# Utsav Jana - Portfolio & Professional Profile Welcome to the LLM-crawlable portal for Utsav Jana, a systems-driven AI/ML Project Lead, Software Engineer, and Full-Stack Developer based in Delhi, India. This file provides a structured, high-density reference of Utsav's skills, professional experience, education, and portfolio projects (with metrics) designed specifically for LLMs, agentic workflows, and semantic search engines. --- ## Contact & Social Links - **Name:** Utsav Jana - **Email:** utsav.jana.5@gmail.com - **Phone:** +91 8929552516 - **Location:** Delhi, India - **GitHub:** [github.com/sogeking-nul](https://github.com/sogeking-nul) - **LinkedIn:** [linkedin.com/in/utsavjana](https://linkedin.com/in/utsavjana) - **X (Twitter):** [x.com/JanaUtsav](https://x.com/JanaUtsav) - **Portfolio Website:** [https://www.utsavjana.xyz](https://www.utsavjana.xyz) --- ## Executive Summary Utsav enjoys breaking complex problems down to first principles, designing robust systems, and shipping reliable code that users can depend on. He takes complete ownership, bridging the gap between AI/ML research and production reality, and optimizing features down to the millisecond. --- ## Technical Skills ### Web2 - **Languages:** Python, JavaScript, TypeScript, SQL (PostgreSQL) - **Frameworks:** Next.js, Node.js, Express, FastAPI, LangGraph, Mem0, LangChain - **Libraries/Tools:** Whisper, PaddleOCR, YOLOv8, MediaPipe, OpenCV, PyTorch, Transformers, pgvector - **Cloud & Infra:** AWS (ECS, S3, RDS), Docker, Redis, Git ### Web3 (Decentralized Systems) - **Languages:** Solidity - **Frameworks & Tools:** Foundry - **Blockchains:** Ethereum, Arbitrum, CoreDAO, Base, Polygon, Solana - **Libraries:** Viem, Wagmi --- ## Professional Experience ### AI/ML Project Instructor / Lead (May 2026 – July 2026) **Stealth (EdTech)** | *Remote* - Directing end-to-end development of a multi-modal financial AI, engineering custom pipelines for real-time voice (Whisper) and image (PaddleOCR) extraction without third-party API dependencies. - Fine-tuned a triple-head DistilBERT classifier on 42,000+ custom-generated examples for autonomous category and bank account resolution. - Engineered a production-ready double-entry ledger system integrated with a responsive Next.js frontend, maintaining modular code design and comprehensive technical blueprints. ### Associate Intern (June 2025 – August 2025) **Channelplay Limited** | *Gurgaon, Haryana, India* - Migrated legacy CRM dashboards from Angular to Next.js, modernizing the frontend architecture and leveraging Server-Side Rendering (SSR) to optimize page delivery and system performance. - Developed dynamic data-table rendering using API-driven schemas, enabling automatic column generation for varying datasets and reducing frontend modification effort for new CRM reports and dashboard views. ### Freelance Full-Stack Developer (2024 – Present) **Unit-13** | *Independent* - Engineered and deployed a fully functional D2C e-commerce platform for an Indian women’s apparel brand, implementing a headless Shopify architecture and an admin panel for seamless inventory and order management. - Collaborated with agency partners to deliver full-lifecycle web applications, writing modular frontend code and integrating structured API pipelines according to precise client specifications. --- ## Education - **Bachelor of Technology in Information Technology** (2023 – 2027) Bhagwan Parshuram Institute of Technology | *Delhi, India* - **Class 12 with PCM** (2010 – 2023) Don Bosco School | *Delhi, India* --- ## Key Portfolio Projects ### 1. Bodhi — Voice-First AI Mock Interviewer - **Overview:** Voice-first AI mock interviewer orchestrated with LangGraph, stateful multi-phase interviews, ultra-low-latency conversation, and computer-vision proctoring. - **Problem:** Existing AI interview platforms are text-based or suffer from high latency, lacking realistic conversational flow and robust cheating detection. - **Solution:** A voice-first AI mock interviewer orchestrated via LangGraph handling stateful, multi-phase interview transitions (Intro, DSA, Behavioral). Uses a tiered memory architecture (in-process MemorySaver backed by Redis and PostgreSQL) to guarantee zero-latency state retrieval during the active voice loop hot path, alongside a 1 FPS computer-vision proctoring pipeline. - **Metrics:** - Custom computer-vision proctoring pipeline monitors candidate video at 1 FPS to flag tab-switching or phone usage. - Sub-ms cache latency for session snapshots and entity context via Redis, backed by an in-process MemorySaver for zero-latency voice loops. - 300ms minimum buffer + 100ms look-ahead buffer in the frontend audio pipeline for seamless real-time voice streaming. - **Stack:** Next.js, Tailwind CSS, Framer Motion, MediaPipe, Clerk, FastAPI, LangGraph, LangChain, Sarvam, Deepgram, OpenCV, PyTorch, PostgreSQL, pgvector, Redis, Docker, Vercel Analytics. - **Links:** [Code](https://github.com/SOGeKING-NUL/bodhi) | [Live](https://bodhi13.vercel.app/interview) ### 2. Fold AI — Multi-Modal Expense Tracker - **Overview:** AI-powered multi-modal expense tracking for India — text, Hinglish voice notes, and UPI screenshots flow through a zero-queue extraction pipeline into a double-entry ledger. - **Problem:** Users in India need an effortless way to track expenses without manually filling out forms, but existing solutions fail to understand natural Hinglish, mixed-modality