Projects
Selected work, most recent first.
Channels · P-1 AI
A platform for an AI agent to be present in the tools engineers already use — Slack, Teams, Linear, Planner, and email — rather than only its own UI.
Every surface is an adapter behind one contract, so adding one means writing an adapter, not changing the core. A message is persisted before it is acknowledged, the core drives an agent session, and the answer is rendered back in the surface's own idiom: Block Kit, Adaptive Cards, Linear activities, email threads. Users answer approvals and questions in place, and every action is attributed to a real user with their own permissions.
Capabilities · P-1 AI
The supply chain that lets an AI engineering agent learn a new domain: an authoring SDK, a catalog, an image factory, and a credential gateway.
A capability bundles a Python library, the skills that tell the model how to use it, and its tests. An SDK and CLI define the authoring contract and scaffold, validate, and publish. The catalog is immutable records over content-addressed blobs in a bucket, run as triggerable jobs with no resident server — the bucket layout is the whole API. An image factory resolves an agent's capability set to exact digests and builds one immutable image, so the sandbox installs nothing at runtime. The zero-trust gateway below brokers every credential that code touches.
Dataset Pipeline · P-1 AI
Packages raw data into content-addressed OCI artifacts through pluggable sources, transforms, and sinks.
Hugging Face is one of the pluggable sources. Runs are orchestrated on Kubernetes by a custom resource definition and its controller.
Archie BYOC Platform · P-1 AI
Deploys an AI engineering agent into customer-controlled cloud; each install assembles from independently versioned components.
P-1's engineering agent runs inside customer cloud environments rather than ours. Each install is assembled from independently versioned components, so deployments can differ per customer without forking the platform.
Zero-Trust Agent Gateway · P-1 AI
A phantom-token pattern — agent code in ephemeral sandboxes holds only opaque, request-scoped tokens; the gateway keeps every real credential.
Agent-written code runs in ephemeral sandboxes and never holds real credentials. Sandboxes get opaque, request-scoped tokens; the gateway holds the real ones and enforces authentication, authorization, scoping, and rate limits.
Observability Foundation · P-1 AI
OpenTelemetry instrumentation across every service, feeding a self-managed Grafana LGTM stack.
Logs, metrics, and traces from every service live in one place, with alerting and on-call tooling.
Duplicate-Invoice Detection · Amazon
Real-time duplicate-invoice detection (up to 500 TPS) with pluggable rule and ML engines, part of Amazon's fraud-prevention suite.
Real-time detection service handling bursts up to 500 TPS, with rule-based and ML engines behind a common interface so new detectors ship without touching the pipeline. Runs on Lambda, SageMaker, and OpenSearch.
EU Compliance Ingestion · Amazon
Processes roughly one billion book titles per year, extracting required metadata and storing compliance artifacts at 500 TPS.
Pipeline on Fargate, SNS/SQS, and Lambda processing roughly a billion book titles a year — extracting required metadata and storing compliance artifacts at up to 500 TPS.
Transaction Tagging Engine · Ampla
Re-implemented with Aho–Corasick; runtime went from hours to 90 seconds.
Re-implemented Ampla's transaction tagging around Aho–Corasick string matching, taking runtime from hours to 90 seconds.