
Adwaith R Nair
Blockchain and GenAI engineer. B.Tech CS and AI, MITS Kochi, class of 2027.
I build systems that have to be trusted.
Four systems, one question: how do you prove a machine did the right thing? Praman governs autonomous agents spending real money. Honora proves legal evidence was never altered. AegisAI decides when an AI may act alone and when a human must sign off. Assetize makes property ownership records independently verifiable.
Six systems, one question.
Every edge is a concern two projects genuinely share, in the code, not in the copy.
- Praman and Honorahash-chained append-only integrity
- Praman and AegisAIdeciding when an agent may act alone
- Praman and Nexusagentic tool-use loop design
- Praman and Assetizemoney correctness as integer minor units
- Honora and Assetizeon-chain records as independent audit trail
- AegisAI and Nexushuman-in-the-loop before a consequential action
- Nexus and Zyrarunning local models as a runtime
Case studies
Pramanagentic commerce control plane
A control plane that lets an AI agent spend real money while no model is anywhere in the authorization path.
What I ownedEverything. Solo build: architecture, mandate format, policy engine, ledger, two-phase executor, evaluation harness, tests and the decision log.
- Money moved under injection, 21 runs per arm
- 0 cases
Honorablockchain evidence management
Evidence hashes and every custody transfer recorded on a public chain, so the record is verifiable by anyone, not just by whoever runs the server.
What I ownedBlockchain, backend and testing: the EvidenceRegistry contract, the Express API, the integrity pipeline and their tests.
- Deployed and source-verified on
- Ethereum Sepolia
AegisAIgovernance layer for agentic AI
A decision layer that returns RESPOND, ESCALATE or REFUSE for every agent action, with confidence computed rather than guessed.
What I ownedwatsonx.ai prompt design and the reasoning logic: the RESPOND, ESCALATE and REFUSE decision layer, the policy-confidence model and the rule ordering.
- Demo scenarios verified end to end
- 3
Assetizetokenized fractional real estate
A Phase 1 investor demo of fractional property ownership where every demo action produces a real, explorable transaction hash.
What I ownedEverything. Solo build at SupeAI: data model, repository and service layers, API, on-chain integration, documentation and the bug log.
- Boundary violations caught by self-audit
- 3
Also built
Nexusmulti-provider CLI AI agent
Four providers behind one interface, three agentic tool-use loops written directly against the Anthropic, Gemini and OpenAI REST APIs with no agent SDK.
Everything. Solo: the provider abstraction, the three tool-use loops, the confirmation gate and the CLI.
Zyralocal-first smart-home voice assistant
A five-mode fallback hierarchy that degrades from full conversation down to the relay board's own access point, so it keeps working when the server, the hub or the Wi-Fi fails.
ESP32-S3 firmware in ESP-IDF and C, the Python AI server, and the system architecture. Two-person collaboration with Abhijith A, whose account holds the primary repository and commits.
Open source
Google Summer of Code 2024, Oppia Foundation.
Stack
Hover or focus a technology to see where it was used. Hover a project to see what it was built with.
- Praman
- Honora
- AegisAI
- Assetize
- Nexus
- Zyra
Built with
- TypeScript
- Node.js
- Bun
- Next.js
- Express
- PostgreSQL
- Prisma
- Supabase
- MongoDB
- Solidity
- Hardhat
- ethers.js
- IPFS
- Ed25519 signatures
- Python
- FastAPI
- ESP-IDF (C)
- Zod
- Tailwind
- GitHub Actions
- Vercel
- Prompt design
Worked with
- watsonx.ai
- Razorpay API
- Anthropic API
- Gemini API
- OpenAI API
- Polygon Amoy
- Ollama
- Faster-Whisper
- Kokoro TTS
- Home Assistant
- shadcn/ui
- Framer Motion
- @react-pdf/renderer
- Recharts
Exploring
- Rust
- Solana
- Zero-knowledge proofs
- Trusted execution and attestation
How I build
I direct AI agents the way I would run a small team. I write the spec, decide the architecture and the boundaries, and hand out single-purpose tasks. Praman has 142 commits that each do one thing because each one was one task.
Decisions are written down before code. Praman carries a 24-entry decision log and Assetize about 3,900 lines of documentation, so anyone, including a model with no memory of yesterday, can pick up the reasoning and not just the code.
Nothing consequential runs unchecked. Nexus asks before every file write and shell command. Praman's authorization path has no model in it at all. The same rule applies to how I work: agents propose, tests and I verify, and I own what ships.
Verification is measured, not asserted. Held-out test splits, ablation runs, a self-audit that caught three real boundary violations in Assetize, and a logged bug list. If a number is not on this site, it is because I did not measure it.
Say hello.
Graduating 2027. Reachable now, for AI and agentic engineering, blockchain and backend work.
adwaith.r.nair189@gmail.comKochi, Kerala, India. B.Tech Computer Science and AI, MITS Kochi, 2023 to 2027, CGPA 8.5.