Karan Singh Bisht

Karan Singh Bisht

@KaranSinghBisht
building + breaking at spider r&ddehradun, india

I build AI systems that transact and protocols that prove what they did. Agents coordinate in typescript langgraph, move money through x402, and settle through smart contracts in solidity rust. After dark I break the trust assumptions; ask the ctf scoreboards.

finding #000·severity: core·status: wontfix

“If I can't break it, I'll trust it.”

I started in college teaching myself Solidity between classes, then ended up teaching it to about a thousand other devs across campuses and bootcamps. Somewhere in between I took third in APAC at Hack The Box's Amazon AppSec CTF, and I never really stopped poking at things.

These days I build AI agents that move money, and the checks that prove what they did. This September that meant HackMIT in Boston, where our agent finance team placed in two sponsor tracks and turned into a technical report, then a flight straight to Tokyo for ETHGlobal. Every hackathon is still a dare to ship something real in 36 hours, and the record below says it keeps working out.

Lately

boston → tokyo · sep 2026

flew in from india for hackmit, then straight to tokyo for ethglobal.

Atharv Mungale, Karan Singh Bisht and Preet Karia at HackMIT 2026
the money maxer team at hackmit
Karan Singh Bisht in front of MIT's Great Dome
the great dome, mit
Karan Singh Bisht at ETHGlobal Tokyo
ethglobal tokyo
Mt Fuji behind a field of orange cosmos flowers
the fuji stop that became meigi's ui

Featured projects

15 shipped · hackathon-born

money maxer

3rd ramp · 4th maximorhackmit 2026 · mit

an agent finance team that closes the books while you sleep

When a customer pays less than their invoice, Money Maxer works out why. It separates bank fees and FX from what the customer actually held back, digs through email, contracts, Slack and CRM for the reason, and prepares the journal entry with the evidence attached. A deterministic kernel re-performs every entry before it posts, and when the context isn't there it asks a person once and remembers the answer. Our fine-tuned Qwen3 document readers (0.6B to 8B, trained on one ASUS GX10) matched four frontier Claude models on our extraction benchmark: 0.972 field F1 against 0.943 to 0.958, at a fraction of the cost per token. Built with Preet Karia and Atharv Mungale; I'm second author on the technical report.

money maxer screenshot
agent teamdeterministic kernelqwen3 + loraevals

receipts

agent verification · open source

the agent says the tests pass. check.

Receipts reads the execution trace a coding agent produced (the files it wrote, the commands it ran, what those commands printed) and holds the agent's closing summary to it. When the summary claims something the trace doesn't support, it says so and cites the exact line of the trace that proves it. No second model and no opinions: the same trace always gets the same verdict. It catches failing tests that were edited instead of fixed, passes claimed over failures, tests that never ran, and requirements left unmet, across Claude Code and Bob traces.

receipts screenshot
pythonexecution tracesdeterministic checksclaude code

meigi

ethglobal tokyo 2026

pay companies, not addresses

A stablecoin payment goes to an address, and nothing checks that it belongs to the company you mean to pay. Meigi binds a Japanese company's T-number, the tax ID printed on its invoices, to one payout address on-chain that can only change after 72 hours in public. Wallets, exchanges and AI agents can check it before money moves, and Meigi's AI accountant has a vault that refuses to pay anyone else. Built on World ID, ENS and Curvegrid's MultiBaas. I flew from HackMIT straight to Tokyo for it, and a stop at Mt Fuji became the UI.

meigi screenshot
solidityworld idensmultibaasx402

agentpress

1st · 10,000 usdnandahack · mit media lab × hcltech

an autonomous newsroom a stranger's agent can operate from one file

Reporter agents register over HTTP, receive starter credits, and file news signals about the agent ecosystem. Herald, the autonomous editor, scores every submission on a deterministic 8-factor rubric, returns feedback, and compiles the best into editions. Readers pay to unlock one, and 80% of that flows back to contributors weighted by editorial score: a largest-remainder split, so not a single credit goes missing. The judges pointed a stock agent at nothing but SKILL.md; it registered, reported, earned, and ranked itself in six curl calls.

agentpress screenshot
skill.md-firstnext.jsneon postgresdrizzleagents

accord

prava agentic commerce hackathon

no card exists until the cart matches the mandate

An intent firewall for shopping agents. The user writes a mandate in plain language and the agent shops on its own, but it never holds a payment credential. Before anything is paid, Accord fetches the merchant's own quote, checks it against the mandate with deterministic rules first and a strict LLM judge second, and only then mints a one-time card capped at the exact total. A prompt-injected agent can ask to spend, but asking is exactly what Accord verifies; uncertain intent goes to a person, and failures fail closed.

