Engineering High-Performance
Backend Systems & AI
Associate Software Engineer specialized in optimizing database Aggregations for 35% latency reduction, and building intelligent PDF/media search engines using vector indexing.
About Me
Profile Dashboard
A dashboard-style overview of my core metrics, technical specialties, and featured products.
Harsh Dahiya — Associate Software Engineer
I build reliable, low-latency APIs, database architectures, and generative AI systems. My experience spans logistics software engineering at Oryx, where I owned backend features from requirements to containerized deployments, and Web development at Ivory. I love performance optimization, clean code reviews, and automating complex workflows.
SpendSight
An AI-powered startup spend auditor. Built with Next.js 14, Supabase (with strict tenant isolation via Row Level Security), and Anthropic Claude. Features a deterministic TypeScript computation engine.
Bengaluru, India
Based in India's tech hub. Collaborating globally on performance engineering, API backend development, and generative AI RAG integrations.
Career
Work Experience
A progression of my software engineering career, showcasing key achievements and impact metrics.
Oryx International Logistics
Delhi, India (Bengaluru remote/hybrid)- Designed and built RESTful API modules for shipment tracking and vendor management using Node.js and Express.js, owning each service from requirements and design through deployment and production support.
- Implemented JWT-based authentication with role-based access control (RBAC), enforcing least-privilege access and securing cross-service API communication across all backend services at production scale.
- Optimized MongoDB aggregation pipelines for high-traffic reporting endpoints, achieving ~35% reduction in query response times through strategic indexing and pipeline restructuring.
- Containerized backend services with Docker and configured GitHub Actions CI/CD pipelines, standardizing deployments across development and production environments.
- Designed independently communicating backend microservices, applying microservice principles to isolate domain responsibilities across the production system.
- Actively used AI-assisted development tools (GitHub Copilot, Cursor, Claude Code) to accelerate development workflows throughout the engagement.
Bharat Electronics Limited (BEL)
Ghaziabad, India- Programmed Arduino-based microcontrollers using C/C++ to perform digital signal processing and sensor telemetry analysis for communication equipment.
- Troubleshot hardware and software integration issues in prototype systems, validating circuit designs and verifying signal integrity through oscilloscopes and hardware debuggers.
Ivory Communication
Delhi, India- Translated product and design requirements into responsive React.js UI components integrated with Node.js REST APIs, delivering polished frontend features end-to-end.
Portfolio
Featured Projects
A filterable showcase of my engineering projects, demonstrating systems engineering and product architecture.
Parsing large video, audio, and PDF documents, aligning page-level vector search citations, and preventing API bottlenecks during high volume concurrent requests.
Engineered a Python/LangChain platform with pgvector cosine similarity retrieval. Configured Redis for rate limiting (slowAPI), chat TTL caching, and map-reduce caching, achieving 95%+ backend test coverage.
Creating a secure financial audit pipeline that guarantees client math checks are 100% deterministic (no LLM hallucination risk) and isolates tenant data completely.
Created a pure-TypeScript computation engine bypassing LLM logic for math audit tasks. Structured separate anon/admin Supabase clients, enforcing tenant security using Postgres Row Level Security (RLS) policies.
Benchmarking and proving database latency improvements under concurrent high-density simulated logistics workloads.
Built a monorepo middleware with a Performance Lab. Benchmarked standard application-layer query maps (~200ms) against optimized MongoDB Aggregation pipelines ($match ➔ $group ➔ $project), displaying 35%+ speedups.
Intercepting LLM prompt injection threat vectors (OWASP LLM01), credential leaks, and runaway API billing without introducing significant network overhead.
Engineered a dual-layer guardrail featuring local regex scans (<3ms) and a Gemini Flash JSON classifier prompt (~250ms). Configured serverless rate-limiting via Upstash Redis sliding window.
Managing real-time bi-directional synchronization between user-directed manual form logging and complex multi-tool agentic chat commands.
Designed a React-Redux app syncing state with a FastAPI server running a LangGraph agent. Bound 8 specialized QMS tools (information extraction, risk assessment, duplicate detection) to orchestrate workflow checkpoints.
Processing high-concurrency transaction CSV ingestion asynchronously to avoid network request blockages while applying statistics and LLM diagnostics.
Architected a containerized FastAPI + Celery + Redis pipeline returning quick job execution status. Built a 5-stage worker process executing Gemini API batches of 20 with exponential retry backoff.
Mapping 20+ canonical financial statement items out of unstructured PDF documents with high accuracy and exporting formatted spreadsheets.
Designed a FastAPI backend using pdfplumber for coordinate mapping combined with Claude Sonnet extraction prompts. Coded an openpyxl formatting engine for confidence audits.
Cleaning noisy scans and performing cross-document verification of details like name, DOB, and formatting checks on Aadhaar/PAN IDs.
Coded a 7-rule cross-document validation pipeline in Python using OpenCV filters for scan preprocessing and GPT-4 prompts for name variation matching, achieving 95%+ extraction accuracy.
Designing interactive mobile-first content recommendations and classification with real-time natural language category tuning.
Built an Expo mobile client with interactive SVG trendlines. Linked Gemini API for zero-shot text classification and compiled categories weight tuning. Won Hackathon.
Enabling reactive live state updates, global room presence indicators, and concurrent notification syncing with low-latency socket pipelines.
Designed a combined websocket model — using Socket.io on the Node.js/AWS backend for raw low-latency sockets, paired with Convex on the Next.js frontend for live data state synchronization.