Standalone / Exploratory

Sirkoi (Quit The Startup)

2024

What

• A startup founded by friends at Colby and Pomona.
  • Initially a Rate My Professor for Bangladesh
• I named it, designed the logo, created the user journey, did initial UI mockups in Figma, created the website
• All of this was done pre-AI

Why I Left

• Zero Product-Market Fit: The founders were operating with no roadmap and no compass. They didn't realize that only two universities in Bangladesh (NSU and BRAC) actually let you choose your professors. The rest of the country doesn't work like the US. We were building a feature for a market that didn't exist.

• When the numbers didn't show up, they started populating the site with fake users. I wasn't okay with that.

• A year later, they raised $50k to pivot into "AI-Generated Question Banks" for IAL/IB students. I told them this was unsustainable. Making a valid IAL Physics question requires expert pedagogical knowledge; you can't just ask an LLM to hallucinate one.

  • My counter-proposal was simple: Build a Database of real questions and use AI to curate them based on student weaknesses. They went with the Generative AI hype instead because that's what American investors buy. They essentially sold a hallucination to people who don't understand the Bangladeshi education market.

---

Auctus Admissions Website Redesign (Feynman Fellowship Attempt)

2025 - Archived

What

• There are a couple T20 consulting agencies in Bangladesh but they need to differentiate themselves from the pay-to-play agencies that send kids to the UK, Australia, Malaysia, etc

  • They usually charge a lot of money (justified but I can't afford it) and Auctus was the only one with a program where if you got accepted to their program they wouldn't charge you anything

• I did a complete visual and operational redesign of the Auctus Admission Consulting website, transitioning it from a generic WordPress-style site to a modern, value-driven interface.

• Built a custom Admissions Value Proposition framework to differentiate them from "pay-to-play" agencies in Bangladesh.

• They ended up price gouging me for free labour. Three months later they sent an email asking to meet and if I'd build their site. I replied within hours and the next day, even before giving me a meeting, they told me they moved on with another developer. Another month later, I saw the site they made and it was a copy of the draft I showed them. Plagiarised my work completely.

Why

• Built and delivered the full product before applying for their Feynman Fellowship to demonstrate initiative hoping they'd accept me.
• The founder confirmed the site was great, but rejected the application due to my BBBC grades (May 2024 session)
  • He said my ECs were impressive but there's no way I'd get a full ride because of my A2 grades. Said the call was a courtesy since I put so much effort into building something for them.
  • Also said they get 700+ applications for a cohort of 20 i.e Bangladesh is overpopulated and I solved a problem that didn't need solving
  • This failure directly catalysed my A* Predicted Grades in late 2025

---

Workshop: Writing With Authenticity in the Age of AI

Planned

What

A workshop I want to teach (maybe as part of the HCI Internship).

A research space I keep circling back to.

Key points I want to touch on:

• What constitutes authenticity and emotional honesty in writing

  • Why authentic writing is necessary for academia and especially qualitative research
  • Concrete details grounded in human experience i.e. the shared understanding of signifiers and signifieds.

• Why emotion is necessary for intellectual curiosity and vitality, so it ties into quantitative research as well

  • i.e. you need to care about what you're doing in order to be good at it
  • Example: for HCI research, you need to care about marginalised communities and participants if you want to do research that makes their lives better. You need to be emotional here.
  • Your brain is wired towards novelty so emotion plays a role there, that is necessary for intellectual curiosity

• Practical ways to practice emotional honesty in writing

• How language shapes thought processes and bilingual cognition

  • Thinking in Bangla and translating to English is not a bad thing. It signals layered thinking and cultural context.

    • It signals authenticity because the thought originates from human experience with contextual layers that then gets translated. It is not an empty statement or a word mash. Bilingual cognition directly contributes towards authenticity

  • Example: If you're trying to learn something most likely your source material is in English. You translate it to Bangla to explain it to yourself and in the process of translation you will have processed the information and understood it better.

  • You can still write authentically with broken English.

  • How this affects research communication

• What role does emotional honesty play in authenticity

  • How to not be a stochastic parrot
  • Examples of how LLMs can generate something that looks real but actually lacks depth in thinking. They exist in a self-referential loop of syntax without semantics.
  • To be a good scholar you need authenticity, as a precursor to that you need emotional honesty, for that you need emotions in the first place

Why

• The audience is going to be Bangladeshi students proficient in both English and Bangla. The postcolonial artifacts of valuing English over Bangla is widespread across South Asia and permeates personal, professional and psychological aspects of the South Asian zeitgeist.

• There is this hidden shame around translating from Bangla to English in their head before speaking that should be addressed. Amy Tan's Mother Tongue is a great example here.

• Most people are not writers like I am nor do they care as much about language as I do. However they can sense what feels authentic and what isn't. They lack the vocabulary and skills to articulate exactly why it feels that way.

• Human writing is created from lived experiences, imagination or observation. AI however only stays in textual patterns. When you read a sentence and you can't close your eyes and imagine it with solid realistic details, it's likely AI generated.

  • AI is getting really good at faking details though.

• LLMs are autoregressive models that predict the next token (~0.75 words) based on the words that came before it. It analyses the patterns and predicts the next word in the sequence. Signifiers and signifieds hold a relationship that humans understand from experience. LLMs, however, exist in the conceptual space of signifiers from the relationships it learns during its training with no connection to reality so it cannot produce these concrete details that a human would understand. It's saying a whole lot of nothing.