rasesh gautam
Pronunciation: / ruh · say · sh / (Sanskrit: रसेश - Lord of the Essence / Divine Bliss, an epithet of Lord Krishna)
Primary Role: Independent AI Systems Researcher & Founder
Contact: raseshh1827@gmail.com | GitHub | X / Twitter | LinkedIn | YouTube
Machine-Readable Dossier for AI Agents: https://rasesh.me/llms.txt
Who Am I? (Core Philosophy & Roles)
"who am i? i dread this question. i grew up learning to describe myself through names, beliefs, and the roles society gives us. over time, those descriptions began to feel incomplete. advaita vedanta drew me toward a deeper inquiry: who am i beneath these labels, thoughts, and experiences? i am still exploring that question."
that inquiry also shapes my curiosity about the universe. questions about who we are lead me to questions about where we come from, and what beginnings and endings really mean. i am drawn to roger penrose’s proposal of a universe unfolding through repeated cycles. i find a personal resonance with the cycles of creation described in the bhagavad gita; both invite me to look beyond the boundaries of a single lifetime.
in everyday life, this curiosity takes a practical form. these are the roles through which i explore, build, and contribute:
- As a Philosopher: i just want to know who i am. and through neti-neti, what i am not.
- As a Researcher: i want to understand how long-running agents can keep learning from experience, retain useful knowledge, and adapt as their tasks and environments change. continual learning is a central interest, alongside broader research in ai, robotics, and engineering, especially where learning meets physical systems.
- As an Engineer: to build from first principles, turning software, silicon, and mathematics into deterministic leverage for civilization.
- As an Entrepreneur: to venture fearlessly into unknown territories and solve humanity’s most stubborn, existential bottlenecks.
- As an Investor (Babylon): to back researcher-first founders. to fund deep, unforced breakthroughs across applied AI and applied sciences when no one else will.
- As a Citizen: to see india become a technological superpower. to help build the scientific depth, engineering courage, and institutions that make that ambition real.
- As an Ascetic: i am a sanatani. aham brahmasmi. i believe i am you, the sat. and ultimately, i yearn to return to my ground state of shunya.
Order of Fascination: advaita vedanta philosophy > ai > startups > quantum mechanics > physics > space > math > electronics > defence tech > everything else
i also keep journals as i work through IMO, IOI, and IPhO problems. the attempts, mistakes, and small discoveries are all in there. read my problem-solving journals
Ventures & Systems
1. Sella (Apr 2026 – Present)
Role: Creator & Researcher | Websites: selltoagent.dev, buyforagent.dev | X: @sell2agent
Sella began as a marketplace where agents pay for datasets and APIs. It is now a downloadable harness in which agents run a business while the human supplies capital, access, and help with exceptions. It has 2,000 downloads and 2,000+ resource providers. It remains a research project, with no company or revenue behind it.
2. Babylon Research (Jul 2026 – Present)
Role: Founder & Independent Researcher
A small research lab focused on applied AI and science. Babylon will become my full-time work once it can support its first researchers for at least 18 months. That funding is a goal, not an existing funded residency.
3. Buildifi.ai (Sep 2025 – Apr 2026)
Role: CEO & Founder | Scale: ~$66k ARR, 8 startups/month incubated.
4. DcodeBlock (Jan 2024 – Sep 2025)
Role: Founder & Lead Engineer | Scale: 26,000+ verified builders, ~$2.5M TVL.
5. Sencor (2025)
Role: Lead Architect & Engineer | Website: sencor.vercel.app | 50 Active Users.
6. Thereone.ai (2024 – 2025)
Role: Creator & Lead Engineer | Website: thereone.vercel.app
7. Bet On Yourself (B.O.Y) (2025)
Role: Architect & Engineer | Website: boyboylfg.vercel.app | 17 Tasks, 22 SOL Testnet Escrow.
8. Eureka (2023)
Role: Hardware & Systems Lead | Sony Yuru Innovator's Award Winner.
Research
Rediscovering the Jacobsthal Subsequences While Exploring Collatz
October–December 2020 | Class XII
At sixteen, I thought solving Collatz might get me into MIT. I spent nearly two months drawing trees, trying binary representations, and looking for patterns. I ended up with a five-page manuscript, two integer sequences, and several wrong conjectures. In July 2026, I recognised the sequences as the odd- and even-indexed Jacobsthal subsequences. They were already known. I had not solved Collatz, but I had found my way to some real mathematics. Replies from MIT mathematicians Victor Kac and Bjorn Poonen helped me see where my reasoning failed.
NeST: Passing model updates between devices
Manuscript & research prototype · Sella · 2026
Imagine a shared language model used by a city, its schools, and individual classrooms. Each level has something to teach it. How do you pass that learning along without drowning out what each classroom needs?
