I BUILD
INTELLIGENT
SYSTEMS THAT
OPERATE
IN THE REAL
WORLD.
I design and implement production-oriented systems across perception, retrieval, reasoning and automation.
System Perception Pipeline
End-to-end multi-sensor autonomous perception architecture engineered for real-time spatial estimation, deep metric re-identification, and downstream reasoning.
Object Detection — Bounding Box & Semantics
Spatial localization, anchor-free bounding box regression, and semantic classification of target entities across dense multimodal feeds.
6.8ms (GPU)
145 FPS
0.924 mAP@50
Systems I've built.
Production and experimental AI/ML systems engineered with mathematical precision, strict validation, and decoupled architectures.
Autonomous Vessel Tracking & Visual Re-Identification System
Production maritime perception pipeline integrating YOLO vessel detection, OC-SORT tracking, and LoFTR + DINOv2 feature descriptors deployed on live RTSP streams via FastAPI.
Real-time vessel tracking on live maritime RTSP video feeds suffers from false-positive track fragmentation, wake interference, wave backscatter, and severe visual appearance changes as vessels change bearing.
PTZ Camera + LiDAR Multi-Domain Sensor Fusion Node
Spatial perception and tracking architecture integrating 3D LiDAR point clouds with PTZ optical cameras via ROS 2 and Extended Kalman Filtering.
Optical PTZ cameras provide detailed semantic classification but lack direct metric depth, while LiDAR point clouds offer accurate range but lack dense semantic labels; calibrating moving PTZ optical axes with LiDAR presents spatial-temporal synchronization challenges.
MedBot: AI Medical Knowledge Assistant & RAG Pipeline
Grounded retrieval-augmented generation pipeline indexing medical documents in Pinecone vector database with LangChain orchestration, OpenRouter, and FastAPI.
Medical Q&A systems require strictly factual, grounded information retrieval from complex medical literature; naive language models hallucinate terminology, dosages, and clinical guidelines without verifiable passage grounding.
Dual-EMA Real-Time Perimeter Intrusion Detection Pipeline
Classical computer vision perimeter security pipeline utilizing dual-EMA background subtraction, contour tracking, and global lighting change suppression on live CCTV feeds.
Perimeter security CCTV feeds experience continuous false alarms triggered by sudden cloud cover, wind-blown foliage, shadows, and camera auto-exposure adjustments when using standard motion detection.
ASTRION: AI Logistics Support Agent
Source-grounded support agent for logistics operations that answers order, ticket and policy questions from SOPs and signed customer agreements, computes fees and credits in code, and changes nothing without confirmation.
Logistics support answers depend on order records, open tickets and contract terms, and figures such as cancellation fees and service credits must be exact. A model that writes those numbers itself, cites nothing, or changes records unprompted cannot be trusted or audited.
Engineering Experience
Verified technical timeline spanning production AI engineering on defense platforms at Tardid Technologies and computer science foundation at VIT Vellore.
AI Engineer
Full-timeMaritime Defense & Autonomous Perception
Engineering perception modules for the Brainbox AI product suite, serving Autonomous Surveillance Vessels and Perimeter Security nodes.
Computer Vision Intern
InternshipNaval Target Detection & Image Restoration
Investigated low-visibility perception enhancements for naval environments.
B.Tech — Computer Science & Engineering
EducationSystems & Artificial Intelligence
Core engineering foundations.
System Stack
Production technologies I deploy for perception, reasoning, and automation.
- LanguagesPython / TypeScript / JavaScript / HTML / CSSApplication & AI/ML development
- Computer VisionOpenCV / YOLOv8Detection, tracking & real-time vision
- Representation & MatchingDINOv2 / LoFTR / CNNs / RANSACRe-ID, feature extraction & image matching
- AI / MLPyTorch / Hugging FaceModel development & inference
- Generative AILangChain / Google Gemini / OpenRouterRAG, LLM applications & model orchestration
- Backend & InferenceFastAPI / Flask / Node.js / Express.jsModel serving, APIs & backend systems
- FrontendNext.js / React / Tailwind CSS / Framer MotionProduction web interfaces
- MobileFlutter / React NativeCross-platform applications
- DatabasesPostgreSQL / MongoDB / Supabase / FirebaseRelational, document & application state
- Vector DatabasesPineconeEmbeddings & semantic retrieval
- ORM / Data LayerPrismaDatabase access & schema management
- Robotics & 3DROS 2 / PCL / EKF / Point CloudsSensor fusion, state estimation & robotics
- Cloud & DeploymentAWS / Vercel / Netlify / RenderCloud infrastructure & deployment
- ContainersDockerReproducible deployment & services
- Automationn8n / WebhooksWorkflow automation & integrations
- APIs & ArchitectureREST / GraphQLService communication & integrations
- Real-Time SystemsSocket.IOReal-time communication
- CV Tracking & SurveillanceOC-SORT / Motion Gating / EMA Background SubtractionObject tracking & perimeter surveillance
| Domain | Technology | Application |
|---|---|---|
| Languages | Python / TypeScript / JavaScript / HTML / CSS | Application & AI/ML development |
| Computer Vision | OpenCV / YOLOv8 | Detection, tracking & real-time vision |
| Representation & Matching | DINOv2 / LoFTR / CNNs / RANSAC | Re-ID, feature extraction & image matching |
| AI / ML | PyTorch / Hugging Face | Model development & inference |
| Generative AI | LangChain / Google Gemini / OpenRouter | RAG, LLM applications & model orchestration |
| Backend & Inference | FastAPI / Flask / Node.js / Express.js | Model serving, APIs & backend systems |
| Frontend | Next.js / React / Tailwind CSS / Framer Motion | Production web interfaces |
| Mobile | Flutter / React Native | Cross-platform applications |
| Databases | PostgreSQL / MongoDB / Supabase / Firebase | Relational, document & application state |
| Vector Databases | Pinecone | Embeddings & semantic retrieval |
| ORM / Data Layer | Prisma | Database access & schema management |
| Robotics & 3D | ROS 2 / PCL / EKF / Point Clouds | Sensor fusion, state estimation & robotics |
| Cloud & Deployment | AWS / Vercel / Netlify / Render | Cloud infrastructure & deployment |
| Containers | Docker | Reproducible deployment & services |
| Automation | n8n / Webhooks | Workflow automation & integrations |
| APIs & Architecture | REST / GraphQL | Service communication & integrations |
| Real-Time Systems | Socket.IO | Real-time communication |
| CV Tracking & Surveillance | OC-SORT / Motion Gating / EMA Background Subtraction | Object tracking & perimeter surveillance |
Research Studies & Technical Experiments
Investigative empirical evaluations examining feature invariance, visual representation granularity, keypoint correspondence, and time-series anomaly detection.
