Founder · Urja Labs
Sagar Gupta
AI Engineer · On-Device LLMs & Edge Inference
Engineer with 10 years building production mobile and streaming systems, now focused on on-device AI. I author open-source LLM infrastructure — a Kotlin Multiplatform engine that runs Gemma models fully on-device, and a private, offline-first RAG app on top of it: embeddings, hybrid retrieval, reranking, and source-grounded answers. Deep background in performance and low-resource inference from large-scale Android TV / IPTV work.
📍 Greater Noida, India · Open to full-time & contract — on-device, self-hosted & cloud AI, plus Android
The bet
Subscription AI turned everything into a metered cloud call. I think a large share of that work belongs on the device — private, offline, zero per-token cost — and that the hard, durable engineering is the runtime layer that makes on-device LLMs actually shippable. So I'm building it in the open: litertlm-kmp + NativeLM. The engine is the asset; the app is the proof.
Experience
Open-source on-device AI infrastructure — litertlm-kmp and NativeLM. Built litertlm-kmp, a Kotlin Multiplatform engine that runs Gemma-class LLMs fully on-device by wrapping LiteRT-LM, with GPU/NPU hardware delegation and model tiering by device capability. Built NativeLM, a private, offline-first local-AI chat app on top of it — no account, no upload, no telemetry. Implemented a full on-device RAG pipeline (EmbeddingGemma embeddings, hybrid dense + lexical retrieval, reranking) returning source-grounded answers, plus grounding fixes that stopped the model answering from the wrong document. Extracted the AI core into a reusable KMP library; added in-chat charts and an adaptive phone-to-tablet UI. Privacy-first by design — local encrypted backup (Argon2id → AES-256-GCM), no server; dual-licensed (AGPL-3.0 + commercial).
Airtel Xstream Play (IPTV) & Xsafe on Android TV, STB & mobile. Owned performance optimization for a large-scale, multi-module IPTV app on low-end (2 GB) devices — profiled and eliminated I/O chokepoints (JSON parsing, DB access). Architected a stale-while-revalidate “Smart PagingSource” to work around Remote Mediator limits and boost data loading. Engineered core Android TV / STB components including the Electronic Program Guide (EPG) and an OTA update system. Optimized Xstream streaming for +25% playback stability; led the migration to Jetpack Compose + Clean Architecture, cutting unit-test time by 30%.
Secure financial modules for the TataNeu super-app (Credit Line, Personal Loan) — biometric authentication + KYC flows (Karza, E-KYC) meeting banking-grade security standards.
Consumer apps for Maruti Suzuki (Rewards, True Value).
IoT / BLE integrations with connected hardware and printers.
A location-based services discovery app.
Skills
Education
MCA — Computer Software & Media Applications
Integral University, Lucknow · 2016–2019
Languages
- English (Professional Working)
- Hindi (Native / Bilingual)
Certifications
- Deep Learning Specialization — Coursera
- Android Developer Nanodegree — Udacity
- Design Patterns (Advanced / Creational) — LinkedIn Learning
Services
Available for full-time and contract work through Urja Labs:
On-device & edge AI
LLMs, RAG, and vision that run fully on phones and edge hardware — offline, private, no server. Android & iOS via Kotlin Multiplatform.
Self-hosted & private AI
Open models (Gemma and friends) and RAG deployed inside your own infra or VPC — data-sovereign, compliant, no third-party API.
Cloud AI applications
LLM features, RAG pipelines, and agents built on cloud APIs when scale or model size calls for it.
Android & mobile engineering
High-performance native apps and Kotlin Multiplatform libraries, backed by years of large-scale Android TV / IPTV / streaming work.
AI architecture & advisory
Help teams choose between on-device, self-hosted, and cloud — and design the pipeline end-to-end.
Get in touch
Hiring, or need help on a project — on-device, self-hosted, cloud AI, or Android? sgupta8874@gmail.com · LinkedIn.