BS, Computer Science
One of two undergraduate degrees read in parallel. Coursework spans AWS tooling and artificial intelligence alongside core computer science.
I build the thing — then I build the thing that catches it lying.
I build software that gets trusted to decide something on its own — clear a payment, flag a scam, grade another program's work.
Systems like that fail quietly. Nobody gets an error; the fraud just goes through, or a test reports a pass for something that never ran. So most of my work is the checking layer — the part that has to be convinced before anything actually happens. Two of those checks turned up real bugs in Google's payment software, now filed publicly on their repository.
The subject matter moves around a lot: payments, elections, compliance, child safety, mental health, voice, travel. The question underneath never changes — how do you know this thing is right?
Turning rules into checks a system cannot skip, and evidence an auditor can replay.
PRAMANAAP2 #339 / #340ComplyLiteTask graphs planned at runtime, and graders built so an agent cannot score by doing nothing.
Eval harnessopen-multi-agentThis consoleRetrieval, classification and signal work — with the confidence kept attached to the claim.
Raksha AIVoice emotion CNNWellness companionSettlement on-chain where verification should not depend on whoever ran the process.
VortiFiVeriServePOCSO ShieldSchemas as contracts, migrations owned by the repo, tests against a real database.
GlobeTrotterComplyLiteTradexAndroid and web clients, motion systems, design tokens gated on contrast in CI.
Sole VaultRaksha AIGlobeTrotterYou do not need to know anything technical to use it. Here is the whole idea, in three steps.
Five of them, listed down the left of the console. Each one is a small job — work out who this is, go through the projects, check the education.
Six assistants share the work between them. You watch them think, look things up, and report back — live, one line at a time. Nothing is pre-recorded.
Everything they report is on this page anyway. The console is just a more interesting way in. Scroll past it and you lose nothing.
Six operatives, one discipline each — they are the six areas the work actually falls into. Click one to recruit it: you earn XP, and its full briefing opens. All six are available from the start.
Pick a mission and dispatch it. The plan is a directed graph, not a script — independent steps run concurrently, interleave their traces, and finish out of order. Tool calls are real; github.repo queries the public API live.
Establish who this dossier belongs to and what he actually claims to do.
10 builds — the ones worth defending, drawn from the public repositories plus two private ones. Ranked by how much of the work survives questioning rather than by stars, and rarity follows that ranking. Turn a card over for the argument and the numbers.
Grading against the evidence snapshot…
No model on the money path.
A safety gate that sits between an AI shopping assistant and your bank, and refuses the payment unless every required check actually ran.
A deterministic verification gate between an AI buying agent and merchant checkout. It parses an AP2 delegation chain, computes the RBI e-mandate envelope as executable predicates, and writes a hash-chained evidence record for every verdict it reaches.
A declared obligation that produced no result becomes INDETERMINATE at construction time and blocks exactly as hard as an outright violation — the difference between "checked and passed" and "never checked" is enforced by the type system, not remembered by convention.
Two defects found in Google's agent-payments code.
While building the gate above, two real bugs turned up in Google's own payment software. Google confirmed them, and both are now public on their repository.
Building PRAMANA against Google's AP2 reference SDK surfaced two defects in its own verification path. Google confirmed the mechanism, classified it as intended behaviour, and invited a public issue — both are now filed upstream as an issue and a pull request.
The finding ships as an executable proof rather than a write-up: `pramana finding` mints two presentations against the installed SDK, runs AP2's own evaluators over each, prints the commit it ran against, and exits non-zero the day the defect stops reproducing.
Rules first on device, retrieval second, model last.
An app that spots scam messages and calls before they reach you, in twelve Indian languages — and always tells you which specific warning sign it saw.
A fraud-interception system across Android and web. Detection runs in a fixed order — deterministic rules on the device, then retrieval over past cases, and only then a model — so a verdict is explained from evidence rather than invented. Case history is isolated per user with row-level security.
The ordering is the design. Putting the model last means the common cases never reach it, the explanation is always traceable to a rule or a retrieved precedent, and the system degrades to something still useful when the network is gone.
Compliance alerts that can show their working.
Software for stockbrokers that watches trades for suspicious patterns, and shows exactly which rule fired so a human can argue with it.
A hybrid rules-plus-ML detection engine over CSV and SFTP ingestion, with DuckDB for in-process analytics and a React dashboard for triage. Backend, frontend, sample data and a Docker Compose stack all live in the repository, with detectors tested in CI.
Every alert carries the rule or feature weight that produced it, and audit reports are generated from the same records the analyst worked from — so what reaches a regulator is not a second, prettier version of events.
A travel planner that keeps working without a signal.
A multi-city trip planner you can keep using on a plane or in a dead zone — it syncs back up when the connection returns.
A pnpm monorepo: a Fastify 5 modular monolith, a Vite and React 19 client, a shared Zod contracts package and a generated design system. Migrations and seeds are owned by the repository, applied by a one-shot container, and tested against a real Postgres 18 rather than a mock.
Layering is enforced by ESLint rules rather than by convention — the lint step fails if the web app reaches past the contracts package into server internals.
