Kelden Dradul Dorji
Designing Accountable Collective Systems
Research, engineering, and community-building work · MIT Media Arts & Sciences applicant · Thimphu, Bhutan
Portfolio thesis
I build systems that help people and intelligent agents coordinate under real-world constraints — spanning AI products, computer vision, digital platforms, multi-agent research, and community-building in Bhutan.
- Human communities
- Meet → learn → build. Share knowledge, create trust, build local capability.
- AI-agent collectives
- Communicate → delegate → act. Coordinate roles, detect failure, enable supervision.
Coordination Observatory
Camera Culture · Multisensory Intelligence · Cyborg Psychology
The problem
Agentic AI systems can plan and delegate — but more agents don't automatically mean a reliable collective. Agents duplicate work, pass errors downstream, or become opaque to the humans supervising them.
What I built
- A research question: how do communication structure, adaptive roles, and incentives shape coordination and failure in heterogeneous AI-agent systems?
- A reproducible sandbox and observability layer for decentralized agent teams
- Evidence about why teams succeed, fail, or become unsafe — connecting artificial life models of coordination under scarcity to working multi-agent systems
Scale & outcomes
- Targeted research groups
- 3 labs
- How structure shapes failure
- 1 question
- Reproducible by design
- Open
AI Receipt Recognition
Founding engineer · stealth startup · 2024 – 2025
The problem
Small businesses needed a practical way to extract and organize data from paper receipts within existing workflows. The challenge was not only image recognition — the system had to work with real users, variable receipt quality, and unreliable connectivity.
What I built
- An AI-powered phone receipt-detector: capture, recognition, and structured export inside the workflows owners already used
- User correction flows — model performance alone was insufficient; operational reliability determined whether the product was useful
- Deployment coordination across business owners, frontline users, ERP providers, and telecom engineers — local constraints treated as design inputs, not edge cases
Scale & outcomes
- Paying customers
- 20+
- Transactions per month
- 20,000+
- End-to-end ownership
- Idea → launch
Artifacts
Applied ML & Scientific Software
Bhutan Data Scientists · 2017 – 2019
The problem
Scientific ML work only becomes useful when practitioners can understand its limits — it demands precise questions, documented assumptions, and inspectable results.
What I built
- Contributions to a DenseNet-based computer vision project identifying early tuberculosis signs
- Data cleaning and ML model research that worked inside real scientific workflows
- Data and inference infrastructure applied to 200,000+ professional-profile records
Scale & outcomes
- Records in inference pipelines
- 200,000+
- Healthcare CV in research use
- TB detection
- Inspectable, transferable results
- Documented
Reliable Digital Systems at Scale
Sky · AND Digital · 2022 – 2024
The problem
Contributing to the transition from Sky Glass to a new streaming platform while ~20 million users were live — where product, engineering, data, operations, and user behavior constantly shape one another.
What I built
- Frontend systems with Node.js, React, GraphQL, and Contentstack
- Experimentation infrastructure with Optimizely
- Observability with Grafana — making decisions inspectable so teams could learn from failures
- Clear ownership boundaries for systems changing under live conditions
Scale & outcomes
- Users served
- ~20M
- Migration under load
- Platform
- Zero-downtime change
- Live
AI Training Infrastructure
Alpine Sana · computer vision startup · 2024
The problem
The key challenge was not a single model run — it was creating reliable, repeatable workflows for data, training, iteration, and team collaboration inside a growing startup.
What I built
- Core AI-training infrastructure as early employee, leading a small team
- Reproducible annotation and training pipelines with Label Studio and AWS SageMaker
- Translation of ambiguous requirements into concrete technical work
Scale & outcomes
- Funding progression
- Seed → A
- Training workflows
- Repeatable
- Hiring and collaboration
- Team lead
Thimphu Tech Meet
Thimphu, Bhutan · Ongoing
The problem
Bhutan had motivated builders but no low-barrier space to bring unfinished ideas, share work publicly, and find collaborators.
What I built
- An open weekly gathering for developers, researchers, tinkerers, and builders — meet, learn, build
- Shared public notes, peer learning, and mentorship connecting Bhutanese builders to global networks
- Convenings spanning AI, agents, open-source, and entrepreneurship — including an AI roundtable with UNICEF participation
Scale & outcomes
- Meetings and public notes
- Weekly
- Mentor and speaker connections
- Global
- Collaborations sparked locally
- Prototypes
An Independent Builder Layer for Bhutan
With Loden & fellow builders · In development
The problem
Bhutan has institutions, companies, students, and motivated builders. What's often missing is a durable, open layer for learning in public, contributing to shared work, and turning curiosity into prototypes or ventures.
What I built
- A builder-led ecosystem designed to complement government and industry, not replace either
- Design principles: open by default with responsible boundaries, small projects completed publicly, local relevance with global standards, community ownership over individual dependence
Scale & outcomes
- By default
- Open
- Small projects shipped
- Public
- Owned, not dependent
- Community
Why MIT — and what's next
- Camera Culture — Raskar
- Agent communication, trust, delegation, and systemic risk
- Multisensory Intelligence — Liang
- Agents coordinating across text, images, tools, and human feedback
- Cyborg Psychology — Pataranutaporn
- Interpretable interaction and oversight of agent collectives
At MIT, I will develop the Coordination Observatory into a research platform — making multi-agent AI more cooperative, inspectable, and accountable. Long-term, I aim to reinvest experience and networks into Bhutanese researchers, open-source contributors, and founders.