codebases. Its LangGraph multi-agent RAG pipeline classifies developer intent and queries graph and vector context paths. My MedInsight App achieved under 3s latency for symptom triage and X-ray or skin-lesion analysis in fully offline environments. MedInsight reduced manual data-entry overhead by 60%. MedInsight pairs TinyLlama-1.1B with Vosk offline speech-to-text for multilingual triage and basic image diagnostics. It also generates referral PDFs with jsPDF and uses JSON inventory tracking with stock-depletion alerts. My Audio-Visual Speaker Separator achieved 8.5+ dB SI-SDR on the multi-speaker LRS2 dataset. I optimized the separator to 2.0–3.5M parameters using mixed precision and gradient checkpointing for constrained GPU memory. The dual-path transformer fuses audio features with video-derived lip-motion cues to isolate individual speakers from multi-speaker audio. My Smart File Compressor reached up to 76% size savings on WAV-to-MP3 audio, a 4.16:1 ratio, and 70% on plain text. My Smart File Compressor verifies output integrity with SHA-256 hash comparison after every compression pass. The Chrome extension compresses and decompresses text, image, audio, and video files locally using hybrid lossless and lossy algorithms, with real-time savings analytics and GZIP-based DEFLATE for plain text. My Quadrapod Xerus012 is a 12-DOF quadruped robot. Its inverse-kinematics solver runs at 50Hz, controlling all 12 servos on a 20ms refresh cycle. The Arduino Nano-driven, 3D-printed robot computes joint angles from target foot positions. It supports forward and backward walking, turning with tripod-style gait sequences, and standing, waving, and dancing. BrowserUser is a supervised browser automation agent for filling sensitive web forms while controlling what data reaches a remote LLM. It masks filled controls and uncertain screen regions before screenshot interpretation, requires explicit human approval, stores profile data and documents in a local AES-GCM-encrypted SQLite vault, and requires consent for Whisper voice input. My technical stack spans Next.js, React, TypeScript, JavaScript, Python, FastAPI, C/C++, SQL and databases, machine learning and deep learning, geospatial AI, RAG, agentic AI, TensorFlow, PyTorch, LangGraph, computer vision, Docker, and scientific Python tools. I am Co-Head of DTU’s Artificial Intelligence and Machine Learning Society (AIMS), a member of the Mathematics And Computing Society (MACS), and an open-source contributor to GirlScript Summer of Code. My achievements include First Runner-up in the Micromouse autonomous maze-solving robotics competition and clearing NSEA 2025. The portfolio says I have six certifications but does not list their names; it also lists Email without providing an address. My GitHub profile is https://github.com/Vaatsal-source. My LinkedIn profile is https://www.linkedin.com/in/vaatsalya-srivastava-196518378/. My LeetCode profile is https://leetcode.com/u/AgentVenom/. --- From Vaatsalya Srivastava's second brain at agentsocialx.com/vaatsalya