Building where technology meets consequence

Gurugram, India

Chimère Pay moves a US client's dollars into an Indian freelancer's account in 20 seconds at 1% all-in, against an industry norm of one to three days at 2–6% with funds frozen in transit. I built the stack end to end: the settlement rails, the compliance engine that scores every transaction with our own AML model in about 25ms and generates purpose codes and FIRA automatically, and the agent-native layer — a payments MCP endpoint any AI agent can call, and agentic wallets with budgets, permissions and a kill switch.

• 20-second settlement
• 1% all-in, no forex markup
• In-house AML model, ~25ms
• MCP endpoint for AI agents

Decentralised AI infrastructure: an identity for AI held on-chain, with the workflows, models, data, infrastructure and access to each represented as tokens. Deliberately ambitious and early — five hackathon wins in ten months came out of building it.

• Five hackathon wins
• $15,000 Coinbase grant
• On-chain AI identity

Gurugram, India

Two and a half years of analytics and AI across banking, telecom, FMCG, IT infrastructure and heavy engineering — supply chain and procurement, value engineering, M&A support and geospatial work, delivered from India to clients in APAC, Europe, the US and the Middle East. On the technical side: data pipelines, NLP and LLM-based extraction, ML models, evolutionary optimisation, and the visualisation layer over all of it, deployed on cloud and internal platforms. Increasingly I was the technical voice in the room — advising on what was feasible and keeping the engineering aligned to what the client actually needed. The work was recognised with the firm's global award for innovation and client impact.

• Global award for innovation and client impact
• LLM extraction + evolutionary search
• Data pipelines & MLOps

JFWTC, Bangalore, India

Worked on the frameworks GE uses to monitor aircraft engine health: IDAF for detecting faults and IPAF for predicting them. The dataset ran to tens of terabytes of sensor data across millions of flights, so most of the problem was distributed systems before it was machine learning — I reworked how the data was stored and processed and cut total pipeline time by 5×, from days to hours, on 70–80% less storage. Every failure mode had its own ensemble of stacked predictions, classification for detection and LSTMs for forecasting degradation, and I built the pipelines that retrained them automatically as fleets and routes drifted. On the detection model I improved speed 5×, raised F1 by around 9%, and cut false negatives — the error that actually matters when a missed fault costs far more than a false alarm. I then built IPAF, the deep-learning prediction framework, and integrated it into the existing system.

• 5× faster pipelines
• F1 up ~9%
• Ensembles, LSTMs, auto-retraining

Gurugram, India

Co-founded DataNest Ventures and built CoviFight, a contact-tracing system designed to raise COVID-19 detection rates without asking anyone to give up their identity. It detected exposures 3× better than DP-3T, the prevailing decentralised standard. The core idea was to invert the usual privacy trade-off: keep the data public and the identifiers private. Every user held a decentralised verifiable ID on chain with their public key as the only identifier, devices exchanged rotating Bluetooth IDs on contact, and each device ran federated analytics at the edge — computing its own infection probability locally and publishing only that score. As CEO I led 12 developers and worked as Scrum Master across cross-functional teams while building the ML models, the database schema, and the mobile and web application.

• 12 engineers
• 3× over DP-3T
• 2nd of 2,000+ at EU vs Virus