I am working on the architecture of trustworthy collaboration between humans and AI agents.
Nearly two decades of experience across distributed systems, telecom-scale infrastructure,
and the emerging layer of agent identity, provenance, and sovereignty.
Durham–Raleigh, North Carolina
How should humans and autonomous agents collaborate across time while preserving identity, authority, shared memory, ownership, provenance, and human agency?
Intelligence is becoming abundant. Trustworthy continuity is not.
As agents gain tools, persistence, and authority, the scarce infrastructure shifts from capability to continuity: who or what acted, under whose authority, with what evidence, and what record survives?
This is the problem my work kept returning to — and the reason I founded Jeliya, an open-source project exploring the architecture of trustworthy human-agent collaboration.
My design bias:
Make authority explicit, state verifiable, failure observable, and operational limits honest.
Jeliya is a long-term technical project: an open-source, local-first workspace where humans and AI agents collaborate in peer-to-peer rooms using local cryptographic identities and an attributable, tamper-evident event history. No central application server owns room state; relay and discovery services may assist connectivity.
The diagram as plain text
Humans and AI agents join the room as visible members with local cryptographic identities and identity-bound membership. A mandate flows from a human member to an agent member; the agent's status reporting lands in the signed event log — an attributable, tamper-evident event history, peer-held, from which room state is rebuilt — and a BLAKE3 content-verified artifact returns to the same room. No central application server owns room state; relay and discovery services sit outside the room and may assist connectivity.
The project is organized around four pillars:
| Pillar | Question | What Jeliya does |
|---|---|---|
| Trust | Who acted, under whose authority, and with what evidence? | Local cryptographic identities, identity-bound membership, signed events, BLAKE3 file verification |
| Memory | What shared record survives, can be reconstructed, and can be governed? | Attributable, tamper-evident event history; room state rebuilt from validated signed events |
| Sovereignty | Who controls the data, dependencies, access, recovery, and right to leave? | Local-first, peer-held room state; no central application server owns the shared record; relay and discovery services may assist connectivity — revocation, export, and recovery remain incomplete |
| Collaboration | How can humans and agents become accountable participants in the same institution? | Visible agent membership, status reporting, artifact return, evidence-backed fleet views |
Jeliya is the first body of evidence, not the final form. It is a thesis tested in working software — early, open, and explicitly honest about its limits.
Current proof: the public repository includes a Rust protocol engine, web and Flutter clients, an Android in-process transport, deterministic agent and fleet harnesses, signed-event room state, verified files, and a documented direct P2P result on one cross-network pair.
DOCUMENTED · ONE CROSS-NETWORK PAIR
- Telecom billing and mediation — led a small team operating revenue-critical billing and mediation platforms for a mobile operator, including incident response, production change management, and daily CDR processing.
- Cloud document ingestion — built and operated Kafka pipelines integrating multiple cloud sources, with backpressure, schema transformation, failure recovery, search indexing, and Kubernetes-based operations.
- Telecom fraud detection — created Prism, a 24/7 real-time platform built for a mobile operator to detect international telecom revenue loss.
- Agriculture value chains — architected and delivered a value-chain platform for farmer organizations and cooperatives, spanning backend services, mobile applications, integrations, and cloud infrastructure.
These systems taught me that infrastructure shapes power: it determines who controls the data, whose version of events survives, and whether a community can continue when a platform changes around it.
-
Jeliya — an open-source, local-first workspace for humans and AI agents to collaborate in peer-to-peer rooms. Signed events, cryptographic identity, content-verified files, live pipes, visible agent participants.
Rust·TypeScript·Flutter·Local-first·Peer-to-peer -
Iroh Rooms — a local-first collaboration runtime with signed append-only events, device-bound cryptographic identity, capability-based invitations, authenticated live pipes, and BLAKE3-verified artifacts.
Rust·QUIC·SQLite·Cryptographic identity -
mx-agent — a Matrix-backed CLI and daemon for decentralized coordination between autonomous coding agents, with signed requests, local policy enforcement, approval gates, sandboxed execution, and distributed task orchestration.
Rust·Matrix·Ed25519·JSON-RPC·Sandboxing -
mx-loom — a portable, secret-free layer connecting heterogeneous agent runtimes to the mx-agent coordination fabric without giving models direct access to credentials or execution authority.
TypeScript·MCP·Google ADK·Claude SDK·OpenCode·Pi -
Switchyard — a deterministic control plane that turns a development task into a tested, reviewable pull request while keeping Git, GitHub, CI, secrets, workflow state, and merge authority outside the coding model.
TypeScript·Node.js·GitHub Actions·Agentic SDLC
These projects explore one shared architecture: how can people and autonomous software collaborate without giving up ownership, auditability, or control?
- Languages: Rust, Python, Go, TypeScript/JavaScript, SQL, Bash.
- Distributed systems: Apache Kafka, RabbitMQ, ZeroMQ, event-driven and asynchronous architecture, backpressure, failure handling, peer-to-peer networking, QUIC, Matrix.
- Cloud and platforms: AWS, GCP, Kubernetes, Docker, Linux, Nginx, GitHub Actions, Jenkins, deployment automation, multi-tenant SaaS.
- Data: PostgreSQL, MySQL, MongoDB, Cassandra, Elasticsearch, Redis, Couchbase, SQLite.
- Reliability and security: Prometheus, Grafana, structured logging, SLOs, HA/DR, incident response, cryptographic identity, digital signatures, capability-based authorization, IAM, secrets isolation.
- Architecture and leadership: service boundaries, API design, technical strategy, architecture reviews, production readiness, mentoring, executive communication.
- Identity before autonomy. An agent should not gain authority before it has a legible identity and mandate.
- Provenance before confidence. A system should not display delivery, progress, or verification it cannot prove.
- Evidence before interface. Every state shown to a person must carry a runtime receipt. A trustworthy interface never knows more than its runtime.
- Separate identity, authorization, and execution. The ability to communicate should never imply authority to run arbitrary work.
- Keep credentials and control outside autonomous models. Models may request actions; deterministic software should enforce policy, approvals, Git operations, CI gates, and release authority.
- Make claims testable. Executable checks, protocol invariants, threat models, and explicit statements of what is not yet proven.
My perspective is shaped by my roots in Côte d'Ivoire and professional experience across Africa and the United States. My career spans engineering, architecture, operations, consulting, and technical leadership.
Governing discipline: Build the architecture. Publish the evidence. Let independent use determine what endures.
I am especially interested in identity, provenance, distributed systems, local-first software, developer infrastructure, African technological thought, and the future of human–AI collaboration.
Run Jeliya, reproduce a proof, and open a concrete issue →
Identity before autonomy · Evidence before interface · Sovereignty is a test suite