Lohitaksha Patary

Experience

Where I’ve worked.

Computer Vision InternCellium Networks · Oct 2025 – Present · Remote (US-based)

Surround-view perception and multi-camera calibration for autonomous driving.

Technical details

Stitched multi-camera vehicle feeds into synchronized panoramic Bird’s Eye View (BEV) output using OpenCV, homography transformations, and feature-matching algorithms.
Applied YOLO-based object detection on the stitched BEV for real-time obstacle identification in surround-view perception systems, enabling autonomous driving applications.
Implemented Kalman filter-based multi-object tracking across camera frames to maintain consistent object identities during cross-camera transitions.
Developed calibration pipelines for extrinsic camera parameter estimation, reducing stitching artifacts by 35% compared to naive concatenation.

AI / Accessibility InternIBM SkillsBuild · Nov 2025 – Jan 2026 · Remote[Live app]

Built Granite Accessible Assistant, a RAG kiosk for users with visual, auditory, and cognitive disabilities.

Technical details

Built a full-stack RAG-powered accessibility kiosk AI using IBM Granite 4.0 Micro as the LLM backbone, FAISS vector search for semantic retrieval, and SentenceTransformers for embedding generation.
Designed a profile-driven prompt engineering layer that dynamically adapts responses based on disability type (visual, auditory, cognitive) and user language preferences, incorporating few-shot examples and system personas per accessibility profile.
Integrated BLIP image captioning for scene understanding and MediaPipe hand gesture recognition for touchless interaction, enabling users with motor disabilities to interact without physical contact.
Engineered the complete RAG pipeline — document chunking, embedding generation, vector storage, similarity search, and context-aware response generation with source attribution.
Conducted accessibility testing with screen readers and keyboard-only navigation to ensure WCAG 2.1 AA compliance across the entire user interface.

Cloud EngineerGDG Amrita Vishwa Vidyapeetham · Sept 2024 – Present · Bengaluru, India

Google Developer Student Clubs on campus.

Technical details

Lead study jams and hands-on cloud workshops for peers, preparing cloud-native sample projects and deployment guides on Google Cloud Platform.
Mentor fellow students on cloud architecture, API design, and modern web deployment through practical demos and code reviews.
Organize weekly technical sessions covering containerization, serverless computing, and CI/CD pipelines with hands-on lab components.

Event OrganizerDastaan Tech Fest 2024 · Oct 2024 · Bengaluru, India

Primary host for IBM's cloud-native language workshop at Amrita's flagship technical festival.

Technical details

Coordinated end-to-end logistics including speaker liaison with IBM’s technical team, venue setup and equipment testing, participant registration, and live session moderation for 80+ attendees.
Facilitated hands-on coding segments where participants deployed their first cloud applications under guided supervision, troubleshooting in real time.

Projects

Things I’ve built.

ORCA — Optimized Routing & Car-park Assignment

Jan 2025 · Present

A unified multi-objective framework that solves vehicle routing and parking allocation as one connected problem, rather than sequentially as existing Parking Guidance Systems do.

researchPythonNetworkXOSMnxMOGATypst

Challenge & approach

Challenge — Existing Parking Guidance Systems treat routing and allocation sequentially, failing to capture the feedback loop where parking choices affect network-wide congestion.
Approach — Modeled the city as a dynamic weighted graph with the Bureau of Public Roads traffic function. Employed a custom MOGA with smart population seeding, capacity-aware mutation, and Pareto-based tournament selection to minimize driving time, walking distance, parking cost, and congestion simultaneously.
Outcome —
- 14.8% lower walking distance than FCFS while maintaining competitive driving times
- Validated on the real road network of Koramangala, Bengaluru — 3,865 nodes, 5,173 edges, 80 parking spots, 195 vehicles
- Monte Carlo simulations with Kruskal-Wallis and Mann-Whitney U-tests confirm statistical significance (p<0.001)
- Targets UN Sustainable Development Goals 11 and 9

PRISM — Probabilistic Scheduling Mechanism

2025 · 2026

A three-layer hybrid intelligent scheduling framework that makes predictive OS scheduling decisions instead of reactive ones.

