I came to software from insurance claims operations. That is where the bias in everything below comes from: in claims, an unverified assumption has a financial consequence. So I build systems that are deterministic, traceable and auditable.

One rule runs through all of it:

VERIFIED ≠ DECLARED

A task passes only with reproducible proof.

Moving into software full time after building it in parallel since 2021, alongside a full-time claims role. Open to backend, platform and agentic engineering roles — remote, contractor (B2B / Deel) or EOR.

Live badges — the numbers below come from npm, not from me.

Package	Version	Downloads
@hardmachinelabs/index-ai-validator			Validates whether a site correctly exposes an AI-readable layer: /.well-known/index-ai.json, /agent-index.json, clean Markdown endpoints, declared content size, exposure risks.
@hardmachinelabs/env-sync			Syncs .env variables into GitHub Actions secrets and GitLab CI/CD variables. Monorepo-aware. Built from watching local, CI and provider secrets drift apart.
@hardmachinelabs/zod-config			Runtime config validation for TypeScript and NestJS. Zod schemas as source of truth, fail fast at startup, sensitive values redacted in error output.

All MIT. All built from a problem I actually hit.

Status is stated on every line. Some of this is public and inspectable; some is not yet. I would rather tell you which than let you assume.

in development · not yet public

Built on one rule: agent output stays untrusted until deterministic gates verify it.

Designed from a catalogued taxonomy of 40 real coding-agent failure modes — context rot, runaway token spend, compaction loops, unverified commits. It keeps specifications, decisions, jobs, evidence and handoffs outside the chat, separates CTO / Coder / Tester / Reviewer roles, loads only the context each task requires, and reports through an HTML cockpit.

Bounded context routing — in a 500-file reference repository, active context narrows to 8 explicitly routed files. Sealed module contracts, mechanical quality gates, SHA-256 artifact sealing, append-only decision ledgers. Traced with Langfuse and OpenTelemetry, evaluated against Harbor / Terminal-Bench.

I use it daily. The human provides intent and judgment; the system carries the process.

open proposal · reference implementation · public

Websites are built for humans; agents read differently. Instead of forcing them to scrape noisy HTML, index-ai gives them a clean public layer the publisher controls:

The goal is not to promise discovery or citations. It is to make agent-facing content publisher-controlled, testable, measurable and verifiable. index-ai-validator above is the reference implementation.

in development · private

Backend of a platform turning water-damage claim declarations into step-by-step handling directives. I chose a deterministic rules engine over RAG so every output traces back to a named clause of the Convention d'Indemnisation et de Recours des Sinistres Immeubles. NestJS, Mastra AI, SSE streaming, subscription billing with usage metering, SQL analytics.

Comes straight out of the job I still do by hand.

in development · private

Determines whether a severe weather event occurred at a given address on a given date, and returns an adjuster-ready evidence report in under 60 seconds. A Python filtering layer over raw meteorological telemetry, then a NestJS engine applying contractual thresholds. Weekly batch ingestion from Google Drive into Elasticsearch through a three-stage BullMQ/Redis pipeline with Redlock locking and SHA-256 deduplication.

Core        TypeScript · JavaScript · Node.js (v20+) · NestJS · Python
            REST · GraphQL · microservices · event-driven architecture

Data        PostgreSQL · TypeORM · Prisma · Redis + Redlock
            Elasticsearch · Kafka · BullMQ · SQLite

AI/Agents   Mastra AI · MCP (tools & servers) · Anthropic, OpenAI, Google AI
            Langfuse · RAG · Harbor / Terminal-Bench

Delivery    Git · GitHub Actions · GitLab CI/CD · Docker · Dokploy
            Nginx · hardened Linux VPS · Cloudflare R2 · OpenTelemetry

Testing     Vitest · integration & E2E · Zod validation · deterministic CI gates

• I Built a CLI to Stop Missing Env Vars from Breaking Deployments — dev.to

Map the workflow end to end. Expose the failure points and the missing capabilities. Then build the software required to close the loop.

I keep decision records, failure analysis, test evidence and release criteria, so every claim I make can be inspected.

The goal is not to present experiments as traction. The goal is to show shipped work, clear engineering judgment, and honest status.