SUDHANSHU.SYSv1.0
New DelhiUTC +05:30
नई दिल्ली
Status
Sudhanshu Digital LabSystem 001
Senior AI EngineerLLM Systems & RAG Specialist
New Delhi, India नई दिल्लीUTC +05:30
I build production AI systems that actually ship.
AI systems, software products and interactive experiments.
Senior AI Engineer in New Delhi building production RAG pipelines, multi-agent systems and the backends that keep them honest — LangChain, LangGraph, vector search, LangSmith tracing, and a stubborn habit of measuring everything.
Anyone can call an LLM API. Almost nobody can keep it honest in production.
Connectors, scrapers and APIs with retries, backoff and schema-drift guards.
Evidence
Semantic chunking, embeddings and metadata for filterable retrieval.
Evidence
Agent graphs with explicit state, tool use and structured outputs.
Evidence
RAGAS harnesses on faithfulness, context precision and answer relevancy in CI.
Evidence
Traces on every node — tokens, latency, failure modes, guardrails and alerts.
Evidence
Request routing, caching, prompt compression and quantisation.
Evidence
Semantic chunking, cross-encoder reranking and metadata filtering — tuned against RAGAS, not vibes. Pinecone, ChromaDB, FAISS and pgvector in production at ~5K daily queries.
LangGraph-style agent graphs wired to real connectors — Instagram, LinkedIn, internal APIs — with explicit handling for rate limits, scraping edge cases and schema drift.
Distributed Node.js and FastAPI backends, async job queues for compute-heavy inference, intelligent request routing, response caching and prompt compression.
LangSmith tracing across every agent node, Arize AI model monitoring, RAGAS evaluation harnesses and A/B tests on prompt variants and retrieval strategies.
Multi-Agent Orchestration System
A multi-agent platform orchestrating Instagram and LinkedIn connectors for automated content generation and campaign analytics.
Owned the architecture of an AI platform that orchestrates multiple intelligent agents across Instagram and LinkedIn connectors for automated content generation and analytics, built on LangGraph-style agent graphs. Backend connector design consolidates fragmented platform data into a single centralised pipeline, with explicit handling for scraping edge cases, rate limits and schema drift — the three things that quietly break every social-data pipeline. Every agent node is instrumented with LangSmith tracing, and failure modes plus recovery strategies are documented in the repo rather than living in someone's head.
Image, Video & 3D Inference Backend
A unified backend fronting image, video and 3D generative models, with async queues for compute-heavy inference.
Architected a unified backend integrating image, video and 3D generative AI models with async job queues for compute-heavy inference. Provider-level optimisation and intelligent request routing reduced average inference latency by 35%; response caching and prompt compression cut cost per 1,000 requests by 25%. Provider routing strategies were validated with A/B tests rather than assumed.
LangChain + GPT + Pinecone
A production RAG chatbot tuned against RAGAS, not vibes: 85%+ faithfulness, sub-800 ms p95 retrieval at 10K documents.
A production-grade RAG chatbot delivering 85%+ RAGAS faithfulness, up from a 62% baseline, through semantic chunking, cross-encoder reranking and metadata filtering. Benchmarked at under 800 ms p95 retrieval latency at 10K documents and roughly $0.003 per query with caching. Full RAGAS evaluation coverage — context precision, faithfulness, answer relevancy — with documented eval runs and test cases, a LoRA fine-tuning experiment for domain-specific QA, and GGUF/Q4_K_M quantisation experiments on Mistral 7B for cost-sensitive deployments.
Technologies tied to the work they were used for. No proficiency bars — the evidence is the system each one shipped in.
22 product and experiment builds. Units 01–03 are the AI systems above, so numbering starts at 04.
3D Beverage Brand Experience
A fully responsive 3D brand site for a kombucha label, built with Three.js and GSAP scroll choreography.
A hypothetical juice brand brought to life as a 3D web experience. Built with Vite, React, Three.js and GSAP, the site sequences product reveals against scroll position, with a responsive fallback path so the experience holds together on mobile as well as it does on a desktop GPU.