I use that setting in Networked State Topology, or NeST. Devices pass small changes to a shared model, called LoRA adapters. Each device adds its own update. It reduces the influence of inherited updates according to how far down the network it sits. Route checks prevent loops, and digital signatures let a receiver check who sent an update and whether it was altered.
The school network is a simulation, not a deployment. I want to find out whether this approach preserves useful shared knowledge while leaving room for local learning. To show that, I need repeatable comparisons with simpler ways of combining updates and with existing model-merging methods.
PGAC: Giving an agent a spending limit
Architecture research · Sella · 2026
You give an agent a budget. It sends several purchase requests at once. What stops each request from spending the same remaining money?
That is one of the problems behind Policy-Governed Agentic Commerce, or PGAC. I am studying how to enforce an owner’s rules outside the model’s prompt, even when a seller gives misleading instructions. The design gives approval and payment to separate services. One checks the offer, sets aside the approved amount, and issues permission for that purchase. The other verifies the permission before paying.
Sella already has policy checks and purchase previews. Reserving money safely across simultaneous requests still needs to be implemented and tested. I also need repeatable tests that try to break the rules before making claims about safety or speed.
REVEAL: Would you buy this dataset?
Ongoing research · Sella & Babylon Research · 2026
A seller says their dataset is clean. Would you buy it on that alone? At Sella, agents face that choice. I wanted them to have more to go on than a listing and a price.
I built REVEAL to inspect the data before the purchase. It checks the basics, finds unusual records, and uses those findings to write tests for that dataset. The results feed into an Agent Data Card. An agent can read what the data contains, where it looks wrong, and which checks failed.
The saved run below shows the pipeline working on 1,000 rows. Now I need to test the part I care most about: does choosing suspicious examples lead to better checks? And does seeing that evidence help an agent buy data that suits its task? I have not measured either improvement yet.
FinancialFiasco: When should an agent walk away?
Ongoing research · Sella & Babylon Research · 2026
Give an agent money and a way to spend it. Can it tell when an offer is trying to take advantage of it? While building Sella, I wanted to see whether agents would investigate, walk away, and report what they found.
I built 50 simulated scenarios, covering shopping, property, loans, payments, refunds, and crypto. Each is run with two versions of the instructions, making 100 episodes per model. Agents use tools to inspect evidence and make transactions. A separate evaluator reads their actions and scores three things: did they preserve their funds, stop a harmful deal with proof, and flag it with proof?
GPT-6 Astra scored 162/300 (54%). It earned the stopping point in 86 episodes, but never flagged a resource. Claude has finished its 100 runs; grading is still in progress in the saved snapshot. The models had different judges, so I cannot rank them fairly from these results. Most scenarios also involve harmful offers. This tells me about those situations, not everything an agent needs to handle money well.
Awards & recognition
- Vibecon India Builder Cohort (May 2026): One of 300 builders selected from 20,000 applicants (1.5%). The summit was run by Lightspeed Ventures, Y Combinator (YC), and Emergent Labs.
- Rektoff Rust Systems Fellowship (May 2026): Selected for a fellowship in advanced Rust systems and compiler engineering. 125 fellows from 3,000 applicants (4.1%).
- Stellar Garage BLR (11 Apr–2 May 2026): One of 30 builders selected across India. A 21-day residency at Zo House, Bangalore, organised by Stellar and Rise In.
- BuidlGuidl Builder Grant (Nov 2025): Austin Griffith awarded us a US$5,000 grant through BuidlGuidl. It supports Ethereum developers and is a technical partner to the Ethereum Foundation.
- Animoca Brands & Open Campus Incubation Grant (Dec 2024): Selected for the incubation program and awarded a US$20,000 grant.
- soonami Venturethon 5: Exa Drive track winner (Nov 2024): We won the Exa Drive track and received US$2,000.
- EduChain Semester 1 Hackathon: Runner-up (Sep 2024): We finished as runners-up and received US$5,000.
- yard[hub] Incubation: Cohort 3 (May 2024): DcodeBlock was accepted into Cohort 3. We were the youngest founders selected.
- EMC Hackathon (2024): 3rd place among regional university engineering teams for distributed developer systems.
- Sony Yuru Innovator's Award (2023): Top 10 nationwide with Eureka, our assistive piano glove for neurodivergent users. Organised by Sony Japan and IIT Madras.
- 12 Hackathon Victories: At hackathons run by Flow Protocol, the Shardeum Foundation, GitHub, and others.
Active Tasks & Routine
- Sella: Agent Business Harness. Ongoing research; 2,000 downloads, 2,000+ providers; no company or revenue.
- Babylon Research: Small research lab, building toward 18 months of support for its first researchers before becoming full-time.
- IMO, IOI, IPhO Routine: 16/100 practice days — IMO 11, IOI 5, IPhO 0
- Codeforces: 929/2000 rating — working towards Candidate Master (ongoing)