I build backend systems that handle real load — not toy demos.

The work I'm proudest of usually looks boring from the outside: a sharded write path that doesn't fall over at 1B rows, a workflow engine that uses Kahn's algorithm because cycles are bugs not features, a 300M parameter SLM model I trained from scratch on my laptop because I wanted to actually understand transformers — not import them.

I don't care about being right. I care about systems that stay up at 3am.

_{oct '25 — present  ·  Hyderabad, IN}

▸ Built a DAG-based AI workflow engine with Kahn's topological sort, per-step error policies (Fail/Skip/Retry), DLQ for failed runs, and 139+ integrations across 24 providers.

▸ Engineered a multi-tenant agent hosting platform — tenants connect via CNAME, platform dynamically provisions path-scoped Kubernetes routes per tenant, resolving host headers to isolated agent pods.

▸ Designed a human-in-the-loop interrupt system — agent pauses on low confidence, enqueues to SQS, support agent joins via WebSocket, end-user receives live status over SSE.

▸ Hardened Kubernetes Ingress with TLS, environment isolation, dynamic per-tenant provisioning.

Kubernetes LangChain LangGraph SQS WebSocket SSE

_{aug '24 — oct '25  ·  Remote, IN}

▸ Designed an ETL pipeline to ingest 1B+ rows with batched ingestion, transformation, and fault-tolerant loading into PostgreSQL at scale.

▸ Architected hash-based horizontal table sharding in PostgreSQL for writes; replicated to ClickHouse for OLAP reads, with fan-out queries aggregated at the app layer.

▸ Integrated Protocol Buffers for service-to-service contracts → 60% serialization gain.

▸ Designed compact binary encoding to replace JSON in hot paths → 40% payload cut.

▸ Migrated to Redis connection pool, exposed GraphQL APIs, engineered zero-downtime deploys via ArgoCD + OIDC + ECR.

PostgreSQL ClickHouse Protobuf Redis ArgoCD GraphQL

_{feb '23 — aug '24  ·  Remote, IN}

▸ Contributed to a high-throughput Go service powering the core Generative AI image pipeline — refactored goroutine concurrency, reduced median latency under load.

▸ Built an internal branch-preview deployment system in Go — engineers push a branch, get a live URL (Vercel-style, internal infra).

▸ Established a local GitHub Actions sandbox using act — feedback loop cut from minutes to seconds.

▸ Set up ngrok tunnels for webhook + 3rd-party API testing without staging dependencies.

Go GitHub Actions Docker ngrok

SLM Model · LinkedIn Post Generator

 

parameters:    300M
architecture:  decoder-only · custom hybrid
positional:    RoPE
attention:     multi-head causal · KV-cache
ffn:           SwiGLU
norm:          Pre-RMSNorm
tokenizer:     custom BPE (from scratch)
training:      bf16 · MPS · Apple M5

Built ground-up. No HuggingFace weights, no pretrained base. Tokenizer, transformer blocks, training loop — everything written from zero.

Quper · FinOps Platform

 

status:        in production
stack:         go · postgres · clickhouse
role:          led product team end-to-end
adopted by:    The Chatterjee Group
               (45-country conglomerate)
               Cricut (US consumer tech)

Real-time cloud cost intelligence. Took it from architecture to delivery. Currently running inside enterprise infrastructure across multiple Fortune-class clients.

Refactored core modules to NestJS, built repo analytics tooling for the API testing platform.

100+ tests added, full PostgreSQL upgrade, comprehensive API documentation rewrite.