I build the systems that let teams trust their AI: eval pipelines, inference infra, and agentic dev tools β
alongside production payment orchestration for 125+ D2C brands.
At Juspay, I sit on the Xyne Team, building the AI/ML infra layer for the company:
- LLM Evaluation Platform β a GCP Batch pipeline (job queue in Postgres, Dockerized workers, spot-preemption-resilient) feeding a 4-service eval dashboard (frontend / backend / DB / eval runner) with live log streaming across SWE-Bench, Tau2-Bench, Polyglot, and more.
- Model Leaderboards β TPS/latency leaderboards on models like GLM-5, Minimax M2.5, and Kimi K2.5 that directly drive production model-selection calls.
- Xyne-CLI & Xyne Code β a TUI coding agent (Xyne-CLI, opencode-style) and a VS Code copilot-style extension (Xyne Code) built on top of it, both serving models via vLLM on rented GPUs (H200 / B300) through OpenAI-compatible endpoints. I built the
/commandslayer wired into the CLI engine. - LATAM Fintech News Agent β pulls context from internal Juspay repos to surface relevant LATAM fintech news automatically.
Before this, I spent ~1.5 years as a Product Solution Engineer owning checkout integrations for 125+ D2C merchants β the impact numbers below are from that work.
| Metric | What it means |
|---|---|
| 125+ brands | Checkout integrations across Shopify, Magento & WooCommerce |
| 10% conversion lift | vs. native Shopify checkout, via A/B tested routing |
| 95% payment success rate | Sustained through dynamic multi-gateway routing |
| 35% faster MTTR | Kibana + automated API-level debugging |
LLM Inference & Serving
vLLM SGLang β plus working knowledge of flash attention, paged attention, continuous batching, and KV cache internals. Hands-on hosting models on rented H200 / B300 GPUs.
Languages & Data
Python Java Bash SQL (Postgres/MySQL) Shopify Liquid
ML / AI
PyTorch TensorFlow Lite Hugging Face Pandas β model quantization, fine-tuning, agentic benchmarking
Infra & Cloud
GCP (Compute, Batch) Docker Kubernetes (k9s) tmux + SSH for parallel eval orchestration across VM/CPU fleets
π LLM Evaluation & Benchmarking Pipeline β Xyne Team Production eval infra spanning a GCP Batch job pipeline, a 4-microservice dashboard, and TPS/latency leaderboards used for real model-selection decisions. Also built tmux + SSH tooling to fan out evals across multiple machines in parallel.
π€ Xyne-CLI + Xyne Code β Xyne Team
Coding-agent tooling: a terminal agent (Xyne-CLI) and a VS Code extension (Xyne Code) on top of it, serving vLLM-hosted models on rented GPUs. Built the /commands interface wired into the CLI engine.
ποΈ Breeze Checkout & Breeze Automatic β Juspay Payment orchestration for 125+ D2C brands, plus an LLM tool-calling layer (Breeze Automatic) letting merchants configure checkout in natural language.
β Sign Signify β Personal project Android app translating sign language in real time via CNN (palm detection) + RNN (gesture recognition), with model quantization for edge inference.
π± Manas App β Personal project Offline-first Android mental wellness app with custom conflict-resolution sync logic. 1K+ Play Store downloads.


