Skip to main content
petruarakiss
indexwritingaboutprojects
handoff

End of page. Start of the working loop.

If this site did its job, the next step is simple: download the public CV, read the field notes, or check the public proof. I am looking for AI product work where the model is only one part of the system.

Download public CVRead the notes
available for
Staff / Principal AI engineering roles
best work
harnesses, retrieval, evals, workflow control
base
Madrid / remote-first
paths
writingprojectsabout
contact
emailLinkedInGitHubX
context before autonomyevals before confidenceproof beats claimscode must stay legiblesystems over demospublic CV availablecontext before autonomyevals before confidenceproof beats claimscode must stay legiblesystems over demospublic CV available
(c) 2026 Petru Arakisspublic surface for product engineering work
projects

Proof, not claims.

A compact inventory of systems I can talk about publicly: document intelligence, agent runtimes, internal AI assistants, and open source work.

proof inventory
current proof
3
public repos
3
surface
ai systems
base
madrid

production systems

public summary · private internals
state
system
since
type
evidence
●current
BIFROST
Document intelligence layer for messy real-world inputs: ingestion quality gates, semantic chunking, multimodal retrieval, pgvector/HNSW search, caching, source-quality summaries, analytics, and honest no-answer behavior.
2024
production
Document intelligence layer for messy real-world inputs: ingestion quality gates, semantic chunking, multimodal retrieval, pgvector/HNSW search, caching, source-quality summaries, analytics, and honest no-answer behavior.
pythonfastapipostgresqlpgvectordoclingpytorchtransformers
●current
ORVIAN
AI workflow runtime with protected multi-tenant APIs, context assembly, durable memory, deterministic/cached/full-LLM execution tiers, run events, idempotency, queue processing, and human-review metadata.
2024
production
AI workflow runtime with protected multi-tenant APIs, context assembly, durable memory, deterministic/cached/full-LLM execution tiers, run events, idempotency, queue processing, and human-review metadata.
typescripthonopostgresqldrizzlesupabasequeuesscheduled-jobs
●current
Polaris
Internal AI assistant product integrating BIFROST retrieval with MONARCH guardrails, cached safety-to-retrieval handoff, citations, streaming UX, analytics, and suggestion revalidation.
2024
production
Internal AI assistant product integrating BIFROST retrieval with MONARCH guardrails, cached safety-to-retrieval handoff, citations, streaming UX, analytics, and suggestion revalidation.
next.jsvercel-ai-sdkbifrostmonarchdrizzle-ormpostgresql

open source

github.com/Arakiss →
state
system
since
type
evidence
◐public
gommage
Policy-as-code permission harness for AI coding agents: deterministic rules instead of prompt-only safety.
—
rust
Policy-as-code permission harness for AI coding agents: deterministic rules instead of prompt-only safety.
ai-agentspolicy-as-codeguardrailspermissionsharness-engineering
◐public
traceframe
Local-first trace recorder and inspector for AI agent workflows: append-only evidence, hook ingestion, ledger indexing, and CI-verified reports.
—
rust
Local-first trace recorder and inspector for AI agent workflows: append-only evidence, hook ingestion, ledger indexing, and CI-verified reports.
ai-agentsagent-tracingharness-engineeringlocal-firstdeveloper-tools
◐public
vestig
Runtime-agnostic structured logging with context propagation, observability primitives, and automatic PII sanitization.
—
typescript
Runtime-agnostic structured logging with context propagation, observability primitives, and automatic PII sanitization.
loggingobservabilitypiigdprtypescript