I’ve deployed 20,000+ AI agents to production, and I run the infrastructure they live on.
AI agent infrastructure engineer in Surat, India. Open to remote.
At Ampere, I own everything behind each agent: setup, hosting, cloud backups, fleet-wide upgrades, and the LLM proxy they all talk through. That proxy now runs on its own as Proxen and cuts AI bills for teams.
Before that, I led a college IoT project end to end that brought in ₹2.5 million.
AI video pipelineA generation harness that keeps characters, locations and voices consistent from shot to shot, which single-prompt video models can’t do on their own. It plans the shots, carries references across scenes and stitches the result into one film.2026
Endel Digital crane monitoringProject lead, Aug–Dec 2024. Three-phase industrial IoT system built with my college: sensor data collection, intelligent filtering and aggregation, and a real-time cloud dashboard. I ran the system design and coordinated the college team with Endel Digital. Brought in ₹2.5 million for the college.
Hire AIVoice-based mock interview agent using Deepgram, with context-aware follow-ups and adaptive feedback and analytics on Supabase.
Ampere
Running 20,000+ AI agents without a server in sight
OpenClaw agents are powerful, but running one means renting a server, setting it up, securing it and keeping it alive. Ampere does all of that in about a minute. I built the platform and I run it: every agent’s setup, hosting, data, upgrades and AI traffic goes through systems I own.
The Ampere dashboard. One click gives you a running agent.
Setup in 60 seconds, at fleet scale
Each signup creates an isolated container for the agent with its own CPU, memory and disk, wires up networking through Caddy and Tailscale, and starts OpenClaw with the user’s channels and skills. The same pipeline has set up 20,000+ agents, so a new agent is a routine event, not a support ticket.
Hosting that stays up
Agents are meant to run 24/7, so hosting is built around that: isolated containers per agent, encrypted keys, private networking between hosts, and plans sized from 2 vCPU and 2 GB up to 16 vCPU and 32 GB for business setups.
User data in the cloud
An agent’s value is its memory, so losing data is the worst thing that can happen. Each agent’s workspace is backed up to the cloud, users can export everything in one click, and people already self-hosting can import their setup in under five minutes.
Upgrades without breaking anyone
OpenClaw ships fast. Ampere upgrades agents across the fleet automatically, keeping each user’s data, config and skills intact, so users get new features without touching anything.
The LLM proxy
Every agent’s model calls go through one proxy that I built. It picks the right model for each task, meters usage against credits ($1 = 1,000 credits, auto top-up), and lets users bring their own key at zero markup. This is where I learned how much of an AI bill is waste, which became Proxen.
How a request flows through Ampere.
What this means for you
I’ve already made the mistakes that come with running agents for real users at scale: setup, data, upgrades and cost. I can bring that straight to your team.
How I worked
I build with AI agents heavily, including my own orchestrator, SAMURAI. I own the design, the operations and the result.
Claude Code, Codex, Cursor, Cline, the OpenAI and Anthropic SDKs, LangChain
TL;DR
We built the LLM proxy for Ampere first, because thousands of agents calling models every day makes cost a daily problem. Once it worked there, we turned it into Proxen: one address for a company’s AI traffic that cuts the bill without changing the tools or the answers.
Where it started
On Ampere, every agent’s model calls already passed through our proxy. Watching that traffic showed the same waste again and again: whole conversation histories resent on every turn, the top model doing easy work, and the same context paid for twice. We fixed it there first.
How it cuts the bill
Fewer tokens: trim the history and context each request carries to what the model actually needs.
Cheaper tokens: send easy requests to cheaper models and keep the strongest model for hard work, even inside one agent session.
Never pay twice: our own cache answers repeated requests, and providers’ prompt caching is used properly.
Half price when it can wait: tests, reviews, docs and nightly jobs go through batch APIs at half the price.
Many models, one address: Claude, GPT, Gemini and open models behind one endpoint.
The Proxen console: spend by person and team, budgets, and savings from routing. Example data.
Control a CTO actually needs
Every person and every agent gets its own key with a monthly budget, an alert at 80% and a hard stop at 100%. The log records who used which model, how many tokens and what it cost, while prompt bodies are not stored by default. Company keys stay on the server, and the proxy runs in our cloud or yours.
The result
On most setups we aim for a 50 to 70% lower bill at the same quality. We measure it against each team’s own baseline and test quality on their real work before and after every change.
What this means for you
If your product or your team runs on LLMs, I know where the money goes and how to get it back without slowing anyone down.
Orion is an AI “CEO” agent for small businesses running on OpenClaw. Most agents can tell you what they did. Orion is built to tell you why, and to use that the next time a similar decision comes up.
The problem
A new agent is impressive for a day. Then it forgets why it made yesterday’s call, suggests an approach that already failed, and can’t explain itself. It also takes days of configuration before it is useful at all.
Decision 1: four memory layers instead of one store
Different facts change at different speeds, so they live in different places:
Knowledge graph: durable facts, updated nightly.
Daily notes: a timestamped log of today, consolidated each night.
Tacit knowledge: the owner’s preferences and hard rules.
Decision journal: what was decided, the options, the reasoning and the outcome.
The decision journal is the part that matters. When a similar situation comes up months later, the agent can look up what it did, why, and whether it worked.
An example entry from the docs
## DEC-2026-0323-002: Deploy checkout hotfix
Context: 5 failed checkout sessions overnight
Options:(1) hotfix now (2) wait for full test suite
Decision: 1, with a rollback plan
Reasoning: lost revenue outweighs test risk; rollback takes 30s
Owner:"ship it"Outcome: 0 failures in 24h after deploy
Decision 2: trust is tiered, by source and by action
External input is read-only, known contacts get limited actions, and only the owner has full authority. Actions are tiered too: some run automatically, some run and notify, and some wait for approval. Email bodies are always treated as data, never as instructions, which blocks the most common prompt-injection route.
Decision 3: spend only what the task needs
A one-file bug fix gets a single cheap model in a fresh-context loop. A feature that touches five or more files gets a small team: an architect, coders working in parallel and a reviewer. A daily spending cap, $20 by default, stops runaway costs.
Setup in ten minutes
Run the setup script, pick one of five business templates (SaaS, e-commerce, agency, creator or freelance), answer four questions, connect your tools, and get the first morning brief.
Honest limits
Orion is a configuration and prompt system, not a trained model, and it was built with heavy AI assistance under my direction. It is designed and tested on Claude.
B.Tech in IT from L.D. College of Engineering, Ahmedabad (2022–2026), CPI 8.43. Building at Nextbase since June 2025. 800+ algorithm problems solved, national rank 598 on TUF+, 99-day Codeforces streak, Smart India Hackathon selection at college and GTU level, JEE Main 98.4 percentile.
