No. AI-2026.05New Jersey — NY Metro AreaPortfolio · Preprint · May 2026
Systems that retrieve, reason, & remember.
§1
About Me
biography & research focus
Aditya Inamdar
AI Engineer & Developer
1 Felician University
2 Mainly.ai
I am an AI/ML Engineer and full-stack developer with hands-on experience building production AI systems from scratch. I turn research into working products that real people use. My work spans RAG assistants, real-time voice agents, GPU-optimized world models, and LLM alignment pipelines.
I thrive in customer-facing, ambiguous environments and take ownership of outcomes from discovery to deployment. I completed my Master's in Computer Science at Felician University (May 2026).
Target Roles & Availability
AI/ML Engineer · Applied ML Scientist · Software Engineer | NY Metro / NJ (Open to Relocation & Remote)
—Deconstructed Transformer and Hierarchical Risk Model (HRM) architectures using PyTorch to analyze layer-wise attention weights, improving model interpretability for high-dimensional financial and scientific datasets.
—Engineered automated visualization pipelines within a Graph-based IDE using Deep Graph Library (DGL) to track complex data lineage and multi-dimensional system dependencies.
—Led a cross-functional team of 4 developers to build, test, and deploy a secure mobile application using React Native, cutting the launch timeline by 30% using Agile sprints.
—Developed a Bone Cancer Detection model using Convolutional Neural Networks (CNNs), achieving a 15% increase in diagnostic accuracy over legacy baseline models.
—Engineered a high-throughput data preprocessing pipeline, reducing medical image analysis time by 30%.
[CNNs][Medical AI][TensorFlow][OpenCV][GSoC '23]
· 3 ·
§3
Technical Arsenal
taxonomy of frameworks, languages & systems
§3.1 AI & ML Frameworks
[PyTorch] Deep learning & autograd
[TensorFlow] Model training & CNNs
[MLX] Apple Silicon optimization
[Transformers] Hugging Face ecosystem
[LangChain] Agentic & RAG chains
[ChromaDB] Vector indexing & search
§3.2 Core Languages
[Python] Expert (Systems & ML)
[TypeScript] Full-stack applications
[JavaScript] ES6+ / Node.js
[C++] Low-level algorithms
[Java] Enterprise backend
[SQL] Relational data queries
§3.3 Specializations
[RAG Systems] Hybrid vector search
[LoRA / DPO] LLM alignment & SFT
[Hierarchical Models] HRM interpretability
[Real-time Voice] WebSockets & Audio
[Graph ML] Deep Graph Library (DGL)
[WebGPU] In-browser LLM inference
§3.4 Backend & Infra
[FastAPI] High-throughput microservices
[Node.js] Asynchronous backends
[Docker] Containerization
[AWS / GCP] Cloud deployments
[PostgreSQL] Relational storage
[Redis] High-speed caching
§3.5 Frontend UI/UX
[React] Component architectures
[Next.js] SSR & App Router
[React Native] Cross-platform mobile
[Tailwind CSS] Modern styling
[React Flow] Graph visualization
§3.6 Tooling & Ops
[Git & Linux] Version control & shell
[CI/CD] Automated pipelines
[Apple MPS] Metal performance shaders
[ITIL/SLA] Service management
[Jira/Agile] Sprint orchestration
§3.7
Leadership & Community Outreach
community building & student leadership
Jul 2023 — Feb 2024
Pune, India
President — Hack-X MIT-WPU
Science and Technology Division
—
Spearheaded developer community growth, hosted hackathons, initiated the Cyber Socratic podcast, and orchestrated corporate sponsorships.
Jan 2023 — Dec 2023
Pune, India
Core Engineer — Google Developer Student Clubs
GDSC MIT-WPU
—
Organized GDSC WOW & Cloud Native Days, participated in Google Cloud Jam, and delivered guest lectures on distributed systems.
· 4 ·
§3.8
Open Source Velocity
real-time GitHub contribution telemetry
GitHub Contribution Heatmap
· 5 ·
§4
Selected Work
figures 01–06, ordered by technical impact
Fig. 01 — Graph Knowledge RAGHigh ImpactRetrieval System
Nexus Graph RAG
A retrieval-augmented research assistant capable of answering complex queries over document collections while dynamically constructing interactive knowledge graphs from extracted entities.
Fig. 02 — Voice AI AgentLow-Latency StreamConversational Systems
Real-Time Voice AI Agent
Developed a real-time conversational voice agent supporting bidirectional streaming, tool invocation, interruption handling, and telephony integration with Twilio.
End-to-end alignment pipeline for fine-tuning LLaMA models using Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO), optimized for Apple Silicon hardware.
An end-to-end production-ready AI system for automated chest X-ray interpretation using deep learning, explainable AI (Grad-CAM), and LLMs for structured radiology reports.
Implemented a Dreamer-inspired Recurrent State Space Model (RSSM) trained on pixel observations. 5M-parameter model achieving 27.4x training acceleration using Apple Silicon MPS.
Implemented GPT-style transformer architectures and sequence-to-sequence models from first principles in PyTorch to analyze multi-head attention and autoregressive KV-caching.
