Skip to content
SA

Work

Every claim here has a number attached to it.

Three roles across payments, telecom and benefits, plus the two things I'm building at the moment. Nothing on this page is aspirational.

Headline numbers

Messages/day
~10k
through the Message Center
Daily users
~6k
across three portals
Faster P95
40%
7.1s → 4.3s on low bandwidth
Fewer errors
60×
1.2% → 0.02% error rate

Career trace

Seven years, read as a waterfall.

Bar widths are proportional to actual tenure. Each nested span is one shipped outcome, what it involved, and the number it moved.

  1. Trajector

    Senior Software Engineer

    Aug 2024 — Present · 2 yr 1 mo

    • Message Center

      Architected and built a full-stack messaging platform unifying communications for 3 portals across 3 lines of business on a single shared database, shifting phone and email back-and-forth to async messaging.

      ~10k msgs/day · ~6k daily users · ~1,000 support hours saved/month

    • Customer tooling

      Built the Payment Center, VA-Benefits Payment Calculator and Appointment Center in Next.js with SSR and Tailwind, with responsive, accessible UIs across devices.

      +29% page interactivity

    • Next.js
    • Node.js
    • AWS Lambda
    • MongoDB
    • Tailwind CSS

  2. Jio

    Software Engineer II

    Feb 2023 — Jul 2024 · 1 yr 6 mo

    • Design-token pipeline

      Built a Node scraper that pulled design tokens from the Figma API into a generated TypeScript module, eliminating design-to-code drift bugs.

      ~2 hrs/release → near-zero

    • Scheduled digest service

      Replaced a proposed 24/7 EC2 cron with a Python/Django service on Azure timer-triggered Functions, delivering reliable weekly digest emails.

      ~95% cost reduction

    • 504 investigation

      Diagnosed intermittent 504s across a React → Node BFF → Django/Azure chain using OpenTelemetry correlation IDs and HTTP keep-alive tuning.

      error rate 1.2% → 0.02%

    • React
    • Node.js
    • Django
    • Azure Functions
    • OpenTelemetry

  3. Paytm Payments Bank

    Software Engineer

    Jun 2019 — Jan 2023 · 3 yr 8 mo

    • Low-bandwidth performance

      Cut P95 page load for branch staff on constrained networks via Brotli compression and HTTP/2.

      7.1s → 4.3s (40% faster)

    • RBI-compliant testing platform

      Built an employee testing platform with immutable attempt records; shipped 5 days ahead of the regulatory deadline and passed the RBI audit on first review.

      ~1,500 employees onboarded in week one

    • Data-loss fix

      Refactored Mongoose schemas to atomic $set/$inc operations with optimistic-concurrency versioning, ending a recurring class of production incidents.

      ~10 incidents/month → 0

    • Node.js
    • MongoDB
    • Mongoose
    • JavaScript

Building now

Two projects, honestly scoped.

Neither is deployed or pushed public yet, so each card lists only what actually runs today. The moment a repo goes up, its link appears here.

  1. 01

    AskMyDocs

    Retrieval-augmented Q&A over your own documents.

    In progress

    A RAG system built around the part that actually decides answer quality: the retrieval layer. Per-user chunks with 1536-dimension embeddings sit behind an HNSW cosine index, with user_id denormalised onto chunks so a tenant-filtered vector search stays a single index scan rather than a join.

    Running today

    • Postgres 17 + pgvector schema with an HNSW cosine index over 1536-dim embeddings
    • Content-hash caching so re-uploading an unchanged file skips re-embedding entirely
    • A queued → processing → ready | failed ingestion state machine
    • Grounded answers from a locally hosted LLM with context-only prompting and an explicit “not in your documents” fallback
    TypeScriptNext.jsPostgreSQLpgvectorDrizzleDockerOllama
  2. 02

    ChatQ

    The same chat service, modelled two ways: broker and RPC.

    In progress

    A multi-client pub/sub chat service on a RabbitMQ fanout exchange with per-client exclusive queues — then the same contract re-modelled in gRPC, specifically to feel the difference between broker-based delivery and direct RPC rather than read about it.

    Running today

    • RabbitMQ fanout exchange over AMQP/TLS with per-client exclusive queues
    • JSON envelopes carrying a UUID correlation ID through structured pino logs
    • Defensive parsing and self-echo filtering at the consumer boundary
    • A parallel gRPC/Protobuf contract with unary send and server-streaming subscribe
    Node.jsRabbitMQgRPCProtobufpino

Open to work

Hiring for something that has to stay fast under load?

I'm open to senior and staff full-stack roles, remote or in India. The quickest way to start is a short email about the problem you're trying to solve.

shreyalkhaniya97@gmail.comRésumé