Parth TiwariStart a project
MenuClose

About / The person on the sheet

Hi, I am Parth.

I build AI products and the systems around them: retrieval, agents, generation, workflows, evaluation, and the interfaces people actually use.

I care about what happens after a demo works once. Where does the evidence come from? What happens when a worker dies? What is the model not allowed to decide? Those questions shape most of my work.

I like ideas that are a little too ambitious. I just want the system underneath them to stay honest.

Parth Tiwari standing in front of river rapids at golden hour, in a black tee, pale jeans, sunglasses and a cream cap, looking off to one side.
Parth Tiwari / Bengaluru

Path / Education and work

How I got here.

Study and work receive equal space here because both changed what I could build next.

  1. 01Jul 2021 – Jun 2025
    Education

    B.Tech in Computer Science

    IPS Academy · Indore, India

    Artificial Intelligence and Machine Learning

  2. 02Jul 2025 – Feb 2026
    Education

    Post Graduate Program in Data Science

    Great Learning · Bengaluru, India

    Generative AI

  3. 03Mar 2026 – Present
    Work

    AI/ML Intern

    Stick and Dot · Bengaluru, India

    Building generative media systems and the product workflows around them.

Current / Work in practice

Right now, I build at Stick and Dot.

Building generative media systems and the product workflows around them.

  • 01

    Built and evaluated Vivid's script-to-storyboard generation pipeline.

  • 02

    Developed the LoRA data and training workflow used for character and style experiments.

  • 03

    Worked across model evaluation, backend jobs, product surfaces, and failure recovery.

Operating rules / What survives the demo

A few rules I try not to break.

These are not values written for a wall. They are the checks I use when the interesting version and the responsible version pull in different directions.

  1. 01

    Evidence before the sentence.

    If I cannot point to the run, record, or source behind a claim, it does not belong in the public version.

  2. 02

    Keep the failure in the story.

    The interesting part is usually what broke, why it broke, and what the correction still does not solve.

  3. 03

    Give consequential decisions a human owner.

    Models can retrieve, generate, rank, and suggest. They should not quietly inherit authority they were never given.

  4. 04

    Make the useful version work first.

    The readable page, durable job, and honest fallback come before the cinematic layer built on top.

  5. 05

    Start with the smallest real test.

    I would rather learn from one bounded run than hide uncertainty inside a bigger build.