Dense Cross-Attention Feature Matching under Maritime Fog & Low Contrast
Maritime vessel visual re-identification suffers when vessels change bearing angles or operate in dense fog, where traditional sparse corner detectors fail to produce repeatable keypoints.
Evaluated a decoupled matching pipeline pairing DINOv2 vision transformer patch representations with LoFTR coarse-to-fine keypoint correspondence and RANSAC geometric inlier filtering.
Tested on low-contrast maritime video feeds and degraded fog sequences to measure inlier correspondence ratio and false match rejection.
LoFTR cross-attention established reliable keypoint correspondence even across low-texture vessel hulls, while quality-adaptive thresholds effectively suppressed incorrect matches in heavy fog.
Applied SPAD & AOD Principles for High-Speed Naval Target Tracking
High-speed maritime target tracking under degraded atmospheric visibility requires microsecond-level temporal resolution and agile beam steering that conventional mechanical gimbal sensors cannot sustain.
Researched and modeled photon-timing arrival distributions and acoustic frequency modulation to simulate non-mechanical optical deflection for rapid target acquisition.
Analyzed simulated position estimation latency and photon arrival timing precision under low-photon naval defense scenarios.
AOD beam deflection modeling eliminated mechanical inertia delays in simulation, demonstrating theoretical capability for rapid scan pattern repositioning for high-speed tracking locks under low-light naval conditions.
Dual-EMA Background Modeling for CCTV Illumination Shift Suppression
Perimeter security CCTV cameras experience frequent false alarms due to sudden ambient illumination changes (e.g. cloud movements, auto-exposure shifts) when using single-rate background subtraction.
Formulated a dual Exponential Moving Average (EMA) background model maintaining two concurrent temporal rates $(alpha_{ ext{fast}}, alpha_{ ext{slow}})$ paired with whole-frame histogram variance thresholding.
Benchmarked false alarm rates and intrusion detection recall across continuous live CCTV surveillance video under variable weather.
Global lighting suppression eliminated whole-frame false positive triggers while dual-EMA separation successfully flagged unattended objects without alerting on sudden sunlight transitions.
Context Retrieval Grounding & Vector Search in Specialized Document RAG
Generic Large Language Models hallucinate factual statements and specific terminology when answering domain-specific queries without grounded passage verification.
Constructed a RAG testbed using LangChain and Pinecone vector search, evaluating top-k retrieved passage relevance and factual consistency against source medical literature.
Assessed context recall, answer groundedness, and hallucination rates across technical documentation Q&A.
Structuring vector chunking around natural document hierarchy boundaries improved semantic retrieval accuracy and ensured generated responses cited exact source context.
About & Technical Focus
A technical overview of my background in production computer vision, generative AI pipelines, defense surveillance systems, and systems engineering.
Sarthak Roy
AI / ML EngineerI am an AI / Machine Learning Engineer based in Bangalore, India, with hands-on experience building production Generative AI applications and defense-grade Computer Vision perception systems.
At Tardid Technologies, I engineer production ML/CV modules for the Brainbox AI product suite across three core surveillance and situational-awareness platforms: Autonomous Surveillance Vessels, AI Perimeter Security, and Networked Warfare Management. My work includes real-time vessel tracking (YOLO + OC-SORT) deployed on live RTSP video feeds, visual re-identification pipelines combining LoFTR and DINOv2 descriptors, classical CV intrusion detection via dual-EMA background subtraction, and PTZ camera + LiDAR sensor fusion with ROS 2.
In Generative AI, I build grounded Retrieval-Augmented Generation (RAG) pipelines and LLM-powered assistants using LangChain, Pinecone vector databases, OpenRouter, and FastAPI, alongside event-driven workflow automation with n8n.
I hold a B.Tech in Computer Science and Engineering from Vellore Institute of Technology (VIT), Vellore (2021 — 2025).
Developing reliable, accuracy-focused ML and computer vision systems running continuously on live RTSP video feeds and sensor-fused hardware.
B.Tech in Computer Science and Engineering from Vellore Institute of Technology (VIT Vellore, 2021 — 2025).
Grounded algorithmic design, verifiable RAG document retrieval, and transparent trade-offs between model accuracy and real-time execution constraints.
Let's Build Something
Interested in AI/ML engineering, perception pipeline development, or technical collaboration? Reach out directly.
Direct Contact Channels
Open to discussions regarding full-time AI/ML Engineering roles, computer vision perception challenges, or ML systems architecture.