Skip to content
vollko
Main
Homepage Engineering Transformation Whitepaper OSS catalog
Sections
Domains What we build Open source About Engage
Deep dives
knowledge-graphs eval-harness agent-memory protocols agent-identity observability orchestration sensing-ingestion governance agent-frameworks feedback-loops end-to-end Start a conversation

We ship AI agents in production for energy and critical infrastructure operators.

A small senior team behind NUIT and 40+ OSS packages across TypeScript, Python, Rust. 75 MW of solar under daily AI forecasting. Bucharest, global.

Tell us what you need to ship
01 · Where we work

Two domains. One posture.

Reliability, audit, and uptime matter more here than novelty.

Agent infrastructure

The plumbing nobody talks about. Agent identity (DIDs, no blockchain), signed tool-call receipts, deterministic transcripts, content addressing for AI artifacts, agent payments. Twelve packages in agent-*, conformance-tested.

Stack signals: TypeScript, Bun, Python, Ed25519, multiformats, MCP, FastAPI

Energy market AI

Price + capacity forecasting (DAM, JAO cross-border, forward curves). Per-asset solar production. Consumption modeling across 8 countries. ETRM / EPRM / hedging engines on real renewable books.

Stack signals: Python, Rust, CatBoost, pvlib, ENTSO-E, OPCOM, JAO, Polars, SeaORM
02 · What we build

Twelve services. Four databases. One agentic loop.

Same shape, every engagement.

02a · the stack

Four layers, stacked.

twelve boxes, one shape
SURFACE BROWSER UI API GATEWAY SENSORS events · pubsub AGENTIC CORE LIVE AGENT RUNTIME rust · persistent memory review → hypothesis → investigate → verify → synthesize TOOL SURFACE mcp · meta-tool 2 visible · 67 handlers EVAL HARNESS per-tool telemetry versioned flat-prompts DATA LAYER POSTGRES TRUTH PG + PGVECTOR STATE GRAPH DB PATHS SEARCH ENGINE LOGS INTELLIGENCE LLM ROUTER policy · retry · fallback SELF-HOSTED LLM vllm / llama.cpp CLOUD FALLBACK anthropic · openai · gemini KNOWLEDGE GRAPH failure-open
One ember box: the agentic loop.
02b · the cube

Six faces. One engagement.

AGENTIC CORE rust runtime HARNESS INTELLIGENCE FACE 01 · CORE rust runtime · persistent memory FACE 02 · HARNESS flat-prompts · sse · telemetry FACE 03 · INTELLIGENCE router · self-hosted · cloud · KG FACE 04 · DATA postgres · pgvector · graph · search FACE 05 · SURFACE ui · api · background loops FACE 06 · SENSORS events · pubsub · debounce
One engagement, viewed from any side.
02c · the loop

Five phases. One gate.

01 · REVIEW 02 · HYPOTHESIS 03 · INVESTIGATE ▼ the gate 03.5 · VERIFY catches hallucinated names 04 · SYNTHESIZE next iteration
Verify catches what hypothesis missed.
02d · the tiers

Three tiers. Cheap path first.

QUERY /api/v1/chat TIER 1 · REGEX 0–5 ms · zero tokens TIER 2 · AGENTIC LOOP 2–30 s · 8k–60k tokens TIER 3 · CLOUD FALLBACK on demand · warm standby
Cheap first. Smart on demand. Standby always.
02e · the seam

Failure-open by design.

AGENT RUNTIME try / except BREAKER 30s base · 5s per-call KNOWLEDGE GRAPH domain corpus + bge-m3 try ✓   enriched fail ✗   empty → today's behavior
Try → enriched. Fail → empty. Never → page anyone.
02f · the pipeline

Three lanes.

USER UI API ROUTER AGENT MCP LLM TIER 1 · 0–5 MS · ZERO LLM POST /chat forward route_query {intent, extracted} execute_routed {result, summary} TIER 2 · 2–30 S · 8K–60K TOKENS POST /jobs inference tool_calls LOOP · TOOL CALLS × N execute(tool, params) results continue next response webhook FALLBACK · CIRCUIT-BREAKER messages API → narrative stream rendered
Top to bottom: cheap → smart → fallback.
02g · the fan-out

Two. Sixty-seven.

DISCOVER( ) EXECUTE( ) queries 5 handlers analytics 5 handlers entities 5 handlers audit 5 handlers compliance 5 handlers context 3 handlers fuzzy 5 handlers investigate 4 handlers knowledge 5 handlers graph 5 handlers ingest 5 handlers routine 7 handlers patterns 8 handlers ~95% schema overhead reduction
Sixty-seven handlers, hidden behind two.
02h · the event lane

Six in. One out.

SENSORS events · pubsub DEBOUNCE 60-min · per entity deduplicates AGENT ad-hoc job 6 in 1 out
Six in, one out. Pages no one.

Below: in practice.

Active right now

40+
OSS packages (npm, PyPI, crates)
75 MW
under daily AI forecasting
04 · Open source

Forty plus packages. Public proof.