researchPythonNumPyScikit-learnMarkov ChainsRandom Forest

Challenge & approach

Challenge — Traditional OS schedulers (Round Robin, FCFS, SJF) are reactive — they make no predictions about process behavior, resulting in unnecessary context switches and higher turnaround times.
Approach — Designed three layers: probabilistic pre-analysis with Markov Chains and Poisson arrival rates, ML classification via Random Forest into CPU-bound and I/O-bound types, and Weighted Moving Average sequence prediction — wrapped in an online learning feedback loop.
Outcome —
- Reduces context switches by leveraging predictive pre-analysis of process behavior
- Benchmarked against FCFS, SJF, Round Robin (quantum 4ms), and Priority Scheduling on five metrics including Jain’s Fairness Index
- The online learning loop continuously updates the transition matrix and model parameters as workload characteristics change

HandSlash — Touchless Gesture Recognition

2025

An AI-powered touchless interaction system using real-time hand landmark detection for camera-based gesture control.

productMediaPipeComputer VisionPythonTFLite

Challenge & approach

Challenge — Public displays, sterile environments, and accessibility kiosks need touchless interaction that is intuitive and responsive.
Approach — Built a real-time hand tracking pipeline using MediaPipe Tasks HandLandmarker with a TFLite float16 model bundle. Implemented gesture classification for grab, click, and type actions by analyzing 21 hand landmark coordinates per frame.
Outcome —
- Won First Place at Edge-Core Hackathon among 25 competing teams
- Sub-30ms inference on consumer hardware for real-time interaction
- Applies directly to accessibility kiosks, sterile environments, and public displays

GitHub

Granite Accessible Assistant

Nov 2025 · Jan 2026

A full-stack RAG-powered accessibility kiosk AI built on IBM Granite 4.0 Micro, serving users with visual, auditory, and cognitive disabilities through multi-modal interaction.

internshipIBM GraniteRAGFAISSBLIPMediaPipe

Challenge & approach

Challenge — Users with visual, auditory, and cognitive disabilities struggle with digital interfaces that are not designed for their specific accessibility needs.
Approach — Designed a profile-driven prompt engineering layer that adapts responses per disability type and language. Integrated BLIP image captioning for scene understanding and MediaPipe hand gesture recognition for touchless operation, on top of an FAISS retrieval pipeline with source attribution.
Outcome —
- WCAG 2.1 AA compliance across the entire user interface, verified with screen readers and keyboard-only navigation
- Deployed as a responsive Streamlit-based web application
- Full RAG pipeline: chunking, embedding generation, vector storage, similarity search, context-aware generation

GitHub Live

PharmaChain — Blockchain Medicine Tracker

Apr 2024 · Present

Blockchain-powered supply chain visibility for pharmaceutical batches, with tamper-evident provenance from manufacturer to patient.

productRustReactTypeScriptBlockchainMerkle Trees

Challenge & approach

Challenge — Pharmaceutical supply chains lack end-to-end visibility, making it hard to verify batch authenticity or prevent counterfeit drugs reaching patients.
Approach — Built Rust backend services exposing REST APIs to a React + TypeScript dashboard. Implemented role-aware batch provenance and medicine history tracking with Merkle tree verification for tamper-evident auditing.
Outcome —
- Role-based access control for hospital administrators, pharmacists, and regulatory auditors
- Merkle tree verification enables tamper-evident audit trails for every batch
- Real-time batch provenance tracking across the entire supply chain
- Capstone for the INSEAD Blockchain Opportunity Analysis certification

GitHub

IntelliQuery — Natural Language SQL Engine

Sept 2024 · Dec 2024

A natural-language to SQL layer that lets non-technical teams interrogate MariaDB datasets and receive annotated answers in plain English.

productJavaSpring BootLLMMariaDBJDBC

Challenge & approach

Challenge — Operations staff, HR, and academic admins needed instant answers from relational datasets but were blocked by the SQL skills gap and ad-hoc developer requests.
Approach — Built a Spring-powered pipeline that interprets natural language intents, assembles parameterized SQL via a hosted LLM, executes against MariaDB using JDBC, and narrates the resulting data back in plain English.
Outcome —
- Early user testing answered 120+ ad-hoc data questions without developer involvement
- Parameterized query templates prevented injection attempts across 200+ generated queries
- Plain-language summaries reduced follow-up clarification requests by 40%

GitHub

Code Royale — Competitive Coding Arena

2026

A real-time 1v1 competitive coding game — two players solve the same problem under time pressure and the first to pass all test cases wins the match.

productNext.jsSupabaseJudge0Realtime

Challenge & approach

Challenge — Competitive coding feels solitary — plain judges offer practice but no fast, social duel loop with progression that keeps players coming back.
Approach — Built a Next.js + Supabase arena with ranked and unranked 1v1 matchmaking, friend invites, bot battles, and a solo practice arena, judged server-side via Judge0.
Outcome —
- Persistent progression with ELO ratings, leaderboards, clubs, streaks, and badges
- Auth via Supabase with email/password and GitHub OAuth, plus realtime presence
- Friends list with live presence indicators and rich player profiles

