Neural network research shipped on custom streaming infrastructure; a reusable stack that halved development cycles inside a bank's compliance envelope; a deterministic runtime across modular Python with optional Rust and CUDA paths. The engineering is what makes the claims testable.
Where it shows up
3 positions, earliest first.
Founder · Sticheo
Jan 2017 – Jun 2019
Hired and led a team of 4 engineers from IIT Delhi on a neural network research project: soft semantic segmentation of video, restructuring cloud training to cut costs by 90%.
Invented and built the underlying technology — custom streaming servers integrating programmatically targeted product placements directly into video streams.
Secured alpha deployments with 4 enterprise clients and synthesised their feedback into the product roadmap.
Entrepreneur-in-Residence · Citi d10x
Jan 2020 – Apr 2024
Championed rapid prototyping across the studio, halving product development cycles through a reusable code stack and compliant cloud infrastructure — the difficulty being architecture that satisfies a regulated bank's controls without abandoning iteration speed.
Identified and analysed green-field Metaverse and Sustainability opportunities for Citi APAC; formulated the disruptive innovation thesis and advised leadership on where to place strategic bets.
Took 7 nascent products through validation with business heads across the organisation, building value propositions against each unit's priorities to secure executive sponsorship.
Supported the launch of cross-business products processing millions of USD per quarter, including novel financial instruments and specialised fintech infrastructure.
Independent Researcher · Relational AI & Cognitive Architecture
2024 – present
ARCANA — Conceived and built a framework for modelling intelligence as relationships and evolving processes rather than isolated objects. The system represents how patterns form, stabilise, change and recur across agents and groups, with every transition reproducible and auditable.
Research & implementation — Built a falsification-led programme: questions specified in advance, deterministic validators, machine-readable evidence, retained negative results and explicit limits. Direct AI coding agents across modular Python and optional Rust/CUDA paths, promoting only results that survive replay and reuse.