Airstride
An AI partner-ecosystem platform, plus Carmen, the agent that sources, sequences and books partnerships on its own. 14 months as a core engineer.
Airstride is an AI-powered partner-ecosystem platform with two products under one roof. Acquisition finds and wins new channel partners; Managementruns the program and the deal pipeline once they're signed. I joined as a core engineer and spent 14 months building across both.
The through-line is Carmen, the AI agent that automates the partnerships motion end-to-end, from sourcing the right partners to running live outreach. I own large parts of the frontend and reach deep into the backend and the AI layer.
- Role
- Software Engineer · frontend lead & backend / AI collaborator
- Duration
- May 2025 — Jul 2026 (14 months)
- Surface
- 2 products (Acquisition & Management) across 15+ modules
- Focus
- Agentic workflows · UI architecture · API resilience
- Stack
- Next.jsTypeScriptMantineNode.jsAnthropic SDKOpenAI SDKGemini SDKInngest
Partnerships don't
scale by hand.
Running a partner program is a chain of slow, manual jobs: research who's a fit, find the right people, write outreach they'll actually answer, then track every deal and find the overlap between your accounts and theirs. Multiply that across hundreds of partners and the whole motion stalls.
The goal was to turn that chain into software an AI could run reliably: fast enough to feel live, resilient enough to trust with real outreach, and transparent enough that a human can step in at any moment. That meant three product surfaces that had to feel like one.
Multiple LLM providers, each better at different steps, so the system had to route work, not bet on one model.
External APIs (LinkedIn, email, partner data) are flaky. Outreach can't silently fail, so every call needed retries, backoff and observability.
Dense, data-heavy UIs (kanban boards, match tables, live metrics) that still had to match Figma at near-perfect fidelity.
Campaign Acquisition,
run by Carmen.
Carmen turns the top-of-funnel partnerships motion into an autonomous pipeline: sourcing partners, designing the outreach, personalising every message and reporting on what works. Each stage is its own agent step: observable, retryable and interruptible by a human.
Source
Scans the market against your Ideal Partner Profile and scores every match for fit.
Build
You describe the goal in chat; Carmen designs the multi-step LinkedIn + email sequence.
Personalise
Finds the right decision-makers and generates tailored messages for each one.
Engage
Runs the sequence across LinkedIn & email with follow-ups, autonomously.
Measure
Live reporting: messages sent, replies, response rate and weekly progress.
// Each stage is a durable, retryable Inngest step. export const engage = inngest.createFunction( { id: "carmen-engage", retries: 4 }, { event: "partner/qualified" }, async ({ event, step }) => { const draft = await step.run("draft-outreach", () => route(event.data.partner) // pick the best LLM for the task ); return step.run("send", () => sendWithBackoff(draft)); } );
Two more faces.
Deals Management
A drag-and-drop pipeline for every partner-sourced opportunity, from Prospect to Closed Won, with weighted forecasting on top.
- A kanban board with optimistic drag-and-drop across seven deal stages, plus board / forecast / list / needs-review views.
- Weighted pipeline& funnel: value rolls up per stage and against a target you set.
- Deal registration, review queues and owner filters, all on live, synchronised state.
Account Mapping
Discover joint opportunities by comparing account lists with a partner, securely and at scale.
- A match engine that classifies every account as exact, probable or possible. 125 accounts analysed per session, 123 matched exactly in this run.
- Consent-based data sharing: request → approve → map, so neither side exposes its full book.
- Surfaces net-new overlap and shared accounts to target together.
What made it hold up.
Carmen routes each step to the model that does it best: Anthropic, OpenAI and Gemini SDKs behind one interface, so no single provider is a bottleneck or a single point of failure.
Every stage is an Inngest step function: automatically retried, observable, and resumable. A failure mid-pipeline never loses work or double-sends an outreach.
Built REST integrations handling 5,000+ daily requests across LinkedIn, email and partner data, with exponential-backoff retry logic that sustained 99.5% uptime at peak.
Dense, live surfaces (the deals kanban, match tables, campaign metrics) built on React Query with optimistic updates and intelligent caching that cut ~30% of redundant network calls.
Translated Figma into reusable, accessible Mantine components at 95% fidelity across desktop and mobile. The product looks exactly like the design, everywhere.
Carmen took a motion that ran on hours of manual prospecting and made it run itself, across three product surfaces that feel like one.
Campaign Acquisition, Deals Management and Account Mapping, live in the Airstride app today, plus the AI pipeline underneath.
Making autonomous outreach feel safe: retries, observability and human override designed in from day one.
Deeper signals, tighter qualification, and expanding Carmen across more of the partnerships lifecycle.