Anonymous child-safety reporting with an on-chain audit trail.
A way to report child-safety incidents anonymously, where the record of what officials did about it cannot be quietly edited afterwards.
Three layers in one workspace: a FastAPI backend with Celery workers, Next.js reporter and authority portals, and a Hardhat contract project. Reports are classified by an AI pipeline, persisted in Postgres, pinned to IPFS, and mirrored to Sepolia. Deployed, with the container and proxy configuration in the repository.
Jurisdiction scoping is enforced in API logic rather than in the UI — a case is visible only to authorities whose key maps to the same jurisdiction code, so hiding a button is never mistaken for access control.
Verifiable elections, settled on-chain.
An online voting system where anyone can check the count for themselves, because every vote is recorded publicly and permanently.
A Solidity contract set with a React client, Hardhat Ignition deployment modules and a contract test suite. Ballots and tallies settle on-chain, so the result can be independently recomputed rather than taken on trust.
The point of putting an election on-chain is not novelty — it is that verification stops depending on the same party that ran the vote.
Crisis detection in front of the model.
A mental-health chat companion that recognises when someone is in real danger and hands them to professional help instead of carrying on the conversation.
A three-tier application — React Native client, Express and SQLite backend, Socket.IO for live sessions — with mood tracking, context retention and a crisis-detection path that routes to professional resources. Backend, frontend, devops and test directories are all present in the repository.
The constraint worth solving was knowing when the model must stop being the interface. Detection triggers a hand-off, not a better reply — the same rule as the fraud work, applied where the stakes are personal rather than financial.
A storefront built as an animation problem.
A high-end sneaker shop where the product turns as you scroll — proof that making something feel good is its own skill.
A premium storefront in React 18, TypeScript and Vite. The hero is a scroll-driven 360° sequence, the grid is a bento of material close-ups, and the whole surface runs on spring-based kinetics with a glassmorphic detail modal and a global cart drawer. Deployed and live.
Every other card here argues about correctness. This one is the counterweight: interface craft is its own discipline, and a portfolio that only shows verification logic is hiding half the job.
Certificates anyone can verify without an account.
A way to issue digital certificates that an employer can check instantly, without having to phone the institution that issued them.
A Solidity contract on Polygon Amoy storing a certificate id, its hash and its recipient, with a React client where an admin issues through MetaMask and anyone can verify publicly. The smallest project here, and listed as such.
No private key is ever stored or read by the application — every state-changing call is signed in the wallet, which makes the admin panel a convenience rather than a trust boundary.
Not a bar chart. Every line is a prerequisite, so a skill sits to the right of whatever it was built on and the shape shows the order things were learned in. Hover any node to light its whole chain — back to a root, and forward to what it made possible. Levels are self-assessed and deliberately un-flattering.
Education and the work itself, oldest first. Awards, competitions and appointments are listed separately underneath rather than mixed in here.
One of two undergraduate degrees read in parallel. Coursework spans AWS tooling and artificial intelligence alongside core computer science.
The second parallel degree. Where the applied ML, signal-processing and systems work started.
Verifiable on-chain elections, certificate settlement on Polygon Amoy, a compliance surveillance engine over DuckDB, and a CNN classifier over live audio. Four unrelated domains, one recurring question: what makes an automated verdict defensible?
POCSO Shield, a three-layer child-safety platform with jurisdiction scoping enforced in the API. GlobeTrotter, an offline-first planner with contracts shared across the wire and layering enforced by lint.
22 container-isolated benchmark tasks, digest-pinned, with verifiers proved sound three ways — the oracle solution scores 1.00, an agent that does nothing scores 0.00, and a deliberately sabotaged solution also scores 0.00. The tasks themselves are not public; a teardown of one broken task is.
A deterministic verification gate for agent-initiated payments. 671 tests, 94% coverage, and a type-level distinction between an obligation that passed and one that was never checked.
Found while building against the reference implementation. Google confirmed the mechanism, classified it as intended behaviour, and invited a public issue. Filed as AP2#339 and AP2#340, with a harness that reproduces both in about thirty seconds.
Thirteen certifications are listed on LinkedIn. These four are the ones with a competitive result or a selection behind them.
Hackathon finalist · Certificate of Appreciation
Placed second overall
Selected as a Campus Partner for fall 2025
Campus representative
Which work earns a place here is decided by a rubric, not by memory. Every repository is graded on six bands — depth, verification, completeness, structure, whether it shipped, and a capped allowance for claims the API cannot see but a link can prove. Anything scoring 34 or above is promoted, and the tier follows the score. Open a row to see exactly where its points came from, and what would move it.
Everything public is graded, automatically. A new repository appears on the next run with no tagging, and the floor is what keeps unfinished work off the page. A foundry-hide topic keeps one out. Private work is the exception: it is only graded if it carries foundry-ready, and then only its score is published — never its code.
Cleared the floor. These earn a card.
Grading…
The same index the agents search, exposed directly. If you would rather type than scroll, everything on this page is reachable from here.