Project summary

TARA (https://www.taraprogram.org/) started cohort based AI-safety study. India has no structured, cohort-based AI safety training program until now. TARA's Delhi cohort (Round 2, 2026) is the first of its kind: a 14-week technical program using the ARENA curriculum (transformers, mechanistic interpretability, RL, evals, alignment), running September–December 2026.

I have applied & was accepted to this programme. I paid my commitment bond and attended the first online session. At present, I live in West Bengal, and there's no TARA cohort in my state. I need to attend Delhi every Saturday for 13 remaining sessions, ~$190 each in flights and local transport, for a total of ~$2,500. TARA has confirmed they have no travel budget for participants, so this cost falls entirely on me.

Losing this seat means losing one of the only structured entry points into AI safety research currently available to me in India.

What are this project's goals? How will you achieve them?

I bring 13 years of machine learning, deep learning/NLP experience, and 3 years of independent AI safety research to this cohort. I've previously completed the University of Chicago's AI+Science Summer School and NeuroMatch Academy. TARA is the first opportunity I've had to do this work inside a structured curriculum, under mentorship with amazing mentors and peers, rather than alone (https://www.taraprogram.org/curriculum)

How will this funding be used?

Following is the cost breakdown (per session in INR):

1) From home to Kolkata Airport (taxi): approx. Rs. 2000/-

2) Flight between Kolkata and Delhi (direct economy flight, round trip): approx. Rs 14000/-

3) From Delhi Airport to venue The Belier at Incuspaze HQ27, Gurugram (via taxi/metro): approx. Rs. 1000/-

4) Travel expenditure per session: approx. Rs. 17000/- (~$190)

Total for 11 sessions: approx. Rs. 187000/- (~$2090).

Total for 13 sessions: approx. Rs. 221000/- (~$2500).

13 remaining sessions: ~ Rs 221,000 (~$2,500)

What's your track record on similar projects?

I hold a master's in data science (Distinction) from Birla Institute of Technology and Science Pilani (9.08/10, top 5%). My industry background includes large-scale ML systems at Infosys, IBM Research, and Johnson & Johnson (including a SciBERT-based medical search engine development).

My current independent research is in the areas of mechanistic interpretability methods for transformer belief representations to reduce misalignment, safety-relevant internal states, and layer-wise geometry of hidden representations across model depth. Code and papers:

1) https://github.com/pps121/graft-belief-geometry
2) https://github.com/pps121/torsional-belief-vector-field/
3) https://anonymous.4open.science/r/spinal-64CC/README.md

More: https://pps121.github.io/
LinkedIn: https://www.linkedin.com/in/partha121/

What are the most likely causes and outcomes if this project fails?

This funds programme attendance, not research directly. If funded, I will complete TARA's structured curriculum, gaining mentorship and peer feedback I currently don't have (my research to date has been entirely self-directed). By December 2026, I expect to have completed a reviewed research project I can share publicly and to have started a reading group mentoring other India-based researchers entering AI safety.

India currently has almost no structured, cohort-based AI safety training. TARA's Delhi cohort is the first. If this project fails—i.e., if I can't fund the remaining sessions—the cost isn't just to me: it's one fewer trained researcher in a country with very few entry points into this field and one fewer person able to pass that training on to others. The TARA programme will follow the ARENA curriculum, which covers mechanistic interpretability, alignment, evaluations, and reinforcement learning. To learn along with fellow batchmates will boost and push you towards more impactful research.

This is field-building and individual upskilling, which will eventually improve the AI safety of the community. Post, I will start reading group and will mentor other India-based researchers in the field that will extend the program's impact to reduce catastrophic risk from AI. Field-building in an underserved region is exactly the kind of high-leverage, low-cost intervention that individual donors are well positioned to fund—a small amount here goes a long way, precisely because so little structured infrastructure exists yet.

What would you do without this grant?

I would have to drop out. I've paid the bond and attended the first session—TARA is expecting me to continue. But ~$190/week isn't sustainable for me on top of my existing salary and expenses, and TARA has confirmed there's no institutional travel budget available.

Without funding, I'll keep doing AI safety research independently, alone, as I have been doing without the mentorship, accountability, and collaborative environment that TARA can provide me