helios · helios-repl
Depth 0 · running
Cycle 01
Helios
Secure, sandboxed agentic runtime. Recursive language models that write code, run it in a live interpreter, then call themselves.
View on GitHub Java 25 · Maven Central · MIT Read more
Critical systems,
Forward-deployed strike teams. We design and build your critical software and agentic systems, from architecture to deployment. We just love building. No bloat. No slop.
0×
Faster than traditional agencies
0%
Capital to engineering
IC7+
Talent floor
0 wk
First deployment
Engineering DNA and Operational History.
Decades of tenure at platforms the world runs on. Now deployed in the most regulated stacks in the world.
AI bolted onto legacy software is broken by design.
Most enterprises run on 100+ disconnected tools. Fragmented data. Retrofitted AI. Workflows frozen by what each vendor will permit. The architecture is the bottleneck, not the tooling.
The Solution
Three layers. One platform.
Semantics, reasoning, surface. Each composable, each yours.
Unified data foundation.
Every source: EHRs, claims, transactions, SKUs. One queryable graph modeled the way your domain actually works. Not the way each vendor permits.
Agentic harnesses.
Agents that use recursive, code-based reasoning. A harness that is predictive, real-time, replayable. Built to be audited, not trusted on faith.
Adaptive interfaces.
A unified workspace that adapts to the signal. Most cases auto-route in milliseconds. Edge cases route to humans with the audit chain attached. One surface, every decision.
Spec to platform. In months.
Forward-deployed strike teams shipping AI-native, composable platforms to production. Healthcare and finance.
Clinical Research
Clinical Intelligence Platform
Six modules on a unified platform.
A unified, AI-native control plane across the trial lifecycle. Authoring, review, and analytics with humans in the loop at every step that matters.
Ingestion
Multi-sourceQuality
Rules + riskReview
TriageInsight
GenerativeWorkbench
R + PythonAdmin
Governance12 wk to first deployment · 36 wk to Phase 1 live
Read the engagement
Digital Health
Cadence
Digital phenotyping infrastructure.
Hardened mobile data collection, high-speed ingestion, and an Auto-ML baseline. Production-grade infrastructure for behavioral signals at scale.
SDKsiOS · Android
PipelineIngest
AutoMLPatterns
13 wk to MVP · 4 milestones · ready to scale
Read the engagement
Born in the field. Maintained in the open.
Forward-deployed work doesn’t run on off-the-shelf stacks. We built our own. Helios runs our agents. Sail provisions our dev environments. MIT-licensed, from the first commit.
Sail
Provisions bare-metal servers and isolates AI coding agents in declarative dev environments. Blast radius: one container. Recovery: ninety seconds from a YAML file.
View on GitHub Native binary · Zero deps · MIT Read more
sail
Running
$
Cycle 01
Field Intelligence
Knowledge logs from the field.
Engineering notes, research commentary, and lessons from production deployments.
View all logs
MAR · 03 · 2026 8 min read
RESEARCH & ENGINEERING
Isolating AI Coding Agents on Bare Metal: Incus, Podman, and a $50/Month Server
How Singular uses system containers, rootless Podman, and declarative YAML configs to run isolated, multi-project AI agent environments on a single $50/month bare-metal server.
Read log
DEC · 08 · 2025 30 min read
RESEARCH & ENGINEERING
Notes: Defeating Non-Determinism in LLM Inference
Study notes on why LLMs produce different outputs for the same prompt—even at temperature 0—and the batch invariance solutions that achieve 100% reproducibility.
Read log
JUL · 19 · 2023 5 min read
RESEARCH & ENGINEERING
Instruction fine-tuning Llama 2 with PEFT's QLoRa method
In this blog post, we will discuss how to fine-tune Llama 2 7B pre-trained model using the PEFT library and QLoRa method. We'll use a custom instructional dataset to build a sentiment analysis model.
Read log