Software Engineer at Hero FinCorp · AI + Backend · Delhi
Ramandeep Singh
I build reliable backend systems and practical AI products—from Java microservices and lending infrastructure to agentic workflows and hybrid RAG.
~57% lower third-party cost98% lower repeated-query latency
01 / Selected work
Products and systems built around real problems.
The builder chapter · Product in progress
Requirements change. Test plans restart. Critical bugs slip through.
I’m building Testing Agent, an AI-powered workspace that makes software testing faster, clearer, and more reliable.
Explore Testing AgentHide product details
The problem
Testers often receive incomplete requirements, changing feature details, and scattered context. They spend valuable time deciding what to test, writing cases by hand, finding edge cases, and turning results into bug reports—while important scenarios can still be missed.
The approach
Testing Agent turns product requirements into structured testing guidance. It helps a tester understand the feature, generate relevant test cases, surface high-risk and edge-case scenarios, record evidence, and create clear bug reports developers can reproduce.
01Understand
02Plan
03Test
04Capture
05Report
What teams gain
- Less repetitive planning and documentation
- Better coverage of important flows and edge cases
- Clearer, more actionable bug reports
- Faster reproduction and fixes
- More confident, predictable releases
Human judgment stays central. Testing Agent is not designed to replace testers. AI supports the repetitive work so testers can focus on users, complex behaviour, and whether the product truly works well.
01 / MedOS Care · Healthcare workflow prototype
In progressMedical AI
A full-stack healthcare workflow prototype for finding doctors, booking appointments, joining video visits, drafting clinical notes, and supporting post-visit care conversations. It is built for development and evaluation, not as a certified medical device or production clinical system.
Patient + doctor portals→API gateway→Scheduling + telehealth→Scribe + care companion
- Connected care journeyPatient and clinician portals cover doctor discovery, availability, booking, profiles, video visits, and visit summaries.
- Clinical workflowRecording-event processing creates draft SOAP notes with auto-save, doctor review, approval, and note locking.
- Conversational assistanceAppointment-scoped care support and English, Hindi, and Hinglish voice booking help guide users through common workflows.
- Service architectureFastAPI microservices, gateway role checks, audit middleware, RabbitMQ events, Redis caching, and Qdrant-backed care-plan retrieval.
02 / Natural language BI · Secure data operations
CompletedEnterprise SQL Agent
A secure natural-language-to-SQL dashboard and collaboration platform that lets non-technical users query relational data, recover from SQL errors, visualize results, and monitor important metrics.
Natural-language question→Agent supervisor→AST safety checks→Read-only SQL + visual answer
- Self-healing queriesA LangGraph workflow analyzes database errors, rewrites failed SQL, and retries automatically.
- Security guardrailsSQLGlot AST checks block destructive operations while role-based filters mask sensitive personal data.
- Business-ready outputThe system creates narrative summaries, interactive charts, scheduled alerts, and PDF, spreadsheet, and slide exports.
- Live collaborationShared dashboard canvases synchronize cursors and board updates over WebSockets, with Slack webhook support.
03 / AI document intelligence / 2026
LiveDocuMind
A production-grade document search and Hybrid RAG engine that combines semantic retrieval, lexical precision, stateful memory, and operational guardrails.
Documents→Hybrid retrieval→RRF ranking→Cached answer
- Hybrid searchDense embeddings and BM25 fused through reciprocal rank fusion.
- Semantic cacheIntent-aware vector caching cuts repeated-query latency by 98%.
- Stateful memoryIsolated Redis-backed sessions with rolling two-hour expiry.
- Production controlsRate limiting, active invalidation, citations, and tunable retrieval.
04GitHub Repo ManagerDeveloper toolingCompleted
A feature-rich repository workspace with GitHub OAuth, real-time repository fetching, infinite scrolling, repository creation, detailed metadata, and AI-powered code reviews.
GitHub OAuth→Repository workspace→AI code review→Management actions
- Repository managementBrowse public and private repositories, paginate with infinite scrolling, create repositories, and inspect detailed repository information.
- Developer experienceGemini-powered code reviews, keyboard shortcuts, light and dark themes, configurable tags, file-size controls, and responsive loading states.
05GeoScope NYCGeospatial data visualizationLive
An interactive React and TypeScript map for exploring New York population density from GeoJSON data, with rich OpenLayers controls and responsive geospatial workflows.
GeoJSON data→Density filters→OpenLayers map→Interactive export
- Density explorationPopulation legends, hover tooltips, user-defined density filters, fullscreen viewing, and an overview-map toggle.
- Map toolingGeometry drawing, scale-unit controls, zoom controls, map download to PNG, responsive layouts, and an animated New York navigation action.
- DeliveryAutomated build and deployment workflow with the application hosted as a scalable static site.
02 / Experience
Backend ownership, applied AI, and product delivery.
Hero FinCorp
Software Engineer
Software Engineer working on fintech backend systems, digital lending workflows, microservices, cloud integrations, and AI and voice-assistant prototypes.
Core Backend & Lending Systems
- Worked on the Enterprise Lead Management System (ELM), a centralized lead management platform for digital lending workflows, supporting lead management, status syncs, workflow automation, and backend API services.
- Architected and developed the Banking Analysis microservice, orchestrating the Account Aggregator journey, consent collection, financial history analysis, and BRE decisioning for personalized loan offers and limit enhancement eligibility.
- Owned approximately 80% of the pre-qualified loan journey end-to-end across backend and frontend, including API integrations, orchestration flows, UI changes, testing, and production readiness.
AI & Voice Systems
- Built a real-time agentic voice assistant prototype using LiveKit, TTS/STT, and LLM orchestration for lending-related customer interaction workflows.
- Designed the assistant to understand customer intent, maintain conversational context, invoke backend workflows, and generate structured responses for customer journey support.