GitHub Live

Audelle — AI-Powered Audio Editor

2026

An audio editor where everything is possible by prompting — the AI understands instruments, structure, mood, and speaker intent, then executes the edit from natural language.

productFastAPINext.jsExpoWhisperDemucs

Challenge & approach

Challenge — Traditional audio editors force creators to scrub waveforms and drag handles for every trim, stem isolation, or format cut — slow for podcasts, music, and short-form content.
Approach — Built an ML-backed pipeline (Demucs source separation, Whisper transcription, diarization, song-structure and mood analysis) behind a FastAPI job queue, with a Next.js web app and an Expo mobile client on the same backend.
Outcome —
- Natural-language edits: trim and arrange, stem isolate, inpainting, denoise, podcast filler-word removal
- Manual Editor fallback with waveform select, fades, gain, normalize, and speed control
- Exports stems as ZIP plus FCPXML/EDL for Premiere, DaVinci, and Final Cut

GitHub

Meetriders — Smart College Carpool

Jul 2024 · Present

A route-aware carpooling network that matches students and faculty leaving campus within the same 15-minute window, shrinking evening bus queues.

productNext.jsSupabasePostgreSQLEdge Functions

Challenge & approach

Challenge — Day scholars at my college had no reliable way to coordinate shared rides despite overlapping home routes and schedules.
Approach — Designed a Supabase data model with row-level security for rider and driver info, then added ride pinging, time-based matching, and notifications via Edge Functions and realtime. Matching considers route overlap and schedule compatibility.
Outcome —
- Beta with 60 commuters created 45 verified ride matches in the first two weeks
- Matching respects 15-minute windows and nearest pickup points, cutting average bus-wait time by ~35%
- Live dashboard continues to capture feedback and telemetry for the next iteration

GitHub Live

AutoTrace — Blockchain Vehicle History

2025

A decentralized vehicle maintenance and repair tracking system using blockchain and IPFS for tamper-proof service history.

productBlockchainIPFSTypeScriptSolidity

Challenge & approach

Challenge — Vehicle service records are scattered across workshops with no unified, tamper-proof history, making a used car’s maintenance background hard to verify.
Approach — Leveraged blockchain for immutable service event logging and IPFS for decentralized storage of repair documents, creating a verifiable chain of custody for every vehicle.
Outcome —
- Immutable service history on blockchain prevents record tampering
- Decentralized document storage via IPFS ensures data availability
- Complete vehicle provenance trail for pre-owned car verification

GitHub

Canteen Digitization

Feb 2024 · Jun 2024

Digitized Amrita University’s campus canteen with smart token queues, live kitchen displays, and cashier tooling so students see their orders progress in real time.

productNode.jsExpressMongoDBReact

Challenge & approach

Challenge — Paper tokens and verbal call-outs created 15㈂0 minute bottlenecks every lunch rush, leaving students guessing about order status and staff juggling duplicate tickets.
Approach — Co-led a four-person build introducing digital token assignment, cashier order capture, and synchronized kitchen and TV dashboards powered by Express APIs, MongoDB state, and frontends refreshing every 5 seconds.
Outcome —
- Pilot with 180+ students cut average wait time from 18 minutes to 7 minutes — 61% faster pickups
- Auto token allocator eliminated duplicate numbers and kept next-token predictions accurate to under a second
- Kitchen staff reported 30% fewer missed orders thanks to color-coded READY/PREPARING boards

GitHub

Cartfolio

2026

A cross-platform app that helps users track purchases and delivery status across multiple shopping and food-delivery services in one unified timeline.

productTypeScriptCross-platformExpo

Challenge & approach

Challenge — Online shoppers juggle multiple delivery apps with no single view of all their orders, making it hard to track deliveries and manage purchases.
Approach — Built a cross-platform application that aggregates order data from multiple services into a single timeline view with real-time delivery status updates.
Outcome —
- Unified timeline for orders across shopping and food-delivery platforms
- Real-time delivery status tracking in a single interface
- Cross-platform support for broad device compatibility

GitHub

Digital Porch

2025

A community platform inspired by the idea of a front porch as a place for sharing stories and connecting with neighbours.

