Luminik helps B2B sales and marketing teams turn the events they attend (RSA, Black Hat, Money20/20, fintech summits) into pipeline. Bootstrapped, live with paying customers. I am the only full-time person on the team.

A couple of outcomes from our customers:

• A Series C cybersecurity company attributed $2.4M of pipeline to RSA, Black Hat, and one regional summit, sourced and sequenced through Luminik.
• A Series A identity verification company attributed $2M across fifteen fintech events in the second half of 2025.

• Aura at Bain Founder's Studio: a PE due-diligence platform. Took it from 0 to $3.6M ARR in 18 months. The team grew from five engineers to thirty-three over the same window.
• Mainteny: co-founded. $2.7M seed, European field-services SaaS.
• SnowOptix: Snowflake cost-optimisation tool. Bootstrapped, used by a top-3 global consulting firm.
• BlueJeans (started 2014): backend engineer on the video conferencing team. Where I first watched a real engineering org coordinate, before I had to do it myself with no team to lean on.

A dedicated team of agents, named after Batman side-characters, handles the recurring engineering work for Luminik. Each one is shaped around a specific job and operates under a contract for what it can decide alone and what comes through me.

The orchestration framework is one I built and open-sourced. Alfred OS schedules the fleet, claims one issue at a time so two agents never pick up the same work, isolates every run in its own git worktree, and gates what an agent is allowed to merge without me. Around it: Claude Code does the actual coding via ACP, Hermes Agent (Nous Research) provides the platform layer, and a local gbrain (Garry Tan) instance holds a 107-page knowledge base of specs and CLAUDE.md files queryable from anywhere on the box. The whole thing runs on a Mac Mini.

Full write-up: Building Toward a Team of Digital Employees on a Mac Mini. Alfred OS itself is documented at alfred.luminik.io.

While setting up the agent stack, three things were broken or badly defaulted in tools I depended on. All in public trackers.

• Hermes Agent #12251. Gemini auth regression after the project switched Google AI Studio auth from Bearer to x-goog-api-key. I isolated the repro on the live /v1beta/openai endpoint, and the maintainer landed a clean superseding fix the same day. Merged.
• Hermes Agent #12224. The WhatsApp PLATFORM_HINT told every model "WhatsApp does not render markdown" and silently overrode persona files. WhatsApp does support *bold*, _italic_, ~strike~, code, fenced blocks, and bullet lists. Open.
• gbrain #89. v0.12.0 hardcodes OpenAI embeddings, which silently undoes the zero-marginal-cost story for anyone running Gemini elsewhere. Confirmed the community Gemini-embedding patch on the same 107-page corpus and left a +1 with a repro. Open.

Long-form on solo building, agent infrastructure, encoding judgment for AI tooling, and what coordination looks like when you are the team.

• Building Toward a Team of Digital Employees on a Mac Mini. Hermes Agent + Claude Code via ACP + gbrain, three upstream bugs filed along the way.
• gstack, CLAUDE.md, and the Work That Frameworks Don't Cover. Where in-repo instruction files stop, and cross-session memory needs to start.
• What It Actually Costs to Build a Serious Product Alone in 2026. The cost structure of compressing a team into one desk.
• Building a $3.6M ARR Product Inside a Consulting Firm. The Aura story, five engineers to thirty-three.

All posts at prasad.tech.

LinkedIn · prasad.tech · prasadus92@gmail.com

For Luminik specifically: luminik.io.