Prabhanshu Pandit
LinkedIn GitHub

Principal Backend Engineer · Vancouver, BC

Prabhanshu Pandit

I build backend systems that keep working when the services they depend on don't. 13 years, from founding engineer at two startups to principal engineer at enterprise scale.

Email me LinkedIn GitHub
System overview diagram CLIENTS web mobile partner API GATEWAY token check rate limits correlation ID Rails platform 7 isolated prod regions Job scheduler fair, per tenant Microservices via one service client LLM + RAG pgvector · Qdrant CACHE fallback ↑ off by default. switched on when launch traffic hit 100× (Detail A) CB CB CB = circuit breaker
Isolated production regions
7
Requests / min, peak region
27k
Enterprise customers
~140
Years shipping backends
13

Figures for the enterprise learning platform I work on today.

Selected work

Problems I've solved in production

Six real problems from platforms I've worked on. Open any one for the problem, the fix and the result.

Startups

Zero to one, four times

On the founding team at two startups, technical lead for a third, and now a product of my own.

  1. 2013 – 2017

    Founding engineer · Learning-tech startup

    Built the multi-tenant platform in Rails, Node.js and Django and ran its AWS infrastructure. Moved 75+ customers to a new data model. Came back in 2021 and was there through its acquisition in 2022.

  2. 2017 – 2018

    Founding team · Health-tech startup

    Among the founding engineers of a clinical-trials platform for cancer immunotherapy research. Built its Node.js API services with Okta SSO on Google Cloud.

  3. 2025

    Technical lead · Bio-pharma SaaS startup

    Took the product from prototype to production in six months, leading three interns.

  4. 2026 – present

    Talentale · My own product

    Designed, built and run end to end. Live, with open sign-up. See below

Personal product · 2026 – present

Talentale

Live · open sign-up

An AI career platform that turns your real work experience into story-backed resumes and interview answers.

I designed, built and run it alongside my full-time job. It's live, and anyone can sign up and try it.

Visit talentale.io ↗

The problem

Most resumes are now AI-written and look the same, so keyword matching no longer separates strong candidates. And people with good experience still blank on the right story when an interviewer asks for one.

What it does today

  • Story Bank: a guided AI interview that draws out your career stories and turns them into STAR stories you review and approve.
  • Tailored resumes and cover letters for each job, built only from your own stories and profile.
  • Evidence gaps, not keyword gaps: Python can be on your skills list while no story shows production backend work in it.
  • Interview prep for each application, plus a prep copilot that coaches you using your own stories.
Talentale Skill Gaps screen showing 10 open, 11 claimed and 6 proven skills for a job
Demo dataSkill gaps. What the job needs that your stories don't show yet. Open, Claimed and Proven are set by fixed rules, not by an LLM.
Talentale Story Bank screen with STAR stories as cards
Demo dataStory Bank. Each story comes from a guided interview, then a STAR draft you approve.
Experience bullets from a Talentale resume tailored to a sample job
Demo outputTailored resume. Every suggested line cites a story or profile field.

Architecture

RAILWAY · PRODUCTION Browser Next.js no API keys FastAPI auth on every route all AI calls here credit metering Redis sessions · context Postgres + pgvector Claude Haiku 4.5 · Sonnet 5 OpenAI embeddings only Bugsnag · errors PostHog · analytics Resend · email one client for every call: cost · latency · failures IDs and event names only

Haiku handles fast, frequent steps like follow-up questions and job analysis. Sonnet handles the outputs where quality matters most: STAR stories, resumes and prep. Claude does all text generation; OpenAI is used only for embeddings, because Claude has no embedding model.

Engineering decisions

Candidate content stays in Talentale

Analytics and error tracking get IDs and event names only. A test fails if any prompt or completion text is captured.

Protected by default

The user always comes from the auth token, and auth is applied at the router level. A test fails if an endpoint that touches user data has no auth.

Every AI call is measured

One client for every Claude call tracks cost, latency, failure rate and cache hit rate per user and per feature.

Next: spoken mock interviews and story rehearsal, story-based job matching, and bringing stories in from your own AI chat via MCP.

Experience

13 years, newest first

From founding engineer at a startup to principal engineer on a platform serving enterprise customers in seven regions.

AI engineering and skills

Shipped AI, and what I build with

Side build for a startup · 2025 · Technical lead

Bio-pharma document platform

For a bio-pharma SaaS startup: a platform that lets teams search and ask questions across their documents in plain language. I took it from prototype to production, leading three interns, in my off hours.

6 moprototype to production
3tenant isolation tiers
3interns led

Django multi-tenancy with custom migration tooling, OCR ingestion, and a RAG pipeline on Qdrant and pgvector with a query router.

In my day job

LLMs in production

  • Built three GenAI proofs of concept in one hackathon: vector-based recommendations, RAG quiz generation and AI content creation. All three went onto the product roadmap.
  • Shipped production LLM features for recommendations and content tagging.
  • Use Claude Code daily, including for production work. Built the platform's Ruby service client in under two weeks with it.
  • Lead prototyping of AI agents and developer tooling for my team.

Skills

Languages
Ruby (Rails), Python (Django), Node.js, JavaScript, Java (J2EE)
Distributed systems
Microservices, event-driven design, Kafka, RabbitMQ, circuit breakers, rate limiting and fair queuing, caching, idempotent jobs
Data
PostgreSQL, MySQL, MongoDB, Redis, Memcached, pgvector, Qdrant
Infrastructure
AWS, GCP, Azure, Kubernetes, Docker, OpenShift, Ansible, Jenkins CI/CD
Observability
New Relic (p95/p99 latency, Apdex, error rates), ELK, correlation IDs, structured logging
Security
OAuth 2.0, JWT (HS256/RS256), SSO with Okta and SAML, service-to-service auth
AI engineering
Claude Code, Kiro CLI, LLM integration, RAG pipelines, MCP

Play

Take the pager for a minute

Three small games, each based on a real problem from the work above.

Traffic Router

Requests queue at the gateway. Send each one to a service before it times out. Overfill a service and its circuit breaker opens. Ten dropped requests and your error budget is gone.

Best 0
Keys 1 · 2 · 3 route the next request. Space starts. Tap a service on a phone.

Contact

Let's talk

I'm interested in Staff and Principal engineering roles where reliability and scale matter. Based in the Vancouver area and working remote.

Email me LinkedIn GitHub