Procedural Cinema Seat Simulator
Preview how a film actually looks from any seat, in-browser, before you buy the ticket.
A free cinema seat simulator that shows exactly how the picture reads from any seat — in the browser, before you spend money. Unlike the flat seat maps on booking sites, CinemaView drops you inside a procedurally generated auditorium built to realistic dimensions. You sit in first-person, drag to look around like a real patron, and switch between Standard, IMAX, Dolby Cinema, ScreenX and 4DX to see how premium formats change the experience.
Zero-Install Device Health Dashboard
Check your PC specs, GPU, RAM, battery and network from the browser — no download, no account, no data leaving the page.
A browser-based device health dashboard that reports specs, GPU, RAM, browser, battery and network connection instantly. No software to download, no account to create, no data shared — everything runs client-side through standard Web APIs. Built on Astro for a near-zero JavaScript baseline and deployed on Cloudflare Pages.
Kanban + AI Sprint Assistant
One place to track what you're learning, building and shipping — with an AI assistant that drafts tasks.
Built for developers who want a single place to track what they're learning, building and shipping without losing context. The Kanban board handles execution, the sprint panel enforces focus, and the AI assistant drafts tasks and removes ambiguity from vague tickets. Auth via BetterAuth, data via Drizzle on Postgres, media via Cloudinary.
Multi-modal AI SaaS Platform
Code, image, music and video generation behind one subscription product.
A versatile AI-powered SaaS platform offering code generation, image creation, music composition and video generation, wired to OpenAI and Replicate models. Subscription billing, usage limits and a unified generation history sit on top of a Prisma-backed data layer.
Project & Task Workbench
A bundle of developer utilities — task manager, project tracker and an in-browser code editor.
A collection of developer tools for creating and managing projects, including a task manager, a project tracker and an embedded code editor. Built as a Next.js app with a Prisma persistence layer and shadcn/ui component system.
Real-time Collaborative Canvas
A Figma-inspired design surface with live multiplayer cursors and canvas manipulation.
A collaborative design platform inspired by Figma, with real-time presence and multiplayer editing through Liveblocks and canvas manipulation via Fabric.js. Includes the essential shape, text and layer tooling plus a responsive interface that survives being used on a laptop trackpad.
Course Delivery & Payments
Full-stack LMS with course authoring, video delivery, Stripe checkout and an admin dashboard.
A comprehensive full-stack LMS covering course creation and management, video content delivery, authentication, payment processing and an admin dashboard. Cloud storage handles media, Docker handles deployment, and Prisma handles the schema.
Self-contained Local CMS
A CMS that runs entirely on your machine with its own database — install and start writing.
LocalPress runs entirely on your machine with its own embedded database. No external hosting, no complex setup — install it and start creating content immediately. Built on Next.js with SQLite as the storage layer.
PERN Teaching Workspace
Lets teachers manage classes, students and assignments in one place.
A classroom management application on the PERN stack, letting teachers manage classes, students and assignments from a single workspace. Auth through BetterAuth, admin surfaces through Refine, media through Cloudinary, and motion through Framer Motion.
Personal Finance Dashboard
Expenses, budgets, income and transaction history in a fast analytics-first dashboard.
A polished finance dashboard for tracking expenses, budgets, income and transaction history. Built as a fast Vite and Tailwind app with clean navigation, summary cards and analytics-focused views.
Vehicle Comparison Platform
Compare vehicles, explore specs and track upcoming models with advanced filtering.
A car inventory platform letting users compare vehicles, explore detailed specifications and follow new and upcoming models. Advanced filtering, real-time comparison and a CMS-backed content pipeline for spec data.
Language Learning SaaS
A gamified language platform covering Croatian, French, Spanish, Italian and Japanese.
A SaaS language-learning platform built over 47 days, covering Croatian, French, Spanish, Italian and Japanese with lesson progression, hearts, streaks and a leaderboard on a Postgres backend.
Product Price Scraper
Tracks and analyses product prices across multiple platforms with scheduled scraping.
