➜ ~/ngoanpv · model.fit(curiosity, epochs=∞)

Phạm Văn Ngoan

Pham Van Ngoan · Ngoan Phạm · NgoanPV — AI builder

Xin chào. I build AI systems that listen, reason and hold up in production — multi-agent systems, voice AI, retrieval and the evals that keep them honest.

Multi-agent systems Voice AI Evals & benchmarks

tap the film to pause

How I think about building

Start from the pain, not the stack. A real problem first; the model comes after.
One mind becomes many. An orchestrator that routes to focused sub-agents, each with its own tools — grounded in retrieval that cites its sources.
Lighter and faster. Latency is a feature: caching, routing, quantization, until the conversation feels natural.
Measure honestly. Nothing ships without passing an eval gate — accuracy, faithfulness, latency, safety.
Then let it reach everyone. Company-wide, and in regulated places, with a human in the loop.

The loop

Phạm Văn Ngoan's build loop: find the pain, build with AI, ship it, learn and iterate

Currently exploring

A simulated crowd passing an idea through friend networks, an evaluation gate that concludes or abstains, a small robot keeping and recalling memories, and a model compressed to run on a phone

Simulated crowds — multi-agent personas built from real people's answers, to try an ad, a price or a policy before it reaches real users.
How reactions spread — agents with friend networks who only pass something on when genuinely convinced, to surface second-order effects.
Evals that can say "not sure" — gates that check agreement with real people, stability across reruns and bias, and abstain when the evidence isn't there.
Benchmarking agents — how reliable, consistent and well-calibrated LLM agents really are, beyond the demo.
Memory for agents and robots — what to keep, what to let go, and how to recall the right thing at the right moment, across sessions and in the physical world.
Models on device — small, quantized and distilled models that run privately and offline on phones and edge hardware.