AI/ML Engineer & Systems Researcher
B.Tech in Artificial Intelligence & Data Science · Uka Tarsadia University (CGPIT)
Cumulative CGPA: 9.25 / 10.0 (Ranked 2nd in Department) · Final-Year GPA: 9.78 / 10.0 (Sem VII: 9.57 · Sem VIII: 10.00)
I build robust, reproducible software systems that bridge algorithmic machine learning with production software engineering. My technical focus spans:
- Self-Supervised Representation Learning & Vision: Investigating architecture-native representations in transformer and convolutional vision backbones, multi-scale geometric losses, and attention gradient diagnostics under severe labeling constraints.
- Geospatial & Ecological AI: Engineering multi-source Earth observation pipelines (Copernicus Marine NetCDF, ETOPO bathymetry) paired with spatial group cross-validation to prevent spatial autocorrelation leakage in environmental suitability modeling.
-
Information Retrieval & Agentic Systems: Designing dense vector search microservices (SentenceTransformers, FAISS
$L_2$ -normalized inner product) and stateful multi-agent deliberation graphs (LangGraph, Model Context Protocol) with strict citation grounding to suppress hallucination. - Production Systems & Edge Computing: Deploying high-throughput containerized microservices (FastAPI, Docker, CI/CD automated gates), edge computer vision with Raspberry Pi hardware actuation, and commercial desktop software published on the Microsoft Store.
| Domain | Methodologies & Frameworks | Applied Systems & Architectures |
|---|---|---|
| Machine Learning & Deep Learning | PyTorch, TensorFlow, Scikit-Learn, XGBoost, YOLO (v5/v8/v12), Self-Supervised Learning (SimSiam, DenseCL), Representation Probing (CKA), Isotonic Calibration | Low-budget SSL pretraining, spatial cross-validation, gradient magnitude forensics, classification & detection |
| NLP, LLMs & Retrieval (RAG) | SentenceTransformers, LangChain, LangGraph, CrewAI, FAISS, Qdrant, Model Context Protocol (MCP), PEFT/LoRA | Grounded QA microservices, adversarial deliberation graphs, multi-tenant vector databases, clinical decision support |
| Computer Vision & Geospatial | OpenCV, MediaPipe, Copernicus Marine NetCDF, xarray, ETOPO Bathymetry, Rasterio, GDAL | Real-time optical gesture control, traffic density estimation, multi-temporal oceanographic feature grids |
| Backend, Cloud & Embedded | Python, C/C++, TypeScript, FastAPI, Flask, Docker, DigitalOcean, AWS, GitHub Actions CI/CD, Linux, Raspberry Pi GPIO | Asynchronous REST APIs, zero-downtime CI/CD deployment gates, hardware motor actuation, native C recompilation |
| Project | Domain & Architecture | Key Engineering Highlights & Verified Metrics | Codebase |
|---|---|---|---|
| native-attention-ssl | Vision Representation ResearchYOLOv12n · Area-Attention · PyTorch Hooks · VOC 2007 |
Investigated architecture-native SSL extracting Query/Key relations from YOLOv12's Area-Attention blocks under a fixed 500-image labeled budget with all 4,952 test images strictly quarantined. Across 3 seeds on a 0–100% mAP scale, downstream validation showed an inconclusive delta of +0.0416 percentage points in mAP50-95 (SD 0.2266 pp). Diagnostic gradient audits revealed native attention contributed ~1.16% of dense gradient magnitude, identifying signal attenuation in low-data regimes. | Repository |
| Strong-Seaweed | Geospatial Ecological AIXGBoost · Copernicus NetCDF · ETOPO · Spatial CV |
Fused 23 physical oceanographic predictors across 246 marine occurrences along the Gulf of Mannar shelf. Mitigated spatial autocorrelation leakage via 5-fold Spatial Group Cross-Validation. Achieved internal spatial-CV ROC-AUC of 0.9288 ± 0.1226, spatial AP of 0.9042, and out-of-fold Brier score of 0.0944 with automated CI release gates. | Repository |
| SandalQuest | Multimodal Speech AI & QAWhisper · PyTorch · BERT · IIIT Bangalore |
Colloquial Kannada Automatic Speech Recognition (ASR) and voice QA pipeline preserving indigenous sandalwood knowledge. Fine-tuned Whisper on noisy real-world recordings, evaluating accent and dialect robustness. Awarded Top 7 Finalist at the IIIT Bangalore ML-Fiesta Hackathon. | Repository |
