Konstantinos Zervakis zervakisai

I turn hard problems into shipped solutions

`>_ whoami`

AI Engineer building production ML, LLMs and multi-agent systems for environments where a wrong answer costs the vintage, a wildfire response, or a satellite mission.

Most people ship demos. I ship systems people depend on. A benchmark impresses for a minute; a solved problem holds up when a missed defect ends a space mission. So I don't lead with model names — I lead with the constraint that mattered, the design that removed it, and the number that proves it worked.


Now	Software Test Engineer @ OHB Hellas — aerospace, ECSS V&V/PA
PhD	Autonomous UAV navigation — Hellenic Army Academy
Published	Sensors (MDPI) 2026 — lead author
Base	Crete, Greece — working worldwide

5 production deployments · 0 hallucinated numbers shipped · 71× faster InSAR unwrapping · 450+ experimental runs

🗂 Case Files

Each one, the same shape: problem → solution → result — with a number attached. Full write-ups at zervakisai.github.io.

★ Vinea — Agronomic decision agents for a vineyard

Flagship agentic project · Pydantic AI pydantic-graph FAO-56 output validators uv 12 tagged phases

flowchart TB
    W(["🌦️ hourly weather · Open-Meteo or committed CSVs"])

    subgraph PY["⚙️ &nbsp; PLAIN PYTHON &nbsp;·&nbsp; every number is computed here"]
        F["<b>FeatureBuilder</b> &nbsp;—&nbsp; <i>no model in the call path</i><br/>FAO-56 water balance · ETc = ET₀ × Kc, clamped to [0, TAW]<br/>Delta-T &amp; wind bands → the day's spray windows"]
    end

    subgraph AI["🧠 &nbsp; PYDANTIC AI GRAPH &nbsp;·&nbsp; judges and explains — never calculates"]
        direction LR
        I["<b>Irrigation</b><br/>irrigate, or hold?"]
        S["<b>Spray</b><br/>which window?"]
        C["<b>Coordinator</b><br/>sequences the day"]
    end

    V{"<b>Output<br/>validators</b>"}
    R["depletion echoed <b>verbatim</b><br/>windows ⊆ deterministic candidates<br/>confidence ≤ its weakest leg<br/><i>an oracle recomputes it nightly</i>"]
    OUT(["📋 <b>tomorrow's advisory</b><br/>in the grower's hands by 06:00"])
    BLK["🚫 <b>blocked</b><br/><i>never ships</i>"]
    INJ["☠️ prompt<br/>injection"]

    W --> F
    F ==>|"finished numbers, typed contracts"| I
    I --> S
    S --> C
    C ==> V
    V -.- R
    V ==>|"holds to the numbers"| OUT
    V -->|"a quantity the model invented"| BLK
    INJ -.->|"reaches the words…"| AI
    INJ -.-x|"…never the arithmetic"| PY

    classDef det fill:#0d2a1a,stroke:#3fb950,stroke-width:2px,color:#e6edf3
    classDef llm fill:#10233d,stroke:#58a6ff,stroke-width:2px,color:#e6edf3
    classDef gate fill:#3d2f00,stroke:#d29922,stroke-width:2px,color:#f0d58c
    classDef rules fill:#161b22,stroke:#d29922,stroke-width:1px,color:#c9d1d9
    classDef good fill:#0d2a1a,stroke:#3fb950,stroke-width:3px,color:#7ee2a8
    classDef bad fill:#3d1418,stroke:#f85149,stroke-width:2px,color:#ffa198
    classDef src fill:#161b22,stroke:#8b949e,stroke-width:1px,color:#c9d1d9

    class F det
    class I,S,C llm
    class V gate
    class R rules
    class OUT good
    class BLK,INJ bad
    class W src

    style PY fill:#07160f,stroke:#3fb950,stroke-width:2px,color:#7ee2a8
    style AI fill:#0a1626,stroke:#58a6ff,stroke-width:2px,color:#79c0ff

    linkStyle 1 stroke:#3fb950,stroke-width:3px
    linkStyle 5 stroke:#d29922,stroke-width:1px
    linkStyle 6 stroke:#3fb950,stroke-width:3px
    linkStyle 7 stroke:#f85149,stroke-width:1.5px
    linkStyle 8 stroke:#f85149,stroke-width:1.5px
    linkStyle 9 stroke:#f85149,stroke-width:2.5px

_{Built in 12 tagged phases — persistence, a Postgres queue, FastAPI, Streamlit, a prompt registry, an eval gate, an LLM gateway, RAG citations, row-level security, SLOs and Kubernetes all landed on top of that core without reaching into it. tests/test_core_unchanged.py compares the parsed AST of those six files against the day they were written.}


Problem	"Should I irrigate? Can I spray?" A grower's daily calls hide real cost — a wrong one wastes scarce water, or sprays into the wrong window and risks the vintage. A raw LLM will confidently invent the agronomy.
Solution	One design constraint: every number is computed in plain Python — the model judges and explains, it never calculates. A FAO-56 water balance and spray-window analysis run nightly per block; a small pydantic-graph of typed agents (Irrigation, Spray, Coordinator) judges the borderline calls and sequences the day. Output validators hold the agents to the numbers: depletion echoed verbatim, spray windows a subset of the deterministic candidates, confidence capped by the weakest leg.
Result	An advisory cannot ship a hallucinated quantity — validation blocks it and a nightly eval oracle recomputes it; the same boundary stops prompt injection. Built in 12 tagged phases from CLI to Kubernetes without touching the AST-frozen core — and `uv run vinea` always produces output, even with zero API keys.