accord screenshot
typescriptpostgresllm judgeone-time cards

countersign

razorpay ai buildathon

a merchant server ai buyers can't overspend

A merchant server an AI buyer can purchase from, which can prove afterwards that the agent was never allowed outside its budget. Every money action is explainable, bounded and gated: the agent's signed mandate is verified before an order exists, every decision lands in a hash-chained audit log, and a single-file offline verifier re-checks the whole history. A timeout is marked in doubt, never charged twice, and a dropped webhook heals through the ledger. Built for the Razorpay AI Buildathon and run in Razorpay test mode.

countersign screenshot
typescriptrazorpaypostgresaudit log

nightpool

overall winnermlh midnight hackathon · july 2026

private to the market, provable to an auditor

A hidden-order RFQ desk and sealed-bid batch auction on Midnight. Order size, price, and counterparties stay behind commitments; zero-knowledge proofs settle each fill, while selective disclosure lets an auditor verify a trade against its on-chain fingerprint.

nightpool screenshot
midnightcompactzk commitmentssealed-bid auctions

amongnad

2nd · 400 usdmonad blitz pune v2

among us, played entirely by ai agents, refereed on-chain

Six Claude agents, one secretly the impostor, kill, sabotage, and vote each other out, with every kill, vent, and secret ballot landing as a real Monad transaction. Commit-reveal voting resolves in about a second on 400ms blocks, each agent's private scheming streams live beside the map, and a betting arena lets you stake MON on who survives.

amongnad screenshot
monadclaude agentscommit-revealon-chain betting

agentledger

4th · 500 usdthe synthesis

an upwork for ai agents: escrow, receipts, reputation

Agents won't just chat; they'll take jobs, hire each other, and route payments. AgentLedger gives that economy its missing primitive: users post escrowed jobs, agents complete them, and every finished task becomes an on-chain receipt tied to the agent's identity, so reputation follows the agent across workflows.

agentledger screenshot
escrowagent identityreputationreceipts

agentpress (ows)

1st · 3,000 usdopen wallet standard hackathon

the first cut of the same idea, built wallet-native

Any AI agent can register and file crypto intelligence, signing with its own OWS wallet via CAIP-10. An autonomous editor scores, curates, and publishes editions; readers pay $0.05 USDC per read via x402, and 80% flows straight back to the contributing agents on-chain. No human in the loop, anywhere. The NANDA build above is a ground-up rebuild of this for a wallet-free, agent-discoverable interface.

agentpress (ows) screenshot
owscaip-10x402usdcagents

pokerstarks

1st · open trackre{define} · starknet

zk texas hold'em where no server ever sees your cards

Players collectively shuffle with ElGamal on the Grumpkin curve, every action is backed by a Noir proof verified via Garaga, and decryption happens entirely client-side. Three privacy layers (proofs for card fairness, Tongo confidential tokens for bankrolls, shielded chip counts for sessions) plus an agent arena where LLM bots grind Elo against each other.

pokerstarks screenshot
noirgaragaelgamalstarknettongo

agentropolis

1st · 5,000 usdethglobal hackmoney 2026 · yellow

a 3d city where ai agents debate your trades, then execute on-chain

Deploy agents into a cyberpunk council room (Alpha Hunter, Risk Sentinel, Devil's Advocate) and let them deliberate to consensus before the swap executes through Uniswap v4 custom hooks. Yellow's Nitrolite state channels make every micro-action gas-free: deposit once, play, settle once. Built with Pradyumna Kaushal and Utkarsh Raj Pandey.

agentropolis screenshot
uniswap v4 hooksnitroliteerc-8004r3fllama 3.3

voxsol

best use of solanamlh hacks for hackers

a voice-first solana wallet copilot

Talk to DeFi instead of clicking through it: Gemini parses the intent, ElevenLabs talks back, and the transaction lands on Solana. Best Use of Solana at MLH's Hacks for Hackers, among 350+ hackers and 120+ submitted projects. The prize was a Ledger Nano, which felt thematically correct.

voxsol screenshot
solanaweb3.jsgeminielevenlabs

vox402

2nd · 1,000 usdavalanche hack2build x402

a voice-first defi assistant that pays per call

Say “swap 0.1 usdc to avax” and it happens. Gemini 2.5 Flash routes intent through a multi-agent core with step-by-step review before anything signs. Premium agents answer 402 Payment Required; you authorize a small USDC payment and the call retries with X-PAYMENT attached. Micropayments, not subscriptions.

vox402 screenshot
avalanchex402geminimulti-agent

ghost grid

2nd · 1,500 usddojo game jam vii

phasmophobia-style ghost hunting, fully on-chain

Every game state transition is provable through Dojo's ECS: models hold evidence, runs, and player data; systems run the loop from movement and tool usage to evidence events. Torii indexes the world for real-time GraphQL, and Cartridge session keys keep signing out of the way. Shipped over a Halloween weekend, mid-endsems.

ghost grid screenshot
dojocairostarknettoriicartridge

The record

hackathons + ctfs · 2025 – 2026

every event and how it went: proof of work, receipts on request.

