Final-year Computer Science engineer building machine-learning systems end to end — live 3D reconstruction, video analytics, legal NLP and verified LLM agents, benchmarked on real data. Published at IEEE CONECCT 2026; first author of a paper submitted to NLLP @ EMNLP 2026; secondary track in robotics and PCB design.
Research Assistant — RV University, Core Engineering
Jan – May 2026
ARI Sync — real-time local-first AI coaching engine · Guide: Prof. Monesh N
- Converted instructional video into live form correction, running the entire ML stack locally on Apple Silicon — two 7B LLMs, Whisper and MediaPipe in under 10 GB with no cloud dependency
- Engineered the real-time feedback pipeline: 30 FPS pose tracking, ~2.5 ms median form-check latency, ~1.9 s from error onset to audible correction
- Architected async ingestion on FastAPI, Celery, PostgreSQL and Redis; React front-end running MediaPipe in WebAssembly for client-side pose detection
Research Fellow — RV University, Academics & Research
Feb – May 2024
Quantum Sciences · Mentor: Bharath Manchikodi
- Completed a semester-long quantum computing fellowship, earning the Microsoft Azure Quantum certification and working hands-on with Azure Quantum services and Q# programming
- Studied quantum mechanics, gates and algorithms via the Linux Foundation course; authored a 50-page report on applications in cryptography and optimization
Explainable LSTM Framework for Ego-Contingent Trajectory Prediction in Multi-Agent Systems
Jul 2026
IEEE CONECCT 2026, Bengaluru — presented; IEEE Xplore indexing pending
- Conditional sequence-to-sequence multi-input LSTM encoder–decoder treating a hypothetical future ego-plan as an interventional input, moving motion forecasting from correlation toward causal reasoning, explained with SHAP
- Scaled training from 10 to all 850 nuScenes scenes (1.2M rows) cutting ADE 13% (0.515 → 0.449) at 26.5 ms inference
When the Metric Measures the Bug: Silent Failure Modes in Diachronic Legal Concept Embedding
Aug 2026
First author — long paper submitted to NLLP 2026 @ EMNLP 2026; under revision
- Showed four silent embedding-pipeline defects that inverted a study's conclusion (22 of 25 legal concepts appeared to shift meaning before the fix, 1 of 25 after); pooling choice alone swings the verdict from 1/25 to 24/25
Live3D — live metric 3D reconstruction from any camera
2026
- Grows a metric 3D model live in the browser from a phone, drone or CCTV feed: Depth Anything 3 depth, verified loop closure (LightGlue), GTSAM Sim(3) pose graph, Open3D TSDF, calibrated confidence — no cloud GPU
- 108-run benchmark on 36 datasets (TUM, EuRoC, KITTI): sub-second live delay on 34/36, camera-path error ≤0.52% on TUM/EuRoC; 12-minute 4K street test; privacy blurring, 29-class labelling, GLB export
GLaw / LawVerse — AI legal intelligence for Indian law · team lead, capstone
2026
- Built a 231,908-document (666M-word) legal corpus and a 115,537-case citation graph with good-law labels; a Semantic Movement Index over legal-RoBERTa embeddings tracks doctrine drift across decades
- Shipped a Flutter app, Next.js site and FastAPI gateway with LLM document review in 12 languages: 98.4% type accuracy, 95.8% serious-defect recall on a 60-case synthetic eval set; 590+ tests
One Camera Brain — turns any existing camera into a smart camera
2026
- 21k-line edge-to-cloud video-analytics platform: rule packs, YOLOX/RF-DETR ONNX detection with tracking, VLM alert verification, signed evidence ledger and a virtual ONVIF camera for existing software
- Evaluated on 313 labelled real videos: 98% precision/recall on intrusion, 42/42 tamper events, 9.8 ms per 1080p frame; 439 passing tests; cloud tier load-tested to 8,000 nodes
AI E & PAMO — verified AI agents and local guardrails for coding agents
2026
- AI E: built five role agents (finance, data, HR, QA, DevOps) on a shared runtime with approval gates, sandboxed execution and MCP tools; deterministic engines compute every figure and a second model family must agree before an answer is certified — the data agent scored 98.4% on 250 questions with zero wrong answers certified
- PAMO: built a macOS daemon and Tauri app that answers coding-agent permission prompts using safety rules and a local LLM; across 3,699 real decisions it held 258 risky actions (secrets, payments, irreversible operations) for review
CreeperX & RC Aircraft — robotics and aeromodelling
2024 – 2026
- Designed CreeperX's 142-component, 4-layer controller PCB as code (SKiDL/KiCad) — three 5 A rails, reverse-polarity protection, battery telemetry, disarmed-at-boot servos and a JLCPCB fab package — with TFLite object detection on a Raspberry Pi 4. Designed and built a high-payload BLDC RC aircraft, competing in the Boeing National Aeromodelling Competition at IIT Kharagpur in 2024 and 2025
LanguagesPython, C++, Java, JavaScript, TypeScript, SQL, Dart
ML & AIPyTorch, TensorFlow, Hugging Face Transformers, LLM fine-tuning, RAG, LLM agents & MCP, ONNX Runtime, scikit-learn, SHAP
Vision & 3DOpenCV, MediaPipe, YOLOX, RF-DETR, Depth Anything, GTSAM, Open3D, LightGlue, three.js
EngineeringFastAPI, Flask, React, Next.js, Flutter, PostgreSQL + pgvector, MongoDB, Redis, ChromaDB, WebRTC
InfrastructureDocker, Kubernetes, Prometheus, GitHub Actions, Git, Microsoft Azure
RoboticsROS 2, KiCad / SKiDL PCB design, ESP32, Arduino, Raspberry Pi, sensor integration
B.Tech (Hons.), Computer Science & Engineering — RV University, Bangalore
2023 – 2027
Specialization in Artificial Intelligence & Machine Learning · Minor in Robotics & Automation · GPA 3.745 / 5.00