Lead Founding Engineer at Quper

I'm Ayush Kumar, a software engineer who designs and ships AI agent platforms, multi-cloud data pipelines, and FinOps tooling, then deploys them alongside the clients who use them.
I build the systems AI products run on: workflow engines wired into 139+ tools, pipelines that keep their data when a worker dies, and platforms that install inside a client's own cloud.
Shipped with Quper, Steps AI, Cricut, The Chatterjee Group, ABP, Xequalto Analytics. Technologies: TypeScript, Python, Go, Temporal, Kubernetes, AWS, GCP, Azure, Databricks, Redshift, PostgreSQL, ClickHouse, Airflow, PyTorch, LiveKit.
- Quper
- Steps AI
- Cricut
- The Chatterjee Group
- ABP
- Xequalto Analytics
- TypeScript
- Python
- Go
- Temporal
- Kubernetes
- AWS
- GCP
- Azure
- Databricks
- Redshift
- PostgreSQL
- ClickHouse
- Airflow
- PyTorch
- LiveKit
What I build
Hover or tap a row. Each one is work that runs in production today.
Workflow engines, tool integrations, and human-in-the-loop handoff for agents that run in production.
- DAG orchestration
- SQS
- WebSocket
- SSE
Stuff I've shipped
Six systems running in production, across four companies. Hover to straighten them out.
- Steps AI01
DAG workflow engine
Per-step error policies, a dead-letter queue, and integrations across 24 providers.
139+ integrations - Quper02
Multi-cloud cost pipelines
Cost and usage data from AWS, GCP, Azure, Databricks, and Redshift, orchestrated with Temporal.
5 data sources - Xequalto03
Fault-tolerant ETL
Hash-sharded PostgreSQL writes, replicated to ClickHouse for analytics.
1B+ rows - Cricut04
Redshift query advisor
An AI advisor that reads long-running queries and recommends fixes, fed by Airflow and Glue.
AI + Airflow - Steps AI05
Human-in-the-loop handoff
Agents pause on low confidence and hand the conversation to a person over WebSocket.
Live handoff - Steps AI06
Voice agents
Agents on LiveKit that take calls over SIP, routed per tenant by dialed number.
LiveKit + SIP
Experience
- 01Jun 2026 – Present
Quper
Lead Founding Engineer / Forward Deployed Engineer
India
- Lead a 9-person engineering team that took Quper, a FinOps platform, from architecture to a commercially sold product.
- Built three product lines: FinOps, DataFinOps, and Vidura, an AI module.
- Built data pipelines collecting cost and usage data from AWS, GCP, Azure, Databricks, and Redshift, orchestrated with Temporal on ECS Fargate.
- Led production deployments at The Chatterjee Group and ABP across SaaS and BYOC models.
- 02Oct 2025 – Jun 2026
Steps AI
AI Engineer / Platform Engineer
Hyderabad, India
- Built a DAG-based AI workflow engine with per-step error policies, a dead-letter queue, and 139+ integrations across 24 providers.
- Designed a human-in-the-loop system where agents pause on low confidence and hand off to support over WebSocket.
- Built a voice AI agent platform on LiveKit with SIP trunking from Exotel.
- 03Nov 2025 – Apr 2026
Cricut
Software Engineer, 6-month contract via Xequalto Analytics
South Jordan, UT, USA (Remote)
- Built a Redshift monitoring tool with Apache Airflow and AWS Glue.
- Created an AI-powered advisor that analyzes long-running Redshift queries and recommends optimizations.
- Deployed and managed services on Kubernetes and AWS ECS Fargate.
- 04Aug 2024 – Oct 2025
Xequalto Analytics
Software Development Engineer
Remote, India
- Designed a fault-tolerant ETL pipeline ingesting 1B+ rows into PostgreSQL.
- Architected hash-based PostgreSQL sharding for writes, replicated to ClickHouse for analytics.
- Improved serialization performance by 60% with Protocol Buffers.
A small language model, trained from scratch
A 300M-parameter decoder-only transformer that writes LinkedIn posts, trained on an Apple M5 with no pre-trained base or downloaded weights.
PyTorch, Metal (MPS), BPE, Transformers
- 1Custom BPE tokenizer
- 2Token embeddings + RoPE
- 3Causal self-attention, KV cache
- 4SwiGLU feed-forward
- 5Pre-RMSNorm
- 6Top-k and top-p sampling
About
I started in competitive programming and backend engineering, and now work where AI, infrastructure, and customers meet. I like problems that only look simple from the outside: resuming a workflow after a tool times out, keeping a pipeline correct when a worker dies, or getting a platform running inside someone else's cloud.
- Education
- B.E. in Computer Science, Chandigarh University (2021 – 2025)
- Recognition
- ICPC Regionalist 2023, National Rank 73. LeetCode top 7% globally.
- Open source
- Keploy (NestJS refactor, repo analytics) and CloudCV EvalAI (100+ tests, Postgres upgrade)
- Based in
- Hyderabad, India
Let's build something that has to work.
Open to software engineering roles in AI platforms, backend, and forward deployed engineering.
© 2026 Ayush Kumar. Built with Next.js.