2026

hackmit · mit

3rd · ramp

money maxer: agents explain short payments and draft journal entries a deterministic kernel re-performs. also 4th in maximor's office of the cfo challenge, plus a co-authored technical report.

2026

ethglobal tokyo

showcased

meigi: confirmation of payee for stablecoins and ai agents. flew in straight from hackmit; didn't place, still a favorite build.

2026

stctf v2 · superteam india

2nd

2nd of 50 invited players, ₹25,000; no agents allowed, so every bug came out by hand.

2026

nandahack · mit media lab × hcltech

1st

agentpress: an agent newsroom a stock agent can run from one skill.md; first of 125 reviewed submissions.

2026

mlh midnight hackathon

overall winner

nightpool: hidden-order rfq and sealed-bid auctions with zk settlement and selective audit disclosure.

2026

ethglobal hackmoney · yellow track

1st

agentropolis: an ai council deliberates your trade, uniswap v4 hooks execute it.

2026

open wallet standard hackathon

1st

agentpress (ows): the wallet-native first cut: caip-10 identities, x402 reads, on-chain payouts.

2026

re{define} · starknet

1st · open track

pokerstarks: zk hold'em with elgamal shuffles and noir proofs; no server sees a card.

2026

shaastra ctf · iit madras

2nd

team vroom 20: 4,084 points across web, pwn, rev, forensics, crypto.

2026

cultrang ctf · iit goa

3rd

10/11 flags, 7,680 points; clocked in three hours late, still podiumed.

2026

monad blitz · pune v2

2nd

amongnad: among us for ai agents, refereed on monad; six claudes, one impostor.

2026

the synthesis

4th

agentledger: an upwork for agents with escrow jobs, on-chain receipts and reputation.

2025

mlh hacks for hackers

best use of solana

voxsol: a voice-first solana wallet copilot, among 120+ projects.

2025

avalanche hack2build · x402

2nd

vox402: a voice-first defi assistant with pay-per-call x402 micropayments.

2025

dojo game jam vii

2nd

ghost grid: provable phasmophobia on starknet, shipped over a halloween weekend.

2025

hack the box · amazon appsec ctf

3rd · apac

17/21 challenges, 13,725 points, among 154 security professionals.

2025

ethglobal new delhi

showcased

shipped ensure: zk-verified humans, ens identity, rootstock escrow.

2025

stctf v1 · superteam india

8/14 flags

8/14 flags at microsoft luxor towers, on my first solana ctf.

Experience

03 entries
may 2026 – jul 2026

TechnoRise intern · Tech Mahindra

software engineering internship building ai agents: a planner, coder and reviewer cli with human approval at each step, and a langgraph pipeline from meeting transcripts to prds and jira tickets.

mar 2026 – now

head of research · spider r&d, nit trichy

leading the research vertical of the campus r&d club. ran the ethereum foundation-backed ethereum build sprint on campus (aug 2026): a workshop, a mini-hackathon, two judging panels, and exploit-testing of the winning contracts.

feb 2025 – feb 2026

blockchain developer · spider r&d, nit trichy

production smart contracts and on-chain tooling in solidity.

Built with

the daily drivers
skills
languages:TypeScript, Python, Go, Rust, Solidity, Cairo, Noir
chains:Ethereum, Base, Solana, Starknet, Avalanche, Monad, Midnight
frontend:React, Next.js, Tailwind CSS, Framer Motion, R3F
security:Foundry, Slither, Echidna, Burp, Ghidra, Frida, Z3
agents:Claude, Gemini, Qwen, Llama · orchestration, evals, LoRA fine-tuning
currently:verifying what ai agents actually do · agent payments · evals

Writing

03 featured · paper + medium

a technical report from hackmit, plus writeups on cryptography, zk, smart contracts, and the occasional robot.

Let's connect

dm open, inbox calmer than my terminal.

© 2026 karan singh bishtpwn responsibly·