01
SECOND BRAIN
// the memory
Every decision, document, and learning a company produces — structured, versioned, and queryable by agents as a living, LLM-maintained wiki. The pattern I published on.
FOUNDER, NIGHT ENGINES — ATHENS → LONDON
BUILDING THE AGENT OS FOR BUSINESS.
A second brain and mission control for companies — agents with full memory of your business, executing real work. Not a thesis: I ran an AI lab where my multi-agent systems processed 2,000+ real insurance claims in production.
01 READOUT — PRODUCTION TELEMETRY
0
PRODUCTION CLAIMS PROCESSED
0→0
PER CLAIM, AUTOMATED
0
F1 SCORE IN PRODUCTION
0
NATIONAL HACKATHON WINS
0
PATENT FILED — MEDICAL DEVICE
0
PUBLISHED PAPER — ACM CAIS 2026
02 WHAT I'M BUILDING NOW — MY COMPANY
NIGHT ENGINES.
The company I founded — an AI agency. We give a business its own second brain, then put private agents to work on it: agents with full memory of the company, running approved playbooks on your real tools while you keep the judgment. Built on the same stack I already run in production. The three pieces:
02
MISSION CONTROL
// the hands
A control plane where agents with full business context execute real work — not chat, operations. Runs on Hermes / OpenClaw / Claude Code with skills, crons and MCP tools.
03
THE TASTE LOOP
// the judgment
Agents run the plan asynchronously; my scarce input is evaluation, not typing. Out of the prompting loop, in charge of the call on what actually ships.
03 TRACK RECORD — EVERY PROJECT WAS A BUSINESS
HELLAS DIRECT
The founders scouted me to build their AI lab.
After the IBM win, Hellas Direct's founders asked me to run AI at an insurer with 500k+ customers. Owned the whole loop — architecture, code, deployment, then product ownership.
★ DIGITAL FINANCE AWARDS 2026 — WINNER
IN PRODUCTION TODAY
CLAIMS AUTOMATION — 7-AGENT PIPELINE LIVE
INTAKESECURITYTRANSLATIONINTENTEXTRACTIONNOTESAUDIT
2,000+ REAL CLAIMS PROCESSED
25 MIN → 5 MIN PER CLAIM
99% F1
MCP SERVER
The claims product exposed as a tool layer for any LLM. Their single biggest competitive advantage — made callable.
MCP / FASTAPI / PYTHON
CUSTOM OCR ENGINE
A classifier identifies the document, a fine-tuned extractor per type pulls structured data. Any document the business receives.
GCP DOCUMENT AI / FINE-TUNING
WHATSAPP RAG ASSISTANT
Customer-facing assistant on the channel Greeks actually use. Separate system, also in production.
AWS BEDROCK / LAMBDA / DYNAMODB
PYTHON / PYDANTIC AI / GPT-4O / GCP / KUBERNETES — ran in parallel with a full-time consulting role; both delivered.
MEDIVA-RX
From a basement to Columbia University.
Started as a hackathon hack, pushed as far as a student venture can go: national win, a filed patent, a real market pivot, a finalist seat at Columbia.
NATIONAL WINNER
PATENT FILED
COLUMBIA FINALIST 2026
- 2023THE BUILD — Smart pill dispenser with Face ID. TensorFlow and OpenCV on real hardware, built in a basement in a month for a hackathon deadline.
- 2023THE WIN — 1st place, PreSfhmmy national competition. The hack became a venture.
- 2024THE PATENT — Filed. Protected the core mechanism before talking to the market.
- 2024–25THE DISCOVERY — 70+ customer interviews (pharma, elder care). The consumer product everyone liked wasn't the business.
- 2025THE PIVOT — Consumer gadget → Type 2a medical device for decentralized clinical trials — running stages I and II at home. A market that pays.
- 2026COLUMBIA — Lab 2 Market finalist, NTUA × Columbia. Presented in New York, June 2026.
TENSORFLOW / OPENCV / REACT / DJANGO — the tech served the business, never the other way around.
UNIAI × IBM
A weekend hack that became a job offer from founders.
The clearest version of how I operate: ship something real fast, then let the room escalate it.
1ST PLACE — IBM CHALLENGE
IBM SUMMIT STAGE
- 48 HRSMAY 2024 — ML taxi allocation for Chicago O'Hare. A model predicting aircraft delays by type, airline, time, destination — driving dynamic taxi dispatch.
- THE WIN1st place. IBM's Chief AI Officer invited the team to present.
- THE STAGEIBM Greece & Cyprus quarterly summit — CEO Marios Maniatis handed me the mic in front of the whole company.
- THE JOBScouted to run an AI lab. Hellas Direct's founders saw it; the demo was the pitch; the pitch became the job.
PYTHON / LOGISTIC REGRESSION / IBM CLOUD — won as a student, hired as an architect.
THE LIVING WIKI
I published the paper on the memory my agents run on.
“Schema-Driven LLM Knowledge Bases as Persistent Agent Memory.” An empirical study of the LLM-maintained wiki as agent memory: where RAG is stateless and re-derives knowledge every query, the wiki synthesizes once and compounds — and wins clearly on entity aggregation, the exact thing agents need to know people and projects. The theory the Second Brain above is built on.
★ ACCEPTED POSTER — ACM CAIS 2026
AGENTSKILLS WORKSHOP · SAN JOSE
PYTHON / HYBRID BM25 + DENSE RETRIEVAL / LATEX — read on OpenReview →
04 STACK — TOP TO BOTTOM
END TO END.
One person across the full depth of the stack — the interface people touch, the services behind it, the agents on top, and the infrastructure under it all. Not a specialist bolted onto a team; the engineer who takes a product from pixel to production.
- FRONT OF HOUSE
- React · TypeScript · interface, motion & WebGL — the product people touch
- BACK OF HOUSE
- Python · FastAPI · Java · SQL · distributed services — the systems behind it
- THE AGENT LAYER
- Claude Code · Hermes · OpenClaw · MCP · skills + crons
- MEMORY
- RAG pipelines · LLM-maintained wikis · embeddings + graph
- THE GROUND
- AWS · GCP · Azure · Docker · Kubernetes · Terraform
05 SIGNAL — OPEN CHANNEL