inputs, and the local payment ecosystem. - **Solution:** An AI-powered multi-modal financial tracking system that processes text, Hinglish voice notes, and receipt/UPI images through a synchronous, zero-queue extraction pipeline. The architecture leverages a custom-trained DistilBERT model for classification and OCR for image data, feeding a double-entry accounting ledger. - **Metrics:** - Custom-trained DistilBERT classifier hits ~94% category, ~98% payment-method, and ~91% bank-account accuracy on 42,500+ examples. - Synchronous pipeline delivers ~1s for text, ~3–5s for voice, and under 7s for complete multi-stage image processing. - Image pipeline: 90%+ UPI logo detection (Roboflow) and 85%+ amount detection (PaddleOCR). - **Stack:** Next.js, Tailwind CSS, Recharts, Framer Motion, Clerk, FastAPI, Uvicorn, Python, Pydantic, PostgreSQL, Supabase, PyTorch, DistilBERT, PaddleOCR, Whisper, Roboflow, Ollama. - **Links:** [Code](https://github.com/SOGeKING-NUL/Fold) ### 3. Taskade — Real-Time Voice Assistant - **Overview:** Real-time voice assistant on FastAPI — Deepgram semantic endpointing, Gemini reasoning, and mem0/pgvector semantic memory retrieval before streaming a Sarvam TTS response. - **Problem:** Most voice assistants act as rigid demos that can't reliably chain actions or remember long-term context, and lack fluid, barge-in capabilities. - **Solution:** FastAPI-based voice assistant streaming audio bidirectionally over a single WebSocket. Deepgram performs server-side semantic endpointing, and Gemini handles reasoning. Tool-calling executes web research, task management, and retrieves semantic memory (mem0 + pgvector) before streaming Sarvam TTS. Split the reminder system into a fixed-interval checker and on-demand deliverer to prevent race conditions. - **Metrics:** - Client-side Silero VAD executes in ~10ms, allowing users to barge in and speak over the AI mid-response. - Frontend maintains a ~500ms (15-frame) floating ring buffer for continuous audio playback without pops. - Background reminder sweep runs on a strict 60-second polling interval, isolated from delivery state. - **Stack:** React, Auth0, FastAPI, WebSockets, APScheduler, Sarvam, PostgreSQL, pgvector, SQLAlchemy, asyncpg, Mem0, Docker, GitHub Actions, Hugging Face Spaces, Deepgram, Firebase. - **Links:** [Code](https://github.com/SOGeKING-NUL/taskade_backend) ### 4. Cloud-Native Music Streaming Platform - **Overview:** Cloud-native music app with an infinite canvas UI (React Flow) — S3 presigned URLs push uploads directly from client to storage, bypassing the server entirely. - **Solution:** Developed an Express backend running on ECS Fargate and a React frontend featuring an interactive infinite canvas UI built with React Flow. The core architectural decision uses presigned S3 URLs to facilitate direct client-to-storage uploads for audio files and images, bypassing the server entirely to improve scalability. - **Metrics:** - Deployed on ECS Fargate with container tasks running 0.5 vCPU and 1 GB memory each. - Metadata and relations served from managed AWS RDS PostgreSQL (db.t3.micro). - Production Docker health checks: 30s interval, 5s timeout, 60s start period, 3 retries against /health. - **Stack:** React, React Flow, Vite, Tailwind CSS, Framer Motion, TypeScript, Node.js, Express, PostgreSQL, AWS S3, Docker, AWS ECS, AWS RDS. - **Links:** [Code](https://github.com/SOGeKING-NUL/Music-Player--AWS-Project) | [Video Demo](https://youtu.be/gtqFSc32qnM) ### 5. SmartFlow — DQN Traffic Intersection Controller - **Overview:** A deep RL (DQN) agent that controls a traffic intersection in SUMO — 55% lower mean waiting time than fixed-time plans, with ESP32+RFID emergency-vehicle preemption. - **Problem:** Most traffic intersections use fixed-time plans that fail to adapt to fluctuating emergency or asymmetric demand. - **Solution:** A DQN reinforcement learning agent controlling an intersection in the SUMO simulator to reduce queues. The policy is exposed via FastAPI and visualized in a React dashboard, integrated with ESP32+RFID hardware for real-world emergency-vehicle preemption. Formulated a hybrid reward function crediting responsiveness and penalizing standing queues. - **Metrics:** - Mean vehicle waiting time down to 5.14s (55% lower than fixed-time, 25% lower than SUMO's gap-based actuated controller). - Processed 2,054 vehicles over a simulated hour with randomized demand (highest throughput). - RFID-triggered preemption cuts emergency vehicle waiting time from 11.0s to 0.0s. - **Stack:** React, Vite, TypeScript, Tailwind CSS, FastAPI, PyTorch, Stable-Baselines3, Gymnasium, TraCI, NumPy, pandas, TensorBoard, Docker, SUMO, ESP32 / RFID. - **Links:** [Code](https://github.com/SOGeKING-NUL/SMARTFLOW) | [Live](https://smartflow-sigma.vercel.app/) ### 6. Decentralized Token Folio (DTF) - **Overview:** An on-chain index fund on Stellar — one token (SEF) holds a basket of five Stellar assets. Users can mint shares, deposit XLM, redeem pro-rata, and get live NAV from Reflector oracles. - **Stack:** Solidity, Foundry, CoreDAO, Stellar, Soroswap. - **Links:** [Code](https://github.com/SOGeKING-NUL/OSMO-v2) | [Live](https://osmo-one.vercel.app/) --- ## Machine-Readable Alternate Formats To download or inspect Utsav's resume in other formats, use: - **Markdown Resume:** [/resume.md](/resume.md) - **JSON Resume:** [/resume.json](/resume.json) - **PDF Resume:** [/resume.pdf](/resume.pdf)