0x00POWERED

hari prasath

I build challenging applications and solve challenging problems.

I always give special attention to fundamentals, and am keen to know what a system is made of and how.

I am always eager to explore next-gen ideas and use them to take systems to the next level.

Fascinated by the idea that the entire computer world works on top of 0s and 1s.

mot“If you can't explain it simply, you don't understand it well enough.”

python

git

postgres

docker

aws

vllm

linux

redis

mongo

kubernetes

kafka

terraform

↓

4 COMPONENTS

46 6f 6c 6c 6f 77 20 74 68 65 20 62 75 73 20 74 6f 20 66 69 6e 64 20 6f 75 74 20 6d 79 20 77 6f 72 6b 73 20 61 6e 64 20 69 6e 74 65 72 65 73 74 73 2eFollow the bus to find out my works and interests.

U1RUNNING

video/

Continuously watching video feeds, and delivering the analytics in real time.

⌄ OPENCLOSE

Similar to text data, visual data like video holds the potential to produce a lot of analytical data in the modern world. But going through all of it, continuously watching it, and delivering analytics in real time is a monumental effort. So there was a big opportunity in video analytics: by leveraging ever-advancing visual models, technology that can process them at blazingly fast scale, and the right infrastructure to continuously watch these feeds, we could build a solution that delivers real-time alerts for everything from traffic to public safety and a lot more.

I worked on developing such a video intelligence platform, right from the data models to the real-time infrastructure that continuously ingests feeds from different sources, takes them through a processing pipeline, runs them through visual models, and delivers the analytics back to the user.

It demanded understanding how video streaming and communication actually happen in the real world through WebRTC, RTMP, RTSP, etc., and picking the right frameworks like LiveKit to act as the heart of ingestion and bring the layer to life.

In the platform, the scale is handled through queues and async processing to make sure that no frames that could potentially lead to an alert or incident are lost. And solving this was the interesting part.

U2RUNNING

tools/

Moving from just using LLMs to putting them into action.

⌄ OPENCLOSE

The world was moving from just using LLMs to putting them into action, and the things that made this possible were tools and the harness around them. So when we set out to build an agentic platform, the first component we set up was the tooling flow, without which it cannot be truly agentic. Once the tools were written, we needed a harness to bring the LLM calls, the tools, and the agent orchestration together and connect the dots, and for that purpose, we used the Agno framework.

So far, we had provided the agent platform with what it needed to complete the basic tasks. But when a user wants to perform more complex tasks, they would want to attach their own tools for the agents to access and utilize. To achieve this, I took advantage of FastMCP to host the user-provided tool code, and to manage the scale, I introduced FastAPI entrypoints through which dynamic tool instances are spawned to respond to MCP queries.

With this, the platform was no longer limited to just talking with LLMs — it could now make them reason, act, and use the right tools to see a task through end-to-end.

U3RUNNING

inference/

Running open-weight models in production.

⌄ OPENCLOSE

As AI became a more integral part of our lives, it went from basic chats to coding assistants, AI agents, and automating day-to-day tasks. Alongside the proprietary, closed-source models built by frontier labs, open-weight models also started becoming increasingly popular. Although these models performed well, running them ourselves was a different problem. Larger models can require dedicated infrastructure with multiple high-end GPUs, and the cost of keeping that infrastructure running can quickly add up.

So, I built an inference platform that enables users to host and use open-weight models. The platform is designed to plan and make better use of the underlying resources to achieve cost efficiency. But the solution doesn't just stop at hosting the models — to deliver quick results, I leveraged prefix caching, and to avoid interruptions as the load increases, I architected an auto-scaling setup.

But once we have multiple replicas of a model running, how do we decide where each request should go? Typical load balancing that works for web applications doesn't necessarily work for AI workloads. We need more intelligent load balancing — one that is cache-aware, load-aware, and length-aware. Once we know which model we want, the job is to route the request to the right replica. But when we have to decide which model should handle the task in the first place, that's where intelligent model routing comes in. As AI infrastructure becomes a bigger part of the budget, making better-informed decisions about which model to use becomes increasingly important.

And the landscape keeps changing, with a lot more to come.

U4ALWAYS ON

under-the-hood/

I don't like stopping at the surface.

⌄ OPENCLOSE

Whenever I come across something new, either at work or outside of work, I tend to jump right into it if it interests me. I don't like stopping at the surface — I go all the way, until I understand it well enough to work with it confidently. I won't blindly accept a concept unless I understand it, very similar to one saying that I won't drive a car without knowing how the engine works.

And when something really piques my curiosity, I like experimenting with different ways of approaching it rather than just following how it is generally done. Not all of it might be needed to do the job, but it changes how I do the job.

DATASHEET

I build and own production systems, working across backend, distributed systems, real-time video and AI/ML pipelines.

What I enjoy most about engineering is designing systems for real-world scale, use cases and problems. I’ve always had a keen interest in the fundamentals.

As a founding engineer, I’ve had great exposure to real-world systems, building them from requirements through production, and the vast amount of learning that came with it.

View Resume →