A product scraping application that tracks and analyses prices across multiple platforms, using BrightData for resilient collection and scheduled jobs for price-history snapshots.
Scroll-driven Gaming Site
An immersive gaming landing page built on layered parallax and pinned scroll sequences.
An immersive gaming website driven by layered parallax and pinned GSAP scroll timelines, showcasing sequenced reveals, interactive UI elements and seamless section transitions.
Digital Asset Marketplace
Marketplace for Figma templates, fonts and design resources with a CMS backend.
A digital asset marketplace for purchasing Figma templates, fonts and design resources, built on Radix and shadcn primitives with a CMS backend for catalogue management.
Scroll-choreographed Theme
A responsive theme template built on GSAP page transitions and pinned scroll sequences.
A modern, responsive theme template with GSAP-driven animations, seamless page transitions and a polished shadcn component layer — put together as a reusable starting point for engaging marketing sites without giving up performance.
TMDB-backed Catalogue
Film and series browsing over the TMDB API, with a responsive player-first interface.
A movie and TV streaming front-end backed by the TMDB API for an extensive media library. Search, discovery rails and detail views sit on a shadcn component layer, with a player-first layout that holds up from phone to desktop.
Game Data Aggregator
Aggregates and browses detailed game data through a public REST catalogue.
A gaming information platform that aggregates and surfaces detailed game data through REST integration — search, filtering and rich detail views over a Next.js front-end with a shadcn component system.
Three.js Showcase
An interactive 3D portfolio with camera choreography and loaded scene models.
An interactive portfolio built around Three.js — smooth camera transitions between scenes, imported 3D models and a responsive layout that degrades gracefully where WebGL performance is limited.
In-browser Code Playground
A personal code playground with live preview and shareable snippets.
A code-editing playground built around the Ace editor, with live preview, snippet persistence and sharing — a scratchpad for testing ideas without leaving the browser.
Realtime Messaging
Realtime messaging with image sharing and authentication on a Firebase backend.
A full-stack chat application on React and Firebase — realtime message delivery, image sharing and authenticated sessions behind a clean, focused conversation interface.
12 commercial engagements.
IT, ERP & Enterprise Digital Transformation
Premium Real Estate Consulting
Staffing & Recruitment
Green Coffee Bean Export
Medical Diagnostics Platform
Innovation & Mentorship Platform
Ed-Tech & Professional Skilling
Enterprise Technology Services
Artisanal Fashion E-commerce
Ayurvedic D2C Commerce
Automobile Information Hub
Artisanal Fashion E-commerce
RAG RUN is a playable browser game inside the 3D studio. Text mode cannot run it; open the studio on a desktop browser.
The Lab is a bay of six shader and WebGL experiments inside the 3D studio — GPU-driven and interactive, so they have no faithful text equivalent. Open the studio on a desktop browser to run them.
6 AI-generated short films. All are published on @thinkai.io.
CH 01
Part one of a mythic short — generated cast, generated world.
CH 02
A neo-noir detective piece cut entirely from generated footage.
CH 03
A dark mystery, told in four minutes.
CH 04
A Spider-Man fan short — character consistency across every shot.
CH 05
The moment before the lights go out.
CH 06
Isolation, a signal, and something answering back.
Career history — four stations on the line, most recent first.
Gaup Media Pvt LtdNew Delhi नई दिल्ली
— Presentfull-time
Digitally NextNew Delhi नई दिल्ली
— full-time
Happymonk AIBengaluru बेंगलुरु
— full-time
Happymonk AIBengaluru बेंगलुरु
— internship
| Certification | Issued by |
|---|---|
| Generative AI with Large Language Models | DeepLearning.AI & AWS |
| Deep Learning Specialization | Coursera · DeepLearning.AI |
| IBM Full Stack Application Development | edX · IBM |
| Advanced React and GraphQL | Udemy |
Senior AI Engineer in New Delhi building production RAG pipelines, multi-agent systems and the backends that keep them honest — LangChain, LangGraph, vector search, LangSmith tracing, and a stubborn habit of measuring everything.