If a team claims to ship code, the code should be public. Two suites - the agent-* protocol family (12 primitives × TS, Py, Rs) and a 32-package zero-dependency utility library - plus standalone tools: triz-ci, pvlib-rust, Kindred, dsl.

agent-* suite

Twelve protocols, three languages, layer by layer

TypeScript, Python, Rust. Byte-deterministic cross-language conformance via JCS, Ed25519, CIDv1, Noise-XK, DSSE.

npm scope: @p-vbordei
Identity
agent-id
v0.1.5 · 67 tests
Self-custody DID + Capability VCs for AI agents. did:key, did:web. Three functions, zero blockchain. Imported by 7 sibling packages - keystone of the suite.
Artifacts
agent-cid · agent-scroll · agent-rerun · agent-toolprint
Content-addressed manifests (agent-cid). Byte-deterministic hash-chained transcripts (agent-scroll, two implementations produce identical bytes). Portable reproducibility bundles (agent-rerun, SLSA for agent steps). Double-signed tool-call receipts (agent-toolprint).
Transport
agent-phone · agent-ask · agent-rooms (Parley)
Sync agent to agent RPC over WebSocket with Noise handshake (agent-phone). Federated public Q&A protocol (agent-ask). Multi turn cross org rooms with bounded conversation (agent-rooms / Parley, the only Python product in the suite, three packages on PyPI).
Commerce
agent-pay
Reference stack composing L402 (HTTP 402 plus Lightning) with DID signed invoices for agent to agent payments. Server middleware, client lib, retry-with-macaroon flow. Demoable on a Polar Lightning regtest with one docker-compose command.
Operations
agent-fleet · agent-publish · agent-launch
Cron driven multi repo health bot that opens weekly review issues across the family (agent-fleet). One shot multi registry release publisher with OIDC Trusted Publishers and SLSA provenance (agent-publish). Drafts platform native release announcements from CHANGELOG and commits (agent-launch). The maintenance loop that maintains the suite, eaten as our own dog food.
All packages: spec first, conformance vectors green, release-please plus Renovate plus Claude PR review on every repo.
utility suite

Thirty two zero dependency TypeScript packages

The small, sharp tools the team reaches for on every project. Apache-2.0, dual ESM and CJS builds, property tested with fast-check, coverage gated.

AI safety
  • prompt-shield
  • pii-detect
  • secret-sniff
  • tool-call-validator
  • llm-stream-parser
  • token-est
  • markdown-streaming
Resilience
  • circuit-breaker
  • async-mutex
  • pretry
  • pmap-bounded
  • pqueue-tiny
  • token-bucket
  • cancellable
  • ws-reconnect
  • typed-event-bus
Time and cron
  • cron-describe
  • cron-next
  • nl-cron
  • business-days
  • human-duration
  • relative-time
  • tz-clock
Web and util
  • etag-mini
  • content-negotiation
  • cookies-tiny
  • hmac-sign
  • dotenv-tiny
  • lru-tiny
  • parse-args
  • slug · tiny-validator
triz-ci
Python · MCP · Plugin

TRIZ engineering co-pilot

The Theory of Inventive Problem Solving turned into an MCP server, a Claude Code plugin, a Gemini CLI extension, and a skill bundle. RAG over a books corpus, structured workflows, three runtime surfaces. v1.0.0.

pvlib-rust
Rust · Crate

Solar physics for the Rust ecosystem

A Rust port of pvlib for production grade solar modeling. Powers the inference layer of our Rust solar forecasting service. Released so other engineers building energy systems in Rust do not have to wrap Python.

Kindred
Python · Next.js · Live

Shared notebook for AI teams

The shared notebook every teammate's AI reads. Write once, retrieve with provenance. Pages decay if untouched. FastAPI backend + Next.js web + kin CLI + Claude Code plugin (MCP + skill + hook). KAF 0.1 spec.

Full repos at github.com/p-vbordei
05 · About

A small senior team.

Vollko is small and senior throughout. The people you talk to are the same engineers who design the architecture, write the code, and stay on call when it runs in production.

The pattern across all of it is the same. We work on systems where the cost of a wrong answer is real. Power markets. Critical infrastructure. Operations that have to run at 03:00. AI is the current toolchain. The durable skill is shipping change inside complex systems.

GitHub · @p-vbordei vlad@vollko.com
40+
OSS packages across npm, PyPI, crates
75 MW
solar PV under daily AI forecasting
22
panels in the visual whitepaper
12
deep dives across the agent stack
// Vollko on the main stage at Big Data Week 2025: "Mind the Gap: AI in Energy".
06 · Get in touch

Tell us what you need to ship.

Send the brief, the chaos, or the scribbled napkin. We respond within two business days with a short read on fit and next steps. If we are a fit, the call is the same week.

Book a call on Cal.com vlad@vollko.com
Bucharest, RO. Working remotely with teams worldwide.