Cloud, Reliability & Platform Engineering
- Designed a cost-optimised API orchestration sequence for the pre-qualified loan journey, reducing per-customer third-party API spend by approximately 57% compared with the earlier bureau-first flow.
- Implemented Resilience4j patterns including retry, fallback, and circuit-breaker mechanisms to improve fault tolerance across microservices.
- Used AWS SQS and AWS Lambda for asynchronous processing and decoupled computation, improving scalability and service reliability.
Mapsense Technologies
Software Developer
Software Developer working on geospatial platform engineering, deployment workflows, frontend infrastructure, production debugging, and cloud and DevOps practices for location intelligence products.
Cloud, DevOps & Deployment Engineering
- Worked on deployment workflows for geospatial web applications, including environment-based configuration, production build optimisation, release validation, and post-deployment debugging.
- Containerized the frontend application setup using Docker to maintain consistent local development environments and reduce configuration issues across developers.
- Worked with Nginx-based frontend serving and reverse-proxy configurations for static asset delivery, API routing, CORS handling, and frontend-backend integration.
- Collaborated on CI/CD workflows using GitHub Actions to automate build, test, and deployment steps for frontend applications across environments.
- Supported production issue debugging involving API failures, environment variables, authentication flows, static asset loading, reverse-proxy issues, CORS errors, and deployment mismatches.
Performance & Reliability
- Improved performance of geospatial dashboards by optimizing map rendering, reducing unnecessary re-renders, and handling large datasets efficiently in data-heavy views.
- Debugged production UI and API integration issues using browser developer tools, logs, network traces, API responses, and environment-level checks.
Geospatial Product Engineering
- Built interactive geospatial dashboards using React, TypeScript, OpenLayers, Deck.gl, heatmaps, map layers, filters, geometry drawing tools, and data tables.
- Built user and organization management modules with role-based controls, configurable settings, optimized API request handling, and admin workflows.
coding75.com
Software Developer
Worked on frontend development for the crackDSA web platform, focusing on responsive UI, reusable components, routing, and user experience.
Frontend Product Engineering
- Built responsive web pages using React, CSS, Tailwind CSS, and React Router.
- Developed reusable UI components to improve maintainability and frontend development speed.
- Implemented client-side routing with React Router for smooth navigation across the platform.
- Improved layouts, responsiveness, spacing, and visual consistency across desktop and mobile screens.
- Collaborated with the team to collect feedback, fix frontend issues, and improve the overall quality of the website.
03 / Skills
A practical stack across AI, backend, and product engineering.
Backend
- Java
- Spring Boot
- Microservices
- REST APIs
- System Design
Cloud & DevOps
- AWS
- Docker
- Kubernetes
- CI/CD
- Nginx
- CloudWatch
AWS Services
- Lambda
- SQS
- Cognito
- S3
- API Gateway
- EC2
Databases & Search
- PostgreSQL
- MySQL
- MongoDB
- Redis
- Elasticsearch
AI Engineering
- LangChain
- LangGraph
- RAG
- Vector Search
- AI Agents
- LLM Orchestration
Frontend
- React
- TypeScript
- JavaScript
04 / Foundation
Harcourt Butler Technical University
Bachelor of Technology in Information Technology
8.7 CGPAKanpur, Uttar Pradesh
05 / Technical writing
Engineering notes from building AI systems in public.
Practical breakdowns of voice AI, retrieval systems, and the engineering discipline behind reliable AI agents.
Why perceived latency matters more than total latency in Voice AI
A practical look at changing a voice pipeline from batch processing to streaming so speech begins before the complete response exists.
Core idea
Read on LinkedIn Time to first audio shapes whether a conversation feels responsive; streaming STT, LLM tokens, and TTS chunks attacks the delay users actually notice.
Semantic caching for faster RAG without stale answers
The vector-space math, similarity threshold, and cache lifecycle behind DocuMind's intent-aware query cache.
Core idea
Read on LinkedIn An 82.5% cosine-similarity threshold reduced repeat-query latency from roughly 1,500 ms to under 20 ms, while active invalidation protects answer integrity after documents change.
Hybrid RAG: fusing dense and sparse search
Why embeddings alone miss exact identifiers—and how DocuMind combines ChromaDB semantic retrieval with BM25 keyword search.
Core idea
Read on LinkedIn Parallel dense and sparse retrievers, fused through a balanced ensemble, cover both conceptual questions and precise acronyms, filenames, or serial numbers.
Reliable AI agents need a stronger harness, not only a stronger model
A framework for diagnosing agent failures across task definition, context, tools, runtime state, and verification feedback.
Core idea
Read on LinkedIn The repository becomes the executable specification when success criteria, architecture decisions, and verification commands are visible and measurable.
More notes
10 articles- 05Generated TTS streaming is not file streamingVoice AI · LinkedIn
- 06Why a 75 ms TTS benchmark can still feel slowVoice AI · LinkedIn
- 07Five ways real-time voice agents reduce latencyVoice AI · LinkedIn
- 08Why browser voice AI is easier than phone-call AIVoice AI · Telecom · LinkedIn
- 09The repository is the system of record for AI agentsAI Agents · Context · LinkedIn
- 10Agents need better handoffs, not better memoryAI Agents · Continuity · LinkedIn
- 11Finished features beat more generated codeAI Agents · Delivery · LinkedIn
- 12Unit tests are not end-to-end proofAI Agents · Verification · LinkedIn
- 13Context debt: why clean agent handoffs compoundAI Agents · Harness · LinkedIn
- 14Building your first Chrome extension: a guide to manifest.jsonWeb Development · Medium
Open to AI, backend, and software engineering roles
Building something that needs reliable systems and practical AI?
Let's talk about how I can help your team ship it.