Ikenna
Nwakwesi

Questionnaire in. Scored opportunity and three documents out.

First commit 2026-06-06, 19:43 EDT. Deadline 2026-06-11, 20:00 EDT. Submitted 08:42 on deadline morning - eleven hours to spare. First commit to submitted: four days, thirteen hours.

Dia turns a client discovery questionnaire into a scored opportunity and three ready-to-send documents: client profile, opportunity analysis, and proposal draft.

What shipped inside those five days: a fully deterministic scoring and document pipeline; a Gemini LlmAgent that uses those deterministic functions as tools and degrades gracefully without an API key; a dependency-free voice-enabled UI with TTS, dictation and hands-free mode; a pytest suite; an eval directory; a Dockerfile; one-command Cloud Run deploy; a security policy; and secret scanning.

Code: github.com/mdvnavy/Dia

Two tools built to avoid doing things by hand

PRESS - first commit 2026-06-06, 23:43 EDT, four hours after Dia's first. A second product started the same night as the first. A content-formatting engine built to produce the submission materials themselves. 47 commits over five days; its last commit before the deadline was the asset drop - thumbnail, headers, title card.

The capture client - 2026-06-09, 00:54. Rather than record the demo by hand, a direct obs-websocket client wired into the agent runtime, landed at 01:01 and constrained to local-only by 02:37. Dependencies pinned to the patch version, 26 passing tests in the same commits, and a deliberate contract: after a successful intake, fire a screenshot and a replay-buffer save, and swallow failures so capture can never break the response.

Two nights later, on deadline morning, the recording itself was agent-driven too - Desktop Commander at the controls, a scripted take driver in the repo, seven failed takes including full-screen captures coming out pure black, until the take landed at 02:46: live agent, correct score, on camera. The narrated cut was done by 03:05 - the delivery is a commit in the repo, timestamped 03:04. The tools did exactly what they were built to do.

Private agentic infrastructure - LTM mesh over Tailscale

A 3-server VPS agent fleet on a private Tailscale mesh (WireGuard, ACLs, MagicDNS) running multiple agent runtimes per server, with self-served artifacts to mobile and integrated search and scraping services.

I connected a vendor’s long-term memory product over the tailnet before their integration guide existed, rejected their public-tunnel recommendation in favor of a zero-exposure architecture, and debugged an undocumented TLS SNI requirement along the way.

Agent Eggs - a portable skill store

A distribution model for agent capabilities: skills packaged as versioned SKILL.md modules with runtime metadata, published to a Hugging Face dataset registry, and consumed by agents across the fleet.

Production skills covering the Google Workspace surface, plus a registry-of-registries index. Effectively a plugin model and developer on-ramp for agents.

Bespoke - the fleet’s voice

A desktop voice suite in three commands. speak: a TTS screen reader that streams Edge neural voices directly into mpv’s stdin - playback starts in seconds regardless of text length, with instant pause/seek over an IPC pipe. ask: a conversational agent on a verified Vertex AI pipe - ADC auth, no API key anywhere. live: real-time speech-to-speech on the Gemini Live API - sub-second turnarounds, barge-in mid-sentence, both sides transcribed to a dated log.

Born 2026-07-18 as a complaint that arrived with its own spec: the vendor screen reader was cooked - slow, fragile, a pause that didn’t pause - and the first message of the night already named edge-tts, because the experimenting was already done. The core was built overnight, in hours, to the 02:09 baseline commit. Everything after was chrome: a designed intercom family by the next afternoon, live cut and smoke-tested that evening, “the intercom, resident” merged 23:48. In daily production as the channel the agent fleet speaks its updates through.

Learning what code can do

Coming to code from design and data, you don’t see a developer’s tool - you see something with no ceiling: administration, organization, research, a dozen uses nobody thinks to apply it to. And on the moment it clicks: I can figure this out. No roadmap, just reading, researching, and pulling threads.

Agents as the interface

Using your computer primarily through agents as the interface - you orchestrate the agent, and the agent navigates the system for you.