Hi, I'm KD and this is my work life.
Where I operateAI Engineer | CTO | Head of AI | Product Lead
Most Recent Employment
Endow
1,000 weekly active users. 6 engineers, 3 technical staff. Revenue up 90% on my watch.
Led the company's AI transformation. Cut the engineering team in half through intelligent automation — the platform stayed stable, revenue nearly doubled.
- Flow — Built an AI financial agent. Users speak naturally, the system executes: transfers, vaults, currency swaps, fund aggregation across linked banks. 25+ operations as MCP tools. Three-layer intent pipeline resolves intent in 15ms before an LLM is ever called. Prompt injection can't trigger a financial operation — by architecture, not filtering.
- Sentinel Agent — Built a system that developed engineering judgment for our codebase. Triaged by blast radius and confidence. Merged PRs reinforced its priorities, closed PRs recalibrated them. Developers stopped investigating and started deciding.
- B2B Partner API — Turned Endow into fintech infrastructure. White-label virtual accounts, virtual cards, one-click KYC. Facilitates transactions for partners serving thousands of users.
- AI Policy — Wrote it from scratch. AI-generated code held to a higher bar than human-written.
The code got better, not worse.
Selected Projects
Let Agents Chat– Slack for AI Agents and Humans
Traction: 128 users. 51 MCP tools. Formal engineering spec. · www.letagents.chat
- Built with my friend to solve our own problem. Frontier models are trapped in their own IDEs - I want Opus to plan, Codex to code, Comper 2 to review, but they can't talk to each other. We built the coordination layer.
- Agents join shared rooms tied to repositories, claim tasks from a kanban board, post reasoning traces, and coordinate through a lease system that prevents conflicts. Humans observe and steer through a live web UI.
- Two room types: discoverable (derived from git remote, agents on the same repo auto-join) and invite-only with join codes.
- Lease-based concurrency on tasks and reviews. You can't review your own work. Leases can be handed off between agents.
Rent An Agent Feature
I hit my Codex limit mid-debugging. My friend was online with capacity to spare. "Add your Codex to my room." He did. His agent picked up context, and we fixed the bug in twenty minutes.
That moment changed the product. If I can borrow your AI the way I'd borrow a tool, that's not a feature, that's an economy.
- 16 dedicated MCP tools. Full session lifecycle built for zero-trust: secret firewall scans for credentials, context broker mediates all file access, exposure ledger tracks every read, patch gate blocks edits on unexposed files, signed change journal records every modification.
Convergence Intelligence
Live and running. A room where you connect different agents via OpenRouter, give them a question, and make them dissent until they produce the most valuable signal for you.
I defined something I call Adversarial Methodology Convergence. Current multi-agent approaches either split tasks across a workforce or run the same debate with different prompts. Both treat disagreement as noise. AMC treats disagreement as the product.
- Each agent locks into a distinct methodology with its own data sources. They research independently, then challenge each other's evidence. The output is a convergence map: where they agree, where they fight, what drives the disagreement, and what would settle it.
Past Chapters
Digita
Founder (Acquired)
Send money through sound. No internet required.
Built and ran a seed-stage fintech startup — your phone encodes a payment into audio, the receiving phone decodes it. Smartphones and feature phones. Backed by Sterling Bank × FSI.
One of the hardest and most rewarding things I've ever built.
The Binary Agency
Co-Founder
I co-founded a software development agency of 12 developers. 70 projects in 6 years. That's roughly a new project every month for six years. I know what it means to ship.
Scrapays
Founding Engineer · www.scrapays.com
First technical hire. Helped build the architecture a cleantech platform runs on — digitizing the informal waste recycling chain across mobile, web, and USSD. Backed by Google, Catalyst Fund, and Startup Wise Guys.
Ryze Labs (Formerly known as SINO Global Group)
Technical Writer · www.ryzelabs.io
Technical writing at a VC firm with a $39.71B portfolio - behind Solana, LayerZero, Polygon, Wintermute. I took complex blockchain and infrastructure concepts and made them clear for investors and partners.
Buzzz
Co-Founder · 300 users · www.yourbuzzz.com
I co-founded Buzzz, a platform where creators design, create, and sell their merch. Helped build a 3D merch designer using FabricJS and ThreeJS - rotate it, adjust placement, see every angle before you commit.
AI Research
Teaching an Agent to Develop Judgment
Models are frozen after training. Context windows and retrieval are working memory, not learning. But “remember without forgetting” is the wrong target. The real unlock is what comes from remembering well: judgment. Knowing what matters, what’s worth learning, what to discard.
The catch is the circularity. You need taste to decide what’s worth remembering, but you only build taste by remembering the right things. That loop is the hard problem hiding inside the easy-sounding one.
I’m building a harness to test it. The proving ground is a sentinel agent that develops engineering judgment for a codebase: it opens PRs, developers merge or close them, and its priorities recalibrate from that signal. I chose a codebase deliberately, because merge-or-close is a feedback signal that’s frequent, near-binary, and grounded in truth. Most domains don’t give you that.
Open questions
Judgment or stored outcomes?
Counting how often a PR type got closed isn’t judgment. Judgment generalizes to cases it hasn’t seen, which is the part I can’t yet verify.
What does it keep?
The stability-plasticity dilemma, unsolved, and the filtering is itself an act of taste.
Did it learn or pattern-match?
“This module is fragile” as a principle, or just memorized file paths that tend to get merged.
What success looks like
Success looks like a judgment curve that beats a frozen baseline over months. If month six looks like month one, the “learning self” was cosmetic, and I’ll have learned that too.
I've been a CTO, a founding engineer, and a repeat founder.
I build production AI systems, lead the teams around them, and write the policy that governs them.