Cẩm Huỳnh qcam

Senior Backend Engineer (Elixir)

Fully remote · Flexible hours

Come build the part of AI nobody's building yet.

Most AI companies are building tools. We're building relationships, and that's a harder problem than any API integration.

At Iris, a conversation is a write, not a read. We build memory that persists after the session ends, and identity that holds up across months. We use real voices and real faces. Every message adds to the relationship instead of disappearing.

Nemotron 3.5 ASR Streaming 0.6B — On-Device Feasibility (Android/iOS/Web)

Model: https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b

Model facts

Architecture: Cache-Aware FastConformer encoder (24 layers) + RNNT decoder, 600M params — streaming with configurable 80ms–1120ms chunk latency
Formats: safetensors, PyTorch, GGUF (q8_0) via NeMo-Speech.cpp (C++ runtime)
Official hardware/OS support: NVIDIA GPU archs only (Ampere/Blackwell/Hopper/Jetson/Lovelace/Turing/Volta), Linux / Linux4Tegra. No ONNX, Core ML, or TFLite export path officially provided. No mobile or web runtime published.
40 language-locales, license OpenMDW-1.1

Iris Channel Performance — 2026-04-25 (full day)

Both nodes (a1 + b1), full day data (4693 log lines).

Channel Joins (1840 total)

Channel	n	avg	p50	p95	p99	max
ChatChannel	1297	49ms	37ms	140ms	191ms	316ms

Iris Knowledge Pipeline — Review & Improvement Proposals

Current Architecture (as-is)

Data model

persona_knowledges — knowledge metadata: name, instructions (text[]), context/operator scope
persona_knowledge_chunks — individual chunks: content (text), embedding (pgvector 384/1024), position
personas_knowledge_graphs — extracted entity/relationship graph per persona (single JSONB blob)

	require "msgpack"

	class DateTimeExtension
	def self.unpack(data)
	case data.size
	when 4
	seconds = data.unpack("N")[0]
	Time.at(seconds).utc

	when 8

	defmodule GCounter.Node do
	use GenServer

	def start_link(node_name) do
	GenServer.start_link(__MODULE__, node_name, [name: node_name])
	end

	@impl true
	def init(node_name) do
	init_counter = %{node_name => 0}

	[
	{
	"source": "quan-cam.com",
	"url": "https://quan-cam.com/rss",
	"category": "personal"
	},
	{
	"source": "medium.com/@hqc",
	"url": "https://medium.com/feed/@hqc",
	"category": "personal"

	defmodule RFC822 do
	import NimbleParsec

	day =
	choice([
	string("Mon"),
	string("Tue"),
	string("Wed"),
	string("Thu"),
	string("Fri"),

	<?xml version="1.0" encoding="UTF-8"?>
	<breakfast_menu>
	<food>
	<name>Belgian Waffles</name>
	<price>$5.95</price>
	<description>Two of our famous Belgian Waffles with plenty of real maple syrup</description>
	<calories>650</calories>
	</food>
	<food>
	<name>Strawberry Belgian Waffles</name>