_{12 phases · 0 hallucinated numbers}

📡 SAR Unwrap

Training-free InSAR phase unwrapping on a consumer GPU

_{InSAR GPU weighted IRLS integer LP training-free}

Problem → Solution → Result

P — Phase unwrapping is the long pole of InSAR: SNAPHU is accurate but takes minutes per scene with no GPU path, while deep-learning unwrappers degrade off-distribution — exactly where an autonomous disaster responder meets an unseen scene.

S — A hybrid that is fast, accurate and training-free: a GPU multi-scale warm-started weighted-IRLS solve (FFT-preconditioned conjugate gradient) recovers the fringe pattern in milliseconds, plus an exact region-contracted integer-LP leveling stage — the totally-unimodular dual of minimum-cost-flow — to remove residual offset drift.

R — 71× faster than SNAPHU on airborne UAVSAR at 99.5% per-component agreement — up to 130× on Sentinel-1 — and SNAPHU parity at low coherence (99.3% at 0.35) where the fast path-follower collapses to 70.3%. Single RTX 3060.

_{71× vs SNAPHU · 99.5% agreement}

🔥 FLARE Benchmark

Risk-parameterised UAV path planning under uncertainty

_{reinforcement learning benchmark risk coefficient ρ statistics Python}

Problem → Solution → Result

P — Safety-tuned planners paradoxically increase total mission cost by 23–38% when timing constraints exist — and nobody could audit the safety-vs-throughput trade-off.

S — A deterministic wildfire-response benchmark for risk-parameterised UAV path planning, with a configurable risk coefficient ρ that makes the trade-off explicit and auditable.

R — 450+ experimental runs · Friedman χ²(4) = 190.8 (p < 0.001), Wilcoxon (Bonferroni), Cliff's δ. Submitted to Sensors (MDPI) 2026 — lead author.

_{Sensors / MDPI 2026 · 450+ runs}

Also public — 🤖 AgenticRAG multi-agent orchestrator + MCP server · 🏭 SmartFactory-RAG industrial RAG + predictive maintenance · ⚛️ QBench quantum-classical ML benchmark suite

From the archive

Earlier case files — smaller scope, same shape.


Formula Student TUC	Forecast-first ECU pipeline predicting Lambda (λ) and RPM spikes with LSTM, ARIMAX and ETSformer — preventing lean spikes and over-rev failures before they become mechanical damage.	`<5 ms inference` `live telemetry`
LightningText	Real-time text analytics over 14M BookCorpus paragraphs on Apache Spark — reservoir sampling, Count-Min Sketch and wavelet compression instead of brute force.	`20 KB RAM` `<2% error` `8× leaner storage`
HAR-DRive	PCA vs LDA on UCI HAR: a 5-dimensional LDA with an SVM-RBF cuts features by 98% and still scores macro-F1 0.984 — activity recognition that fits a wearable.	`F1 0.984` `−98% features`

🛠 Stack

Agent framework of choice: Pydantic AI — typed contracts, graphs, one-layer retries, offline-testable.

Generative AI & Agents

RAG & Retrieval

Machine Learning

MLOps & Cloud

Quantum ML

Languages

📊 GitHub Stats

🎓 Research & Education


🎓 PhD(c)	Autonomous Navigation of UAVs — Hellenic Army Academy
📜 MSc	Machine Learning & Data Science — Technical University of Crete
📐 BSc	Mathematics & Applied Mathematics (241 ECTS) — University of Crete
📄 FLARE	Decision optimization under uncertainty — Sensors (MDPI) 2026, lead author
🧠 AI4Edu	AI for personalized learning — co-authored, deployed platform

⚡ Beyond the Code

🤺 Discipline
2× Silver Medalist — National & Global Fencing

_{Épée teaches what shipping does: read the opening, commit fully, and own the outcome in real time.}

🎙️ SKAI Crete FM
Live national radio on AI & autonomous systems

_{Featured guest on the trajectory of AI, autonomous systems, and emerging aerospace career paths.}

🎧 Long-form podcast
Translating agentic AI for non-specialists

_{Turned complex agentic-AI concepts and current industry shifts into something a general professional audience could actually use.}

✍️ Medium essay
"From the Bronze Age to the Periclean Era"

_{Argues the executive KPI of AI is shifting from productivity to orchestration integrity — introducing FLARE's ρ as a managerial tuning knob.}

💬 Got a problem worth solving?

The good ones don't wait.

_{Based in Crete, Greece · Working worldwide · Open to multi-agent systems and production ML engagements}

Navigation Menu