productTypeScriptReactCommunity

Challenge & approach

Challenge — Local groups needed an inclusive digital hub that felt as welcoming as in-person porch conversations while staying lightweight for varying connection speeds.
Approach — Designed a clean feed with role-based moderation, instant notifications, and responsive layouts so neighbours of every age can share announcements and resources with ease.
Outcome —
- Community pilot onboarded 5 neighbourhood groups with 200+ shared posts in the first month
- Moderation queue cleared 95% of flagged posts within an hour thanks to role-based tools
- Lightweight asset strategy keeps page weight under 750 KB for users on 3G connections

GitHub Live

Monza Motors

2025

A motorsport-inspired ownership dashboard unifying vehicle specs, maintenance logs, and documents for dealers and drivers, with a 3D vehicle customizer.

productReactJavaScript3D Web

Challenge & approach

Challenge — Car owners and service partners lacked a streamlined way to access maintenance logs, documentation, and performance insights without juggling paperwork.
Approach — Architected a centralized vehicle hub with editable service logs, dealer integrations, and a document vault within a UI that echoes motorsport precision.
Outcome —
- Service reminders helped test users schedule two missed maintenance visits in the first fortnight
- Document vault cut file lookup time from minutes to seconds for dealer support teams
- Telemetry hooks primed the roadmap for predictive maintenance scoring in v2

GitHub Live

GitHub

A year of commits, across three accounts.

@lohitaksha06

@Einheit-Zenkai · @coding-royale

24 public repos19 followers

LessMore

259 contributions in the last year · snapshot from 30 Sep 2026 · github.com/lohitaksha06

Research

Published, and in preparation.

ORCA — Optimized Routing and Car-park Assignment via Multi-Objective Genetic Algorithmsin reviewLohitaksha Patary, Aditi Singhal, Nithilan Rameshkumar, Murali KarXiv, 2025 · extended version in review at IEEE T-ITS / ACM Computing Surveys · DZone write-up

A unified multi-objective optimization framework that jointly solves vehicle routing and parking allocation as a single connected problem — a gap not addressed by prior work. Existing Parking Guidance Systems treat routing and allocation sequentially, failing to capture the feedback loop where parking choices affect network-wide congestion. ORCA models the city as a dynamic weighted graph and employs a Multi-Objective Genetic Algorithm to discover Pareto-optimal solutions that simultaneously minimize driving time, walking distance, parking cost, and system-wide congestion.

Read the abstract notes

Written in Typst (IEEE format) with source code and preprints openly hosted on GitHub.
Addresses UN Sustainable Development Goal 11 (Sustainable Cities & Communities) and Goal 9 (Industry, Innovation & Infrastructure).

PRISM — Probabilistic Real-time Intelligent Scheduling Mechanismin preparationLohitaksha PataryarXiv, 2026 · target venue IEEE ICCA / ICCCS

A three-layer hybrid intelligent scheduling framework that addresses the fundamental limitation of traditional OS schedulers — their reactive nature. Conventional algorithms like Round Robin, FCFS, and SJF make no predictions about process behavior before execution, resulting in unnecessary context switches, cache misses, and higher turnaround times. PRISM combines Markov Chain probabilistic pre-analysis with Poisson arrival modelling, Random Forest classification of CPU-bound vs I/O-bound processes, and Weighted Moving Average burst forecasting, with an online learning feedback loop.

Read the abstract notes

Written in Typst (IEEE format) with source code and preprints openly hosted on GitHub.
Addresses UN Sustainable Development Goal 11 (Sustainable Cities & Communities) and Goal 9 (Industry, Innovation & Infrastructure).

How MOGA Solves the Urban Parking Assignment Problem: Inside the ORCA SystempublishedLohitaksha PataryDZone, 2025

A technical article bridging ORCA’s academic work and industry practice. It details how Multi-Objective Genetic Algorithms apply to urban parking optimization — graph modelling, objective function design, GA parameter tuning, and practical deployment considerations — written for software engineers and smart city developers.

Read the abstract notes

Written in Typst (IEEE format) with source code and preprints openly hosted on GitHub.
Addresses UN Sustainable Development Goal 11 (Sustainable Cities & Communities) and Goal 9 (Industry, Innovation & Infrastructure).

Achievements

Things that went well.

2025

First Place, Edge-Core Hackathon

Led a team of four to build HandSlash in under 36 hours — an AI-powered touchless hand gesture recognition system using MediaPipe Tasks HandLandmarker with a TFLite float16 model bundle. Won among 25 competing teams, recognized for technical innovation and clean implementation.

