Software Engineer - Agentic Delivery Platform
Summary
Build and harden an AI agent-based delivery platform that turns product specs into PRs, handling sessions, previews, and multi-agent workflows for a fintech portfolio product.
About the role
We are building a fintech portfolio platform (encrypted client-side data, market data integrations, bank statement import) and a growing set of supporting services.
Our current delivery bottleneck is the Agentic Delivery Platform (coding layer): a production system where product managers and engineers drive feature work through AI agents - from clarification and prototyping to multi-agent development, preview environments, quality gates, and PRs.
This is not a chatbot wrapper. It is a stateful agent runtime: orchestration, sandboxed worktrees, Docker preview stacks, streaming UI, model routing, interrupt/resume, and human-in-the-loop controls. When this layer is unreliable, product velocity stalls. Making it production-grade is the priority of this role.
What you will own
Primary (blocker)
Design and harden the agent delivery runtime (sessions, locks, work-state SSOT, recovery after restarts)
Evolve the multi-agent pipeline: plan → audit → challenge → execute → test/code review → gates → commit
Own preview lifecycle: lazy per-session stacks, idle suspend/wake, compile repair, publication gates
Improve agent protocols: structured tools, prompt policies, read-only vs write modes, iteration loops
Operate the coding-layer production path: deploy, isolated agent runtime, preview isolation, reliability
Model routing: presets, per-agent model roster, allowlists, cost/latency/quality trade-offs
Secondary (ecosystem)
Contribute across a multi-repo stack when needed: main web app (React/GraphQL), Python microservices (LLM parsing, market/FX data), shared libraries
Keep cross-service contracts stable while agents and humans ship changes safely
Domain
Portfolio / wealth-tech product surfaces
Client-side encryption constraints (server never sees plaintext sensitive data)
Market data and statement-import workflows
Internal AI-assisted delivery tooling used as the main path to ship product work
Requirements
Must-have
Strong TypeScript across Node and React
Solid Docker experience (Compose, networking, volumes, prod vs local)
Experience building stateful backend systems (sessions, locks, idempotency, restart recovery)
Hands-on work with LLM/agent systems (tool calling, custom orchestrators, agent CLIs, or similar)
Comfort designing contracts and SSOT (APIs, SSE/events, persisted state) instead of brittle chat heuristics
Ability to work across multiple services/repos, not only one application package
Strong plus
Multi-agent orchestration (planner/executor/reviewer patterns), evals, tool-use reliability
Git worktrees / sandboxed agent runtimes / preview-in-the-loop workflows
Python services (FastAPI or similar)
Production ops: CI/CD, reverse proxies, separate app image vs agent runtime
Interest in current agentic engineering: model routing, AI-native DX, human-in-the-loop systems, failure taxonomy
Nice to have
GraphQL + MongoDB
Applied crypto awareness (RSA/AES, client-side encryption constraints)
Fintech or market-data experience
Observability for agent pipelines (structured logs, step telemetry)
What success looks like
The Agentic Delivery Platform stops being the delivery blocker
Preview and agent runs are predictable under failure (interrupt, orphan processes, flaky builds)
PMs/engineers can go from intent → reviewable preview/PR with clear gates and recoverable errors
Multi-agent workflows improve iteratively without collapsing into prompt spaghetti
