I design and build backend systems that are resilient under load, safe with data, and predictable in failure.

My work spans multi-tenant SaaS platforms, asynchronous distributed architectures, real-time event systems, and cloud infrastructure on AWS and GCP. I care deeply about correctness, data integrity, and observability not just shipping features.

I think in transactions, queues, and service boundaries. Every system I build is designed to fail safely, scale predictably, emit useful signals when something goes wrong, and be maintainable six months from now by someone who isn't me.

Multi-Tenant Medical Shop SaaS

A production-grade B2B SaaS platform built for small medical shops across India. Designed with correctness and data integrity as first-class concerns — not bolted on after the fact.

Architecture Highlights

• Row-level multi-tenancy via shop_id across all guards, services, and queries
• RBAC with OWNER / MANAGER / STAFF role hierarchy
• Atomic stock deduction via PostgreSQL transactions
• FEFO batch rotation — First Expiry, First Out
• Integer-based monetary storage in paise — eliminates float precision errors entirely
• JWT auth with rotating refresh tokens
• Redis-backed async background processing
• Grafana dashboards for queue depth, job failure rates, and API latency
• Fully containerized for local dev and production deployment

Multi-Tenant RAG Platform

A self-hosted, multi-tenant retrieval-augmented generation platform: upload documents per tenant, query them, get cited answers — with strict tenant isolation and a fully local model stack, so tenant data never leaves the infrastructure and there's no per-call LLM API cost.

Architecture Highlights

• FastAPI backend with a layered service architecture (API → services → models → async workers)
• Tenant isolation via row-level filtering — every query scoped by tenant_id at the data-access layer, enforced through a shared auth dependency and covered by dedicated cross-tenant tests (a tenant cannot see or delete another tenant's documents, chunks, or API keys)
• PostgreSQL + pgvector for embedding storage and cosine-distance similarity search
• Retrieval pipeline: pgvector similarity search → cross-encoder reranking (local sentence-transformers, CPU-only) → generation
• Embeddings and generation both served by a local Ollama instance over HTTP — no external LLM API dependency
• Async document ingestion via Celery workers, with MinIO object storage and PyMuPDF / python-docx / BeautifulSoup parsing for PDF, DOCX, and HTML
• API-key based auth with create / list / revoke endpoints
• Prometheus metrics and Grafana dashboards
• 60 automated tests, including dedicated cross-tenant isolation coverage across the auth, document, chunk, and query endpoints

Status: Local-only — not yet deployed publicly. Foundational stack, auth/tenant provisioning, and the ingestion + retrieval pipeline are built and tested; public deployment and load testing are next.

Agentic Internal Support/Operations Assistant

An internal AI agent for support/operations staff that answers questions about accounts, orders, tickets, and SLAs by combining policy-document retrieval with live operational data — with account/role-scoped authorization and a confirm-before-execute workflow for the one state-changing action it can take.

Architecture Highlights

• Hand-built agentic tool-calling loop (not a framework's black-box runner) — needed to pause for human confirmation on one specific action, expose a UI-visible tool-activity trace, and stay provider-agnostic (Groq or Anthropic, swappable via one config value)
• RAG layer: Qdrant + local embeddings (fastembed, BAAI/bge-small-en-v1.5 — no GPU, no per-call API cost) over policy/SOP/product/agreement documents, each chunk tagged with a source-authority ranking
• Structured-data tools are parameterized Postgres lookups — no arbitrary SQL ever exposed to the model
• A deterministic calculation tool handles all timestamp/SLA-elapsed math — the model never computes a time delta itself
• Authorization is enforced at the tool layer, independent of what the model asked for
• Confirm-before-execute: the only state-changing tool stages a pending action; a human must explicitly confirm before it's created, with P1/high-value escalations additionally gated to a manager role
• 37 automated tests covering authorization, retrieval ranking, time calculations, the full confirm/reject action lifecycle, and the agent's tool-calling loop — run against a scripted fake LLM provider, so the suite needs no live API key

Status: Deployment-ready — Dockerfile and hosting config verified locally against a live stack (Postgres, Qdrant, and a live LLM-backed agent). Not yet hosted on a public URL.

opencode — AI coding agent, 200k+ GitHub stars.

Merged

• Fixed a dead ecosystem doc link (Daytona plugin reference) in the docs site
• Fixed a missing language link in the Arabic README

Submitted

• TUI event-stream reconnect after an OS sleep/lock disconnect
• Diff-view rendering fix for uppercase artifacts
• RPM package naming fix (packaged as open-code instead of opencode)

Design for scale before optimizing prematurely.
Transactions are the first line of defense for business logic.
Validation belongs at system boundaries — not scattered throughout.
Data models determine long-term system quality.
Services must fail safely, loudly, and observably.
Async processing reduces coupling and improves resilience.
You can't fix what you can't measure — instrument everything.