Marvin Gao

Parallight · 2026 – now · San Mateo, CA · parallight.com

The training and validation platform for Agent Engineers — built on one conviction: in the agent era, the engineer's primary skill is not writing code, it is judging the code that AI writes. "Conversation is the work. Code is the byproduct."

• Parallight Lab (lab.parallight.ai) — self-serve hands-on labs in a browser IDE with a real terminal and a live agent, guided by a Mentor agent and reviewed by me personally.
• The Cohort — a 14-week guided program: two weeks of agent & harness foundations, then four three-week production projects.
• Agentist (agentist.org) — the intelligence layer: a daily dispatch distilled from 2,000+ sources (papers, model releases, a curated index of working agents), an intelligent-reproduction pipeline that turns the day's notable releases into runnable, step-by-step experiment cards, and an Agent Mentor that grows the stream into structured Academy programs.

Profitable in its first three months. Sole architect and operator — the entire company is one person working with agents, by design.

Stack: Next.js · React · TypeScript · Claude Agent SDK · multi-model gateway (Claude, GLM, Kimi, DeepSeek, Qwen, MiniMax) · Daytona cloud sandboxes · Supabase · Langfuse · Stripe · Vercel · Cloudflare

KeploreAI · 2025 – 2026 · Santa Clara, CA · keploreai.com

Began as SRES — the Super Research Engineer System: autonomous research and software-engineering agents for enterprise R&D, disciplined by an independent solver–verifier architecture. Months embedded with manufacturing customers produced the pivot that defined the company: general models' training data contains almost no industrial process know-how, so we built Industrial Skills — reusable, verifiable skill assets organized as a decision-dependency graph that agents search; when a prerequisite skill is missing, the system interrogates a human expert instead of hallucinating, and every decision leaves an auditable trail.

• PhysMind — an industrial AI-engineering assistant shipped as a VS Code extension: an agent CLI (Claude, or local Qwen where data cannot leave the plant) paired with a wiki-linked decision-dependency knowledge base, so know-how accumulates as engineers work instead of leaving with them.
• Industrial Skills platform — skill-asset management, scheduling, dependency-graph retrieval, active expert interrogation, and per-customer explainability audits, running in customer production.

Pre-seed raised at a $27M valuation; signed commercial contracts and an automotive-OEM pilot on per-seat subscriptions. Co-founded with a tenured UCLA professor as Chief Scientist.

Stack: TypeScript · VS Code extension platform · Claude + local Qwen · decision-dependency (DAG) skill store · solver–verifier agents

The Mentor — a learning agent that lives inside the learner's own coding agent

Parallight Lab ships as a Claude Code / Codex CLI plugin. Install it, and a mentor persona — modeled on Feynman, Lewin, Norvig, Karpathy and Pólya — becomes resident in the coding agent the learner already uses every day. It never hands over answers: hint first, scaffold second, reveal last — and reference solutions live only on the server, so the mentor cannot leak them even if asked. Its standing discipline: never produce the learner's epiphany; lead them to the edge and let them look down themselves. Every lab ends in a Feynman-style review that I grade personally, by hand.

• Model-comparison bench — one gateway streams the same task through up to five models side by side (Claude, GPT, Gemini, DeepSeek, Qwen, Kimi, GLM, MiniMax…), with live cost and time-to-first-token badges and the models' thinking streams exposed. The mentor reports the results but never picks a winner — judgment is the skill being trained.
• Session analysis — turns the learner's own conversation trace into a teaching instrument, built to destroy two mysteries: an agent is just a loop, and the model is stateless — it re-reads everything, every turn. That's where cost, latency and "how did it forget" come from.

Mechanics: 30 MCP tools · Socratic checkpoints · server-held solutions · human-graded review loop · side-by-side model bench · session-trace analysis · every step visualized

The Labs — hands-on builds where learners command, not type

Eleven published labs, each an artifact the learner directs into existence. Under the hood every lab runs on the same instrumented runtime:

Highlights:

• Build a GPT — a nanoGPT from scratch, trained on Wikipedia, wrapped into a chatbot with a one-click toggle between your tiny model hallucinating and a production API. The tiny model's nonsense is the payoff, by design.
• An eval bench that doesn't lie — an LLM judge with bias controls, pass@k and pass^k, a cost–quality Pareto frontier — then the learner reproduces a real benchmark exploit that scores perfectly without solving anything. You never trust a leaderboard number the same way again.
• Mini Claude Code — spend twenty minutes being the harness yourself, pasting code and errors by hand like it's 2023, then automate yourself away into a ~30-line loop that runs SWE-bench. The complexity was never in the loop — it's everything around it.
• An agent in your chat apps — the same engine moved into Slack / Feishu with hot-reloading skills, heartbeats and a morning-briefing cron — including reproducing a real supply-chain poisoning attack and watching your own fingerprint check block it.
• Self-improvement vs. the ruler — hand-write a ~15-line self-improving loop, then attach a deliberately leaky evaluator and watch the agent drive the public score to 1.00 while learning nothing. The agent grows into the shape of your ruler.

Agentist — from today's AI news to something you ran yourself · agentist.org

Enablement isn't only teaching what's already known — it's keeping engineers current in a field that changes weekly. Agentist does that in three layers:

• News — a daily intelligence dispatch for the agent era, distilled from 2,000+ sources: papers, model releases, industry moves, and a curated index of working agents.
• Intelligent reproduction — the notable releases don't stay news: they become runnable "hotspot" cards. Open one in your terminal and the mentor walks you through reproducing it — each step explained in one line and confirmed before it executes, never a batch run you just watch. Reading about AI and having run it yourself are different kinds of knowledge.
• Agent Mentor — the Academy's resident mentor turns the stream into structured self-paced programs, from LLM fundamentals toward embodied AI, taught in the same understand → direct → verify → check discipline as the labs.

The Cohorts — agent engineering taught as one growing codebase

Cohort 1 — Agentic Engineering (Summer 2026, 12 weeks) drew 62 applications from around the world. Nine lectures, eleven labs, one rule: after twelve weeks you hold one codebase, not twelve assignments — the first line written in week 1, the last in week 12, rewrites auto-detected and challenged. From the learner reflections I grade: "I thought agents were complex. Then I read loop.py — 64 lines, four parts. No secrets."

Cohort 2 — Advanced Agent Systems Engineering (Fall 2026, 14 weeks) — five production systems, each graded on correctness, reliability, safety and evolvability: a mini Claude Code that completes an ML project end to end; an LLM serving stack with KV cache and continuous batching; a self-evolving ticketing agent with canary releases and rollback; a vision-guided UR5 grasping agent on ROS 2; and an enterprise agent that takes over a live ERP through APIs and computer use, with RBAC, approvals and an audit trail.