Ayush Gupta
AI Engineer • Backend Engineer
Building reliable AI systems and scalable backend infrastructure.
Final-year Computer Science student at IIIT Lucknow with experience across FinTech, Healthcare AI, EdTech, and Research. Worked on production-grade backend systems, multimodal AI applications, RAG pipelines, distributed workflows, and intelligent automation.
Where I've Built & Shipped
Production engineering across fintech, healthcare AI, EdTech, and research — from Kubernetes migrations to multi-agent RAG pipelines.
Backend Intern
Groww · Summer 2025
Problem
High-frequency stock reconciliation at Groww relied on 20+ manual operational workflows, causing delays and error-prone financial reporting across batch jobs.
Solution
Automated reconciliation by building a 3-step cross-service pipeline with Kafka event streaming, Redis caching, and Spring Batch processing. Migrated 15 legacy batch jobs into Kubernetes-based workflows with Helm charts. Investigated and resolved JVM OOM issues affecting production stability.
Impact
Eliminated dependency on 20+ manual workflows. Migrated 15 backend jobs to Kubernetes, improving deployment reliability. Resolved production JVM memory issues that caused service degradation. Improved system observability and operational reliability across the reconciliation pipeline.
Tech Stack
What I've Built
Production AI systems, retrieval pipelines, and backend infrastructure — each designed to solve a specific engineering problem.
Multi-agent diagnostic workflow combining vision analysis, hybrid retrieval from PubMed, evidence verification, and LLM reasoning with a clinical safety layer. Generates ICD-10-aware, citation-backed diagnostic reports from chest X-rays.
Pipeline
Retrieval-augmented AI assistant for financial analysis. Retrieves relevant financial context, understands document structure, and generates grounded answers with source attribution.
Pipeline
Real-time AI surveillance system for violence and anomaly detection in public spaces. Custom YOLOv8 pipeline achieving 94.8% mAP@50 at 30 FPS inference. Runner-Up at Hack-o-Fiesta v4, IIIT Lucknow.
Pipeline
End-to-end electronic voting platform with biometric MFA. Combines OCR-based identity extraction, face verification with liveness detection, MetaMask authentication, and blockchain-backed vote storage for tamper-proof elections supporting 500+ voters.
Pipeline
Retrieval-augmented restaurant assistant that answers menu, dietary, and availability queries using FAISS with sentence transformers over 200+ menu datasets. 92% answer relevancy with sub-second retrieval.
Pipeline
Other Noteworthy Projects
AI-powered PDF annotation system that extracts and highlights important information from documents using OCR and NLP. Early experimentation with intelligent document processing pipelines.
Pipeline
Intelligence at Scale
From retrieval-augmented generation to autonomous agents — production AI systems that reason, retrieve, and deliver.
Retrieval Augmented Generation
Production RAG pipelines with vector databases, semantic chunking, hybrid search, and retrieval evaluation.
Agentic AI
Multi-agent workflows with specialized agents for retrieval, reasoning, and verification — not single-prompt chains.
Multimodal AI
Vision-language systems that process medical images and text together for clinical decision support.
Prompt Engineering
Structured prompting, chain-of-thought, and ReAct patterns evaluated against faithfulness and relevancy metrics.
Vector Databases
FAISS, ChromaDB — indexing strategies, retrieval optimization, and latency reduction for production workloads.
LLM Evaluation
RAGAS, faithfulness scoring, answer relevancy — measuring and improving AI system quality systematically.
Agentic AI Architecture
How modern AI systems reason, retrieve knowledge, use tools, verify information, and generate reliable outputs.
Competitive Programming
Strong algorithmic fundamentals verified across 1,900+ problems on three major platforms — the engineering baseline for systems thinking.
Tools & Technologies
Production-proven technologies across AI engineering, backend systems, and cloud infrastructure.
Backend
9 technologiesAI / ML
9 technologiesCloud & DevOps
7 technologiesFrontend
4 technologiesVerified Expertise
Industry-recognized certifications in AI, cloud infrastructure, and data platforms — each independently verifiable.
Track Record
Competitive programming rankings, program selections, and leadership — evidence of consistent performance.
LeetCode Guardian
2200+ Rating
CodeChef 5★
Top Rated Programmer
Codeforces Expert
Expert Rating
Amazon ML Summer School
Selected Participant
Oracle Certified Professional
AI Vector Search
Hack-o-Fiesta Runner Up
94.8% mAP CV System
GDG ML Lead
IIIT Lucknow
Finance Club Coordinator
NIVESH, IIIT Lucknow
How I Approach Engineering
Principles forged from production incidents — not theoretical preferences.
Build systems that know what they don't know.
Grounded retrieval, verification layers, and hallucination reduction — AI systems should be honest about uncertainty, not confidently wrong.
Reliability before intelligence.
Observability, monitoring, and graceful failure handling come first. A smart system that crashes silently is worse than a simple one that fails loudly.
Keep architecture simple.
Modularity over complexity. Every component should have clear ownership, a single responsibility, and be replaceable without cascading changes.
What I Learned the Hard Way
Real mistakes from production systems — and the architectural changes that fixed them. Not theoretical, not from a tutorial.
