Hi, I'm Dev Jhawar.
Retrieval that cites its sources, agents with real state machines, and queues that survive a bad day. Lately I've also been sending fixes upstream to Kubeflow, Prefect and Docling.
Who's behind this
I'm a final-year CS student at KIIT who likes the part of software nobody screenshots: queues, retries, retrieval pipelines and the agents that sit on top of them.
Most of what I build is an LLM system that has to behave in production: research agents that cite what they read, RAG that points at the exact clause it used, and schedulers that survive a Redis flush.
Lately I've been sending fixes upstream to the tools I use, like Kubeflow, Prefect, Docling and OpenLLMetry.
The five I'd show you first
Describe something on the web in plain words and get one alert when it happens.
Evaluation runs a cost ladder: hash gate, learned noise filter, rules, triage model, then the judge model only when needed
A spreadsheet where every cell is researched by its own agent, with sources.
Describe a consumer dispute and see how Indian commissions decided similar cases.
Ask questions of a policy PDF and get answers tied to the exact clause and page.
Reads a messy folder, proposes a clean structure, and moves nothing until you approve.
A year of commits, and pull requests to other repos
Each dot is a day. Bigger dots mean more commits, PRs and reviews.
12 repos. Checked GitHub at build time
Thanks @DEVJHAWAR11, this is the right fix. Bounding the send in the shared connection class means every service that connects through _websocket_connect() is covered without touching any of them. …
@markbackman on pipecat-ai/pipecatThanks for adding this. I reviewed ff4ae04 against #2894, including the provider, schema, template, docs, and tests. On Python 3.12.8, the 15 role-adherence tests and 14 existing conversation-template tests pass. …
@feiiiiii5 on truera/trulens@DEVJHAWAR11 Thank you for your contribution! If you like have a look and work on something else or help others with reviews!
@niccokunzmann on collective/icalendarLGTM - Thank you!
@davidmezzetti on neuml/txtaiWork and school
Work
Kubeflow · Prefect · Docling · OpenLLMetry · icalendar
Fixing things I run into in AI and data tooling, from data-cache connections in Kubeflow Trainer to 64-bit parameter overflows in Prefect's API.
Hacktoberfest
Contributed to web and tooling projects alongside maintainers in the Hacktoberfest community.
Education
KIIT University, Bhubaneswar
CGPA 9.65. Data structures, operating systems, networks, databases, AI.
NIOS, Siliguri
Scored 72%.
St. Joseph's School, Siliguri
Scored 94%.
What I reach for
State machines, tools and retries
Search that cites its sources
Models, from scratch and fine-tuned
APIs, queues and databases
Ship it, trace it, keep it up
Say hello
Bhubaneswar, India (UTC+5:30)
I'm happiest on the backend of AI products: retrieval, agent orchestration, queues and the reliability work around them.
