Compute-Aware Mixture-of-Agents: Verifier-Gated Adaptive Aggregation under a Fixed Token Budget
Gupta, S. et al. Β· ICML 2026 SCALE Workshop, Seoul Β· Jul 2026
Building production multi-agent LLM systems and full-stack applications for enterprise clients worldwide
AI Engineering & Full-Stack Β· USA Β· Japan Β· India Β· DRDO
Shipping production full-stack apps and voice-agent products at Gnani.ai for US & Japanese enterprise clients
From mathematics and research to shipping production AI systems at scale β I design, build, and deploy full-stack applications and multi-agent LLM architectures for enterprise clients across the US, Japan, and India. I turn complex problems into reliable products.
β’ Solely designed, built, and deployed 3 production full-stack web applications in 2 months β customer-facing portals and voice-agent products
β’ Owned frontend (React/Next.js), backend APIs, auth, and cloud deployment end-to-end
β’ Delivered directly to US and Japanese enterprise clients, iterating on feedback to ship quickly and reliably
β’ Built Chatverse, an AI no-code automation platform that turns plain English into real actions across Instagram, Gmail, Calendar, HubSpot, Notion, and 15+ tools
β’ Shipped the full stack with 500+ active beta users and 3,500+ live automations β React + Tailwind on Firebase, backend on Google Cloud
β’ Designed and shipped a production multi-agent LLM architecture with guardrails, tool-use constraints, memory, and evaluation hooks β cutting unsafe responses
β’ Built evaluation and monitoring tooling to diagnose agent failure modes under adversarial and out-of-distribution queries
β’ Improved grounding using user history and live external data
β’ Led team of 5 developers for ServiceNow, Apple, and Amazon clients
β’ Enhanced LLM conversation quality by 35% through prompt engineering
β’ Built scalable dev tools using Python, Postgres, and Neo4j
β’ Led the end-to-end AI department β building and maintaining AI email automation, chatbots, and IDP pipelines
β’ Designed, deployed, and managed scalable AI models on AWS SageMaker, exposed as production-ready APIs
β’ Contributed to production AI systems and agent pipelines under the engineering team
β’ Gained hands-on experience with LLM integration, evaluation, and deployment workflows
β’ Processed 50,000+ unstructured datasets via automated web scraping + RAG pipelines to build a domain-specific AI
β’ Fine-tuned BERT and LLaMA for domain-specific document summarization
β’ Deployed Dockerized AI microservices with logging, monitoring, and CI/CD integration
β’ Developed AI autonomous agents for form-filling and chatbot applications
β’ Implemented NLP techniques for business process automation
β’ Streamlined workflows reducing manual effort significantly
β’ Improved PaddleOCR extraction precision by 20% using NLTK and OpenAI
β’ Applied custom NLP algorithms on financial reports
β’ Automated metrics extraction cutting processing time by 30%
β’ Designed chatbots using Rasa and Chainlit frameworks
β’ Leveraged large language models for natural user interactions
β’ Implemented conversational AI solutions for business automation
9+ hackathon wins and growing research footprint β a glimpse of the highlights.
Gupta, S. et al. Β· ICML 2026 SCALE Workshop, Seoul Β· Jul 2026
Gupta, S. et al. Β· ICML 2026 SCALE Workshop, Seoul Β· Jul 2026
Gupta, S. et al. Β· ICASET-2026 Β· 2026
SuperAI 2026 Β· NEXT Hackathon Β· UCWS Singapore
Back-to-back hackathons in Singapore: selected as 1 of 200 builders worldwide for the NEXT Hackathon at SuperAI 2026, then took that momentum into the UCWS Singapore Hackathon β a global open-source event with 150+ projects. Our team placed in the top 20 and won. Alongside co-founder S Akash, we built re-forge: a hardened operating procedure for coding agents (Claude Code, Cursor, Codex) that turns a single agent into a self-evolving team of specialists with adversarial gates, eval verification, durable memory, and evidence on disk for every decision.
Read on LinkedInMicrosoft Office Β· Pathway Hack Γ Tech Masters India
Won the Best Project Award at Hack for the Green Bharat at Microsoft Office. Pitched ChatVerse β our AI-powered automation platform that lets users automate anything across social media and apps using plain English β and received an internship offer on the spot. An incredible day of innovation, sustainability, and connections with passionate builders across India.
Read on LinkedInMeesho HQ, Bangalore Β· Team CodeShine (solo)
Among 50,000+ participants, Team CodeShine β a one-member team β emerged as a winner in Meesho's ScriptedByHer Hackathon. From a spark of an idea to presenting at Meesho HQ in Bangalore, this journey proved that even a solo team can shine bright with perseverance and passion.
Read on LinkedInNational Β· 5G & Beyond Hackathon
After building an AR/VR + ML prototype that peers called too ambitious, I submitted to the 5G & Beyond Hackathon and heard nothing for months. On Dec 30, an email said I was one of the top winners across India β I thought it was a scam. Lesson learned: back yourself, even when no one else does.
Read on LinkedInAmazon ML Challenge Β· Unstop
Secured All India Rank 87 out of over 75,000 participants in the Amazon ML Challenge 2024 on Unstop β competing at national scale in machine learning and OCR-focused problem solving.
Read on LinkedInCompetition-level Generative AI math reasoning system by fine-tuning 7B LLMs (Qwen2.5-Math, NuminaMath) using QLoRA (4-bit) and LoRA adapters. Integrated SymPy-based symbolic verification into inference to prune invalid reasoning paths, trained on 50K+ problems from a 1.7M OpenMathReasoning Dataset.
Multilingual LLM-powered healthcare assistant using prompt engineering and real-time clinical reasoning for symptom triage and personalized health guidance. RAG pipeline integrating 5+ health parameters reduced hallucinations by 30% through contextual medical retrieval.
Automates job applications directly on LinkedIn with custom filters (personal use; may trigger LinkedIn restrictions).
Secure AI knowledge agent integrating FAISS + Streamlit for interactive query handling.
AI-based Cybercrime Prediction System with 82% accuracy, integrating real-time risk scoring. Winner of Innotech 2023 Award
Fine-tuned PaddleOCR on custom datasets for table detection and content recognition, improving detection and classification accuracy by 17%.
Competition-level Generative AI math reasoning system by fine-tuning 7B LLMs (Qwen2.5-Math, NuminaMath) using QLoRA (4-bit) and LoRA adapters. Integrated SymPy-based symbolic verification into inference to prune invalid reasoning paths, trained on 50K+ problems from a 1.7M OpenMathReasoning Dataset.
Multilingual LLM-powered healthcare assistant using prompt engineering and real-time clinical reasoning for symptom triage and personalized health guidance. RAG pipeline integrating 5+ health parameters reduced hallucinations by 30% through contextual medical retrieval.
Automates job applications directly on LinkedIn with custom filters (personal use; may trigger LinkedIn restrictions).
Secure AI knowledge agent integrating FAISS + Streamlit for interactive query handling.
AI-based Cybercrime Prediction System with 82% accuracy, integrating real-time risk scoring. Winner of Innotech 2023 Award
Fine-tuned PaddleOCR on custom datasets for table detection and content recognition, improving detection and classification accuracy by 17%.
Have a project in mind or want to collaborate? I'd love to hear from you. Drop me a message and let's create something amazing together.