Project summary

I'm a confirmed participant in TARA's Delhi cohort (https://www.taraprogram.org/ : Round 2, 2026), a 14-week technical AI safety programme running September–December 2026, using the ARENA curriculum (transformers, mechanistic interpretability, RL, evals, alignment). I have already paid INR 3000/- as a commitment bond to demonstrate my seriousness about AI safety and my deep interest in attending TARA sessions. I also aim to pursue impactful research projects focused on reducing catastrophic risks. As $2,500.00 is very difficult for me to arrange, I respectfully request some funds (if full funding is not possible) so that I can be part of the Delhi Cohort'26. I already paid my commitment bond and attended the first online session to learn about the overall programme, expectations, logistics, and the projects that need to be completed towards the end.

I'm based in West Bengal, India. There is no TARA cohort in Kolkata, so attending requires round-trip flights to Delhi/Gurugram every Saturday, roughly Rs 17,000/- (~$190) per session for flights plus local transport. I'm requesting funding to cover the remaining travel costs for the 13 sessions so I can complete the programme I've already committed to. I would be glad to receive even some partial funding if full funding isn't possible.

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

I've been accepted into this TARA cohort round and have enrolled to continue working in the remaining sessions under structured mentorship and a project. TARA organizers will take every class's attendance that counts for the last 2 weeks will be dedicated for projects aiming to reduce catastrophic risk from AI.

I have 13 years of experience in ML and NLP and am an active independent researcher in AI safety. I have also completed prior structured programmes (AI+Science Summer School at the University of Chicago, NeuroMatch Academy) that indicate I follow through on commitments.

This is the first time TARA is having its cohort in Delhi, India; therefore, I applied this year and got accepted to attend this field-building programme. The programme runs on a fixed weekly schedule from 5th September through 12 December 2026. 

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).

What's your track record on similar projects?

I have approximately 13 years of experience in machine learning and natural language processing (NLP), with 3 years in AI safety. 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).

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

Some of my prior work can be found at my webpage: https://pps121.github.io/

LinkedIn: https://www.linkedin.com/in/partha121/

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

These funds will be allocated to attend to technical AI safety weekly sessions as part of the TARA programme, a structured technical AI safety training programme that follows the ARENA curriculum, which covers mechanistic interpretability, alignment, evaluations, and RL, rather than funding my research directly. After receiving this funding, I will complete a rigorous technical safety training pipeline, which will better equip me to conduct alignment/interpretability research and mentor others on AI safety projects in India and worldwide. In India, with very few structured AI safety training pathways and no existing cohort-based programme of this kind, TARA is the first of its kind, with an exciting curriculum focused on risk reduction.

Currently, I don't have a structured mentorship and peer feedback group. Since my research (interpretability methods applied to belief vector fields, value misalignment, safety vectors, layer-wise hidden state representations, etc.) has been self-directed, I aim to complete and submit reviewed research projects/publications by December 2026 for public sharing. Also, the ability to start a reading group and mentor other India-based researchers in the field will extend the programme's impact to reduce catastrophic risk from AI. This is field-building and individual upskilling towards AI, going well for humanity.

What would you do without this grant?

I have to drop out of this technical AI safety programme. As I have already paid the commitment bond and attended the first online session, the organizers and mentors will think I will continue the programme. But the weekly Kolkata–Delhi travel cost (approx $190/session) is not something I can sustain on an ongoing basis alongside my existing salary and expenses. TARA's organisers have confirmed that they do not have an institutional travel budget to support me, so there's no fallback financial option for me to seek alternative help to attend this programme.

I will continue my AI safety research independently, as I have been, but without the structured mentorship, accountability, and pair-programming environment TARA provides. It will create more gaps in my AI safety experience. I will also lose the bond amount that has already been paid.