| NyaySetu | Multi-Agent Legal Deliberation PlatformLangGraph · Dense RAG · FastAPI · React · Docker |
Full-stack legal reasoning platform pairing retrieval-augmented generation with a multi-agent adversarial deliberation graph (opposing counsel, neutral judge, client advocate). Features case-session memory, multi-turn state persistence, bilingual support, and Twilio/IVR telephone gateway integration. | Repository |
| Robo (Jerry) | Autonomous Vision RobotRaspberry Pi 4 · YOLO · OpenCV · MediaPipe · GPIO |
Physical autonomous mobile robot equipped with optical vision tracking and motor-driven actuation arm. Maps spoken natural language commands to real-time YOLO visual object detection and MediaPipe hand/face gesture recognition with low-latency GPIO pulse-width modulation (PWM). | Repository |
| mindmesh-traffic-vision | Intelligent Traffic Vision SystemOpenCV · Object Tracking · TypeScript · Vite |
Edge-capable computer vision system for intersection traffic analytics. Processes real-time multi-lane video streams to estimate vehicle density, queue lengths, optical flow velocities, and dynamic congestion indexes with interactive telemetry visualization. | Repository |
| cricket-legacy | Systems Engine RecompilationC/C++ · Android NDK · OpenGL ES · Memory Layout |
Low-level static recompilation and native Android port of a legacy 3D game engine. Reconstructs binary memory layouts, endianness translation, asset packaging pipelines, and cross-compilation toolchains for modern ARM64-v8a architectures. | Repository |
| CertiStock | Commercial Document AI (AinzaTechLabs)Python · OpenCV · Node.js · Microsoft Store |
Commercial production desktop software published on the Microsoft Store for textile manufacturing supply chains. Implements automated OCR alignment and inventory reconciliation pipelines, processing 500+ weekly certification sheets with 92%+ field-extraction accuracy across non-standard layouts. | Commercial Software |
- Google Student Ambassador: Selected through a competitive national process to lead campus AI innovation workshops and Google Gemini technology initiatives.
- National Hackathon Distinctions:
- Winner — Innosprint (2025)
- Top 7 Finalist — ML-Fiesta, IIIT Bangalore
- Top 10 Finalist — Codeversity Hackathon, IIT Gandhinagar
- Competitive Innovation Awards:
- Finalist (2024 and 2025) — The Maverick Effect AI Challenge (NASSCOM / DMFT)
- Shortlisted Finalist — Odoo National Hackathon 2025 (24-hr Offline Finals, Gandhinagar)
- Technical Bootcamps: Short-Term Training Program (STTP) on AI & Robotics (SVNIT) · Drone Technology Bootcamp (SVNIT).
├── 🔬 Research & Machine Learning
│ ├── native-attention-ssl # Self-supervised YOLOv12 area-attention representation learning
│ ├── Strong-Seaweed # Geospatial Earth observation & XGBoost suitability modeling
│ ├── SandalQuest # Multimodal colloquial speech recognition & QA (IIIT Bangalore Top 7)
│ ├── deepfake-detection # Media forensics & deepfake video verification pipeline
│ └── RAG-Conflict # Empirical document conflict reconciliation benchmark
│
├── 🤖 Autonomous Agents & Full-Stack AI
│ ├── NyaySetu # LangGraph adversarial courtroom deliberation platform
│ ├── law-ivr # Twilio telephony & Whisper voice consultation gateway
│ ├── orbix-your-ai-team-navigator # Multi-agent workflow coordination workspace
│ └── prompt-to-animation # Natural language to parametric animation engine
│
├── 👁️ Computer Vision & Robotics
│ ├── Robo (Jerry) # Raspberry Pi 4 autonomous vision robot with GPIO arm
│ ├── mindmesh-traffic-vision # Real-time traffic queue & density analytics platform
│ └── flappy-bird-cv # Optical gesture control interface via MediaPipe
│
└── ⚙️ Systems & Low-Level Engineering
└── cricket-legacy # C/C++ native game engine port for modern Android ARM64
Maintained with rigorous engineering standards, verified reproducibility, and empirical integrity.