Aritro Roy

Hey! I'm Aritro, a backend and AI engineer.

Working atWhatbytes Technologies
Contributing topnpmPlanePatchwork
Buildingreqlite

I like turning messy data into systems people can rely on, and I'm happiest when the work is a pipeline, an API or a search problem. Right now I'm going deeper into Rust, AI systems and system design. Everything I've shipped and contributed is on the projects page.

Outside of programming, I play games, perform music live and lift weights, usually with too much caffeine. I'm based in Asansol, India. In case you're curious, here is the hardware and software I use, and what I'm up to right now. You can also find me on X, GitHub and LinkedIn.

GitHub Activity

1,000 contributions in the last year

OctNovDecJanFebMarAprMayJunJulAugSepOct
Mon
Wed
Fri
Less
More

Open Source

All contributions and projects →

Skills

Languages

PythonSQLJavaScriptRust

Frameworks & Tools

FastAPIDjangoAxumTokioSvelteDRFCeleryPyTorchSentence TransformersAirbyte CDKNangoLangfusePytest

Data Engineering & AI

RAG PipelinesElevenLabs STTVertex AIFAISSCross-Encoder RerankingHybrid Search (BM25 + RRF)EmbeddingsETLSchema Design

Databases

PostgreSQLSupabaseMongoDBMySQLRedisQdrantChroma DBPinecone DB

Cloud & DevOps

AWSDockerGitHub ActionsCaddy

Experience

  • Software Development Engineer

    Whatbytes Technologies | Remote

    • Architected a multi-tenant document intelligence platform for commercial real-estate leases on FastAPI — webhook-driven sync across Dropbox, Box, Google Drive and SharePoint (Nango, GCS) feeding a Vertex AI OCR and extraction pipeline that structures 100+ key points per lease into isolated Supabase schemas.
    • Cut end-to-end ingestion latency 60% by parallelizing Celery async stages and optimising schema throughput.
    • Designed Airbyte CDK ETL connectors (full and incremental syncs, 20+ SaaS entities) and a Django REST backend for an AI red-teaming platform, with zero-code YAML tool registration and dependency-graph onboarding.
    • Uncovered and independently verified the production fix for an OAuth defect blocking Google Drive integration in a Rust (Axum, Tokio, SQLite) governance daemon with a Svelte dashboard.
    • Engineered a Stripe-backed billing engine in Django REST on manual-capture PaymentIntents, atomic rollback and row-level locking (select_for_update), eliminating double-spend and duplicate-redemption races across every purchase flow.
    FastAPIVertex AICeleryAirbyte CDKRustStripe

    October, 2025 -Present

  • Backend Engineering Intern

    UrbanRider Technologies | Telangana, Hyderabad (Remote)

    • Resolved PostgreSQL query bottlenecks and REST API inefficiencies to cut application load latency 22%, restructuring indexing to hold performance under peak traffic.
    • Removed manual release steps by automating validation, regression testing and CI/CD with Python, Docker and GitHub Actions, ending environment-parity bugs between staging and production.
    PostgreSQLDockerGitHub Actions

    July, 2022 -May, 2023

Side Projects

  • 1.

    Voice-Enabled RAG Pipeline with Measured Guardrails

    A live voice-to-answer RAG service over a 99K-passage corpus (ElevenLabs STT -> FAISS -> cross-encoder rerank -> Claude) on AWS EC2, holding a P50 of 96ms against a 200ms budget. Cross-encoder reranking lifted recall@5 from 0.848 to 0.916, with per-sentence groundedness guards catching 100% of ungrounded answers at zero false refusals.

  • 3.

    AgentTuring

    A FastAPI query router dispatching requests across Qdrant vector and structured data stores, with tool-use agentic workflows and async concurrency for multi-user workloads, improving semantic query accuracy 25% over a keyword baseline.

  • 4.

    Prospect Pilot - Cold Email Automation

    An AI-powered cold email platform for automated outreach campaigns, with contact extraction, personalized templating, scheduling, and a response-tracking dashboard.

  • 5.

    OptiResume

    a cloud-native resume analysis tool with real-time PDF parsing and LLM integration.

  • 6.

    CourtScraper

    A Python scraper for extracting structured case data from public court records, built for reliable, repeatable data collection at scale.

Impact

60%

faster document ingestion after parallelizing Celery stages

96ms

P50 latency for the RAG pipeline, against a 200ms budget

25%

higher semantic query accuracy over a keyword baseline

0.848 -> 0.916

recall@5 after adding cross-encoder reranking

22%

cut in application load latency from index restructuring