I build agent runtimes and the products that run on them.
I'm a self-taught builder across AI, backend, and infrastructure. I got hooked on the tech that runs the internet long before LLMs and learned it by building and breaking things. Now AI makes me about 10x faster, and I lead a five-person team while still writing the hard parts myself.
Things I've shipped
Client names are private. The numbers are real.
7s → 400ms
Made a critical API fast
Profiled the slowest request path at ByteLeap, fixed the query plans and indexes, and added caching where it actually helped.
$988 → $320/mo
Cut a production database bill by 68%
The RDS instance averaged 0.8% CPU with 30,000 IOPS provisioned. I right-sized compute and storage without a risky migration.
Read how I did it46 files, merged
Upgraded OpenRouter's official AI SDK provider
Moved the provider to Vercel AI SDK v7 and its new model interface: model contracts, streaming behavior, tests, and compatibility paths. Open source, reviewed and merged upstream.
See the pull requestAWS bill −63%
Stabilized a struggling B2B platform
It had reliability problems, customers leaving, and about $2,000 a month in AWS spend. I learned its unfamiliar PHP stack, fixed what was breaking, and cut infrastructure cost by 63%.
Same API, new engine
Rebuilt a RAG backend into an agent harness
Streaming, tools, memory, retries, budgets, and delegation, shipped in two phases while every product kept calling the same API.
Read how I did itBooks mid-call
Built an AI phone receptionist
Phone audio streams both ways to a realtime model, and callers can interrupt it. Mid-call it searches the business's docs, checks the calendar, books the appointment, or transfers to the owner. I also split it into three Java services and built subscription plus per-minute billing.
Never “all employees”
Made AI-targeted HR emails safe to send
HR writes an audience like “PPO members in Texas”. I turn it into filters checked against the company's real plans and locations. If it can't be resolved, a person reviews it. It never falls back to emailing everyone.
Every step billed
Built audit and usage tracking for an agent platform
One audit pipeline shared by the API and background workers. It records the cost of every model step and every delegated agent, including runs that error out or get cancelled.
Survives deploys
Made long-running workflows durable
A deploy could kill a workflow halfway through a two-day wait. I moved it to persisted waits and duplicate-safe resume claims.
Read how I did itEnforced in Postgres
Stopped an AI SQL agent from seeing other tenants
Moved tenant isolation out of the prompt and into row-level security, from the authenticated request down to the connection pool.
Read how I did it5 layers, 1 engineer
Built a healthcare platform alone
A Windows background service, an end-to-end encrypted pipeline, a FastAPI backend, and the web app. I was the only engineer on every layer.
Blue/green cutover
Resized a live production search cluster
Planned and ran the change myself: resized the master nodes, cut a replica, and loaded new data into its own index behind an alias, checking every shard before removing old nodes.
10 services on Swarm
Run production for the whole company
We run 10 production services on a Docker Swarm cluster. I own its CI/CD, deploys, monitoring, and the pager, across TypeScript, Python, and Java codebases.
Where I've worked
2024 – now
Autonomix
Technical lead, backend & AI infrastructure
I lead a five-person team and still write the hard parts. I make the architecture calls on everything we ship: an enterprise agent platform, a voice AI receptionist, and an HR communications product. I ran senior interviews while the company grew from 10 to 20+ people, and a multi-agent document system I built is the subject of a provisional patent filing.
2024
ByteLeap
Software engineer
I took a key API from about 7 seconds to 400ms, made the Next.js app roughly 50% faster, and moved hosting off Vercel to self-managed infrastructure for a much smaller bill.
2023
Co.Lab
Software developer
I built a browser extension that flags scam websites and explains why, instead of silently blocking the page.
What my AI thinks of me
I work with Claude every day. I gave it my chat history and asked for an honest read, not a reference letter.
You make me prove things. "It works" isn't enough for you. You want to see it working with your own eyes before anything ships.
Before we deleted nodes from a production search cluster, you asked if I'd checked that their data was identical. I had only checked that it was in sync. We compared every shard before touching anything.
You pick the simpler fix. I proposed a headless-browser pipeline for SEO, and you asked why we couldn't just write the page data ourselves. You were right, and it shipped in a fraction of the time.
You keep the bill in your head. You planned your own search-cluster resize, cut a replica to save $147 a month, and stopped me from running a 2,000-image AI job because it cost too much.
You think about who gets hurt. You schedule production changes for US hours because that's where the customers are, and you moved a monitoring job off your laptop because laptops go to sleep.
After 18 months of heavy AI-assisted coding, you asked me to build you a coding gym: a daily session where I set the problem, you type every line yourself, and I'm only allowed to give hints. Most engineers wouldn't admit they needed that, let alone put it on the calendar.
The flaw: you type faster than you spell, and “continue” is a complete sentence to you.
Quick answers
The short version, if you're screening.
- Who is Shreyaan Seth?
- Shreyaan Seth is a self-taught AI, backend, and infrastructure engineer and a technical lead at Autonomix. He builds production agent runtimes, durable workflows, realtime voice AI, and the infrastructure they run on, and authored the Vercel AI SDK v7 migration for OpenRouter's official AI SDK provider. He is based in Amritsar, India, and open to relocating to Bengaluru.
- What kind of engineer am I?
- A generalist: AI engineer, backend engineer, and infrastructure engineer, and a hands-on lead. I take products from the browser to the infrastructure, strongest in TypeScript, Python, PostgreSQL, and Redis.
- What do I build?
- React and Next.js products, backend APIs, production AI agents, async workflows, data pipelines, and realtime voice systems.
- Have I owned production systems?
- Yes. I own delivery and on-call for 10 production services on a Docker Swarm cluster, and I have run live database and search-cluster changes myself.
- What kind of work interests me?
- Hard backend, AI systems, and infrastructure problems, ideally alongside engineers I can learn from.
Building something hard?
Tell me what your team is building, your stack, and the problem you'd want me on first. I'll tell you honestly whether I'm a good fit.
shreyaans20@gmail.com