Founding Developer & Tech Lead @ GeniusCFO · SDE II, AI/ML @ QuadB Technologies
Note
I build AI systems that fail loudly instead of quietly. A model that is confidently wrong is worse than one that refuses to answer. Most of my work is the structure around the model — validation gates, evaluation harnesses, and the escalation path that sends an uncertain answer to a person instead of into someone's books.
GeniusCFO — an AI accounting platform for Indian businesses. Ask a question in plain English; it keeps double-entry books underneath. I have been on it since the first commit and lead the backend and AI architecture.
flowchart LR
Q(["Document<br/>or question"]) --> SUP["Supervisor<br/>plan · route · LangGraph"]
SUP --> AG["9 specialised agents<br/>typed A2A protocol"]
AG --> H{"Accuracy harness<br/>check → repair → flag"}
H -->|"clears every check"| L[("Ledger<br/>double-entry")]
H -->|"anything uncertain"| R["Human review"]
classDef start fill:#E9FCCA,stroke:#1D1D1D,stroke-width:1.5px,color:#1D1D1D
classDef core fill:#FFE7A9,stroke:#1D1D1D,stroke-width:1.5px,color:#1D1D1D
classDef agents fill:#DBF5F0,stroke:#1D1D1D,stroke-width:1.5px,color:#1D1D1D
classDef out fill:#E3F2FF,stroke:#1D1D1D,stroke-width:1.5px,color:#1D1D1D
classDef human fill:#FBEBEA,stroke:#1D1D1D,stroke-width:1.5px,color:#1D1D1D
class Q start
class SUP,H core
class AG agents
class L out
class R human
91.4% |
85% |
5,000+ |
9 |
6 |
|---|---|---|---|---|
| field extraction accuracy |
reconciliation auto-matched |
documents per month |
specialised agents |
microservices in production |
Important
That 91.4% is measured against a golden document pack on the live code path, not in a notebook. The same run surfaced a classification gap I would otherwise have shipped. A number that only ever flatters you is not a measurement.
|
Supervisor orchestrating 9 agents over a typed protocol. Four-engine OCR chain, per-tenant isolated hybrid retrieval, PII tokenised before the model call, and a nightly loop that turns production corrections into versioned prompt rules. |
Clinical RAG over FDA drug labels and ADA/JNC guidelines. Hybrid BM25 + dense retrieval, citation-grounded answers, and a RAGAS gate in CI that blocks low-faithfulness responses before deploy. |
|
Real-time stablecoin monitoring built as a multi-agent system. 3rd Prize at ETHOnline 2025 — Best Use of ASI Alliance / Fetch.ai, global hackathon. |
⛓️ DFinance
Aave-style lending rebuilt for the Internet Computer's canister model. I built the liquidity layer end to end — pool accounting, supply and borrow flows, and the state behind them. |
More — Bitcoin, decentralised infra, and client work
| Project | What it is | Stack |
|---|---|---|
| USDB | A stablecoin issued natively on Bitcoin via Runes. Designed the UTXO and Runes transaction mechanics as second lead backend dev. | Bitcoin Runes UTXO |
| OpenxAI Studio | Deployment console for a permissionless AI network on decentralised GPU compute. Led the India-based full-stack team. | Next.js Web3 |
| Freelance commerce | Three Dubai-based commercial sites delivered solo, brief to launch | Shopify WordPress |
| Core |
|
| AI & Agents |
|
| Data |
|
| Platform |
|
Tip
100% accuracy is achievable — but only in the deterministic layer. No amount of prompt engineering makes a model reliably correct. But arithmetic is. A checksum either validates or it does not. So I make the deterministic layer complete, then route every model output through it. What survives is verifiably consistent. What does not, a person sees.
- Measure before you tune. Every hour spent on prompts before you can score them is an hour spent guessing.
- Sample your successes. Documents that go to review generate correction signal. Confidently auto-posted ones generate nothing — so a confident error is invisible unless you go looking.
- Prefer an honest refusal to a plausible answer. Everywhere, without exception.
- Make every automated improvement reversible. Otherwise you have traded one silent failure mode for another.
| 🥉 3rd Prize — ETHOnline 2025 | Best Use of ASI Alliance / Fetch.ai, global hackathon |
| 🇮🇳 National Finalist — Kavach 2023 | Government of India cybersecurity hackathon; led a six-member team |
| 🏆 Winner — BinaryHacks | Shipped a full-stack app as team lead inside the hackathon window |