Debugging JVM Memory Issues at Groww
Problem
Production reconciliation service kept crashing with OutOfMemoryErrors during high-volume trade processing.
Mistake
Initially assumed it was a heap size issue and just increased JVM memory — which only delayed the crash, not prevented it.
Learning
Traced the root cause to unbounded object retention in Spring Batch chunk processing. The chunk-oriented pattern was accumulating processed items in memory across step execution instead of flushing after each commit interval.
Outcome
Fixed by configuring proper commit intervals and item count thresholds. Eliminated OOM crashes entirely and stabilized the reconciliation pipeline in production.
Reducing Hallucinations in Medical-DSS
Problem
Early versions of the diagnostic system produced confident but medically incorrect outputs — dangerous in a clinical context.
Mistake
Initially relied on a single LLM call with a detailed prompt, expecting the model to self-regulate. It didn't — hallucination rate was unacceptably high for medical use.
Learning
Hallucination in high-stakes domains cannot be solved by better prompting alone. It requires architectural changes: separate retrieval from generation, verify outputs against sources, add a safety verification layer.
Outcome
Redesigned as a multi-agent pipeline with separate retrieval, reasoning, and verification agents. The verification agent cross-checks every claim against PubMed sources before it reaches the clinical report.
What Building My First RAG System Taught Me
Problem
First RAG implementation returned irrelevant results despite using embeddings — retrieval quality was worse than keyword search on some queries.
Mistake
Used naive fixed-size chunking without considering document structure. Embeddings captured paragraph-level semantics but queries needed information spanning sections.
Learning
Chunking strategy matters more than the embedding model. Semantic chunking that respects document boundaries dramatically improved precision. Hybrid search outperforms pure vector search for most real-world queries.
Outcome
Implemented semantic chunking and hybrid retrieval. Answer relevancy improved from ~70% to 92% in the Nugget assistant, and retrieval latency dropped 60%.
44+ Repositories. 5 Focus Areas.
Open source contributions spanning AI systems, backend engineering, computer vision, RAG, and distributed systems.
@Ayushlion8
Open source contributor · AI & Backend focus
Most Used Languages
AI Systems
Medical-DSS
Multimodal Clinical Decision Support System
FinRAG-Copilot
Financial AI Research Assistant with RAG
nugget_rag_based_chatbot
RAG Restaurant Assistant — 92% relevancy
AdaptIQ
Adaptive Learning Platform
CalmPulse
Mental Wellness AI Assistant
Contributions Beyond My Repos
Collaborating on real projects — not just opening PRs for typo fixes.
Calc For Everything
Built an interactive geometry visualization system using React and Plotly.js. Users define lines and coordinate points — the system dynamically determines same-side, opposite-side, and on-the-line relationships with real-time graph updates.
- •Dynamic line-point analysis
- •Real-time visualization
- •Mathematical grouping logic
- •Responsive graph rendering
Hacktoberfest 2023
Contributed to multiple open-source repositories across machine learning, web development, and tooling.
- •Open Source Collaboration
- •Code Reviews
- •Git Workflow
- •Cross-project Contributions
Academic Work
Deep learning research with measurable results — 82% RMSE reduction through attention-based forecasting architectures.
Forecasting Research Impact
Reduced forecasting RMSE by 82%.
Improved forecasting accuracy from 467K RMSE to 84K RMSE through systematic experimentation across traditional ML and attention-based deep learning architectures.
Forecasting Research Notebook
Google Colab NotebookComplete implementation of the retail demand forecasting research pipeline including data preprocessing, feature engineering, model experimentation, evaluation framework, and attention-based deep learning models.
This notebook is hosted as a Google Colab file. Click "Open with Google Colaboratory" at the top of the page to view the complete implementation and experiment code.
Time Series Forecasting with BiLSTM + Attention
Research on improving time-series prediction using bidirectional LSTM with attention. Achieved 82% RMSE reduction (467K → 84K) over baselines. Conducted ablation studies confirming attention mechanism contribution.
Research Resources
From Competitive Programmer to AI Engineer
A four-year evolution — each phase building on the last, each role demanding more ownership and architectural depth.
- Started Computer Science at IIIT Lucknow
- Competitive Programming Foundations
- GDG ML Lead — IIIT Lucknow
- Hacktoberfest Contributions
- PDF Annotation Tool
- Guardian Vision — 94.8% mAP
- Hack-o-Fiesta Runner-Up
- Computer Vision Projects
- Research Internship — 82% RMSE Reduction
- BYOL Academy — RAG Systems + Adaptive Learning
- Groww — Kafka, Redis, Kubernetes, Backend Engineering
- Oracle Certifications (4x)
- Amazon ML Summer School
- VE-Lyra Labs — Healthcare AI + Agentic AI
- Medical-DSS + Multi-Agent Systems
- Graduating IIIT Lucknow
Currently Building In
Active areas of focus — where I'm investing research and building time right now.
Let's Build Something
Meaningful Together
Whether you're hiring for an AI/ML role, building something in healthcare or fintech, or just want to talk about RAG pipelines — I'd love to connect.