I own the AI layer end to end. That means architecture and retrieval design, prompt engineering, RAG pipeline tuning, deployment, evaluation harnesses, observability and cost control — plus the backend the whole thing sits on. In my last two roles I was the sole owner of that layer, which is a very different job from bolting an API call onto an existing product.
Because it is measured. I instrument RAGAS evaluation across context precision, faithfulness and answer relevancy, keep documented eval runs in the repo, and trace every agent node through LangSmith so token usage, latency and failure modes are visible in real time. On one production system that discipline took faithfulness from 62% to 85%+ — you cannot move a number you are not watching.
Both, and that is deliberate. Three-plus years of full-stack delivery means the AI layer I build ships behind real APIs, real databases and real frontends. Node.js, FastAPI, PostgreSQL, Next.js and React are all part of the same toolkit — and I have the Core Web Vitals and SEO work to show it isn't only backend depth.
Intelligent request routing, response caching and prompt compression, validated with A/B tests rather than guesswork. On production traffic that combination cut per-query API cost by 25% and average inference latency by 35% — cheap models for cheap work, expensive ones only where quality genuinely demands them.
Llama 3 and Mistral served via vLLM and Ollama, with LoRA/QLoRA fine-tuning for domain-specific tasks and GGUF/Q4_K_M quantization experiments for cost-sensitive deployments. Methodology and results are documented, not just claimed.
Senior AI / LLM engineering work where I own the layer rather than take tickets on it — production RAG, multi-agent systems, LLM infrastructure. I am based in New Delhi and work comfortably with distributed teams across time zones.
Email — sri.sudhanshu1@gmail.com — or LinkedIn. I reply to anything with a concrete problem attached, usually within a day.
Notes on agile practice and what engineering excellence actually costs day to day.
Pick a line and the form asks only what that job needs, plus a tentative budget. The request lands in my inbox, you get a copy, and usually replies within a day.
Marketing sites, dashboards and client portals in Next.js, React and TypeScript — built the way the client work in the archive was: fast, measured, and tuned until Lighthouse agrees. SEO and structured data come as part of the build, not as an afterthought.
A retrieval-augmented chatbot over your documents, website or database: semantic chunking, cross-encoder reranking, metadata filtering and a RAGAS evaluation harness so you can see how honest it is before your customers do. Delivered as a web widget, an API, or inside WhatsApp or Slack.
A personalised assistant or a small team of agents built on LangGraph-style graphs: it reads your calendar, inbox, CRM or social accounts, drafts, files and follows up, and asks before anything irreversible. Every step is traced in LangSmith, so when it does something odd you can see exactly why.
When prompting has hit its ceiling or the API bill has: a domain fine-tune with LoRA / QLoRA on an open model, a proper eval set to prove it beat the base, and a serving setup — vLLM or Ollama, GGUF quantisation where the hardware is small — that keeps the model and the data inside your walls.
For a team that already has an AI feature and cannot say how well it works or why the bill looks like that: tracing on every call, an eval harness that runs in CI, A/B tests on prompts and retrieval, and the routing, caching and prompt-compression that took a quarter off a production bill without moving the quality needle.
Distributed backends for AI products: async job queues for slow inference, connectors that survive rate limits and schema drift, REST APIs for the frontend, and SLA/SLO ownership once it is live. The multi-modal generation platform and the ThreeZinc connectors were built this way.
Interactive 3D on the web: product configurators, launch pages, portfolios and installations in Three.js and React Three Fiber, with the same performance discipline this room was built with — measured frame times, compressed assets, and a fallback for every device that cannot keep up.
The seven lines above are what I have shipped in production, but the interesting work rarely arrives with a label. Write down what you are trying to do, where it is stuck, and what a good outcome looks like. If it is mine I will scope it on a call; if it is not, I will say so and point you somewhere better.
Send me the problem — the retrieval that keeps hallucinating, the pipeline that fails silently, the bill nobody can explain. I'll tell you straight whether I'm the right person for it.
Usually replies within a day.