2025

Published on DZone

Authored "How MOGA Solves the Urban Parking Assignment Problem" — a technical article on ORCA’s multi-objective genetic algorithm approach, reaching one of the largest developer communities online.

2024

National Level 3rd Runner-up, Global Art Competition

Recognized as Regional Art Champion on the national stage.

2024

Second Place, Inter-school Piano Competition

High school piano competition finalist.

2024

Dastaan Tech Fest 2024 Host

Coordinated IBM’s cloud-native language workshop for 80+ attendees across the university’s engineering departments.

Certificates

Credentials and skill badges.

Generative AI Engineering Professional Certificate

IBM

Sep 2026

16 courses spanning AI, generative AI, NLP, PyTorch, Hugging Face Transformers, and RAG applications with LangChain.

Verify credential · ID 09D3CLZVBYB7

SkillsBuild AI Internship — Accessibility Technologies

IBM

Nov 2025 · Jan 2026

Built the Granite Accessible Assistant using IBM Granite 4.0 Micro, FAISS, SentenceTransformers, BLIP, and MediaPipe.

Blockchain Opportunity Analysis

INSEAD

Feb 2026

Four-course professional certification on blockchain strategy and opportunity evaluation under INSEAD and the Blockchain Research Institute. Capstone: PharmaChain.

Sustainability Virtual Internship

IBM

2026

Virtual internship program focused on sustainability applications of AI and cloud.

AI for Sustainability

IBM

Dec 2025expired Feb 2026

Google Data Analytics Professional Certificate

Google

Jan 2025expired Apr 2026

Nine courses, from Foundations: Data, Data, Everywhere through the capstone case study.

Google Cloud Skill Badges

Google

Oct 2025

Build a Website on Google Cloud · API Gateway · Pub/Sub · Cloud Storage · Looker · Google Workspace Tools.

Machine Learning with Python

IBM · Coursera

Dec 2025

Introduction to Deep Learning & Neural Networks with Keras

IBM · Coursera

Jul 2026

Accelerate Your Job Search with AI

Google · Coursera

Mar 2026

Foundations of User Experience (UX) Design

Google · Coursera

2024 · 2025

Seven courses, finishing with high-fidelity designs and a UX design concept.

Getting Started with Front-End Web Development

IBM · Coursera

Jan 2026

React Basics

Programming Fundamentals with JavaScript, HTML and CSS

Duke University

Jan 2025

Blogs

Things I’ve written.

Studio Ghibli and the AI Controversy: A Broader Warning About Creativity, Ethics, and Plagiarism in the Age of Generative AIReflected on the Ghibli backlash to examine where generative AI crosses ethical lines, and how creators can safeguard originality without halting innovation.Nov 2024 The Future is Decentralized: Understanding Blockchain's Core PrinciplesMapped out the foundational ideas behind decentralized networks, consensus, and why transparent ledgers are reshaping trust online.Oct 2024 Sovereign Cloud and Edge Computing: The New Frontier of Data IndependenceExplored why pairing sovereign cloud controls with distributed edge nodes gives regulated teams data residency, resilience, and low-latency analytics.Sep 2024 From Smart Contracts to Smart Economies: Blockchain’s Role in Building Autonomous SystemsOutlined how composable smart contracts can automate governance and value exchange, paving the way for self-running digital economies.Aug 2024 Building PharmaChain — A Blockchain-Based Pharmaceutical Supply Chain SolutionDetailed the PharmaChain blueprint that uses smart contracts to trace medicines end-to-end, tighten recall windows, and secure provenance evidence.Jul 2024 Smart Clouds, Smarter Code: What Developers Need to Know About AI-Powered PlatformsShared practical takeaways for dev teams adopting AI-assisted cloud services, from guardrails and observability to balancing automation with human oversight.Jun 2024 Project Update: Building PharmaChain – A Blockchain-Based Medicine Supply Chain TrackerDocumented the latest sprint where PharmaChain gained lot tracking, cold-chain anomaly alerts, and a patient-facing verification portal.May 2024 Tackling Blockchain Bloat: Can AI Help Prune Without Compromising Trust?Explored AI-assisted pruning strategies that lighten ledger storage while preserving verifiability through zero-knowledge proofs and shard snapshots.Apr 2024 Meta’s Use of Synthetic Data: Revolutionizing AI in the MetaverseLooked at how Meta trains immersive AI experiences with synthetic datasets, and why governance controls matter as virtual worlds mix real and generated content.Mar 2024