2026-08 · in build
Shipping
Recoup — governance layer for failed agentic payments
When an AI agent's payment fails, someone has to decide whether to retry, block, or escalate — and be able to prove why later. Recoup is that decision layer.
- Problem
- Agentic payment stacks can retry money movement autonomously. Retry logic without a policy gate is an unbounded-spend bug waiting to happen, and without an audit trail nobody can attribute liability after the fact.
- Constraints
- Solo build, 7 days, hackathon submission. No mocks on the core path — real Razorpay test-mode Orders, payment and capture.
- Architecture
- Python 3.11 + FastAPI + Pydantic. Deterministic policy gate in front of every money action. Append-only SQLite ledger with SHA-256 hash chaining. Exactly one LLM call, used for diagnosis only and never near execution.
- What I owned
- End to end — failure taxonomy, policy rules, ledger schema, hash-chain verification, dashboard, demo.
- Honest metrics
- The dashboard shows attempted, recovered, blocked and escalated separately, plus false-positive and false-negative rates. No 100% recovery claim — the point is that the blocked ones were blocked on purpose.
prev—
·sha256computing…
· unverified
2026-08 · live
Deployed on Polygon mainnet
ChainGuard — scam detection for crypto group chats
A Telegram bot that reads links and contract addresses posted in crypto groups and tells the group, in the thread, whether they're known bad.
- Problem
- Scam links and drainer contracts spread fastest in the seconds before anyone checks them. By the time a warning is posted manually, people have already clicked.
- Architecture
- Python (python-telegram-bot) for detection, Google Safe Browsing plus a local blacklist for links, a persistent JSON registry so a scam seen once is re-flagged forever, and a Solidity
ScamRegistry contract on Polygon mainnet — verified — written to via web3.py so reports are public and shared.
- What I owned
- Whole thing, solo. Started as a coursework requirement that only needed a two-page document; the working bot was the part I actually wanted to build.
- What went wrong
- I leaked a private key into the repo during the mainnet deploy. I caught it, moved funds off the wallet immediately, untracked the
.env, and now treat that wallet as permanently burned. Secrets handling on every project since starts from that lesson.
prev—
·sha256computing…
· unverified
2026 · live on Sepolia
Submitted — Zama Season 3
PayParity — confidential salary benchmarking on FHE
Compare your salary against the market without anyone — including the contract — ever seeing your number.
- Problem
- Salary benchmarking is useful precisely because it's comparative, and impossible to do safely because it's personal. Plaintext on-chain is a non-starter.
- Architecture
- Fully homomorphic encryption on Zama's stack — comparisons run on encrypted values, so aggregate answers come back without the inputs ever being decrypted. Deployed to Sepolia at
0x063D78A28F677B7D5Da59f8650F94EB739E57175.
- What I owned
- Solo build and submission to the Zama Developer Program, Season 3 Builder Track.
- Why it matters
- It's the clearest proof I can point to that I can work at the cryptography layer, not just call an SDK.
prev—
·sha256computing…
· unverified
2026 · live
Deployed on Railway
Raksha Sutra — digital safety for Indian elders
An AI assistant that helps older users in India recognise the scam before they act on it, in language they actually use.
- Problem
- Elders are the highest-value target for UPI and OTP fraud in India and the least served by security tooling, which is written for people who already know the vocabulary.
- Constraints
- Built for a Design Thinking course with Madhav Nauriyal — which meant the interface had to survive testing with actual non-technical users, not just work.
- What I owned
- Build and deployment; shipped live on Railway rather than left as a prototype.
prev—
·sha256computing…
· unverified