Lead Applied AI Engineer · ZanderLabs · Berlin

I close the loop between brains, robots, and language models.

Swiss Army knife engineer. Built for zero to one.

T-shaped: breadth across embodied AI, agentic development, and enterprise XR; depth into whichever one the problem demands. Drawn to greenfield projects and cold-starts into new tools. The proof: a robot arm that corrects itself mid-task from a person's brain signals, with no retraining.

See the work Download CV [PDF] or press ⌘K

NAT 2026 FIRST AUTHOR · 4 PUBLICATIONS · 2 PATENT FILINGS · B.SC. MECHATRONICS (1.7)

the NAT 2026 system, abridged · hand-coded SVG, no library SESSION 00:00

< 18h EEG paradigm rollout, down from 3-5 weeks

4 EU labs running my experiment platform

15Hz closed-loop robot control from neural feedback

€300k GPU cluster I specced behind a state grant

4/1 publications, one as first author

3+ continents running my XR platform

CH.02

Selected work

SYSTEMS OWNED END TO END

CH.02.1 NEUROADAPTIVE ROBOTICS SHIPPED · CONTINUES AT ZANDERLABS

NAT 2026 · first author HAL + GCRM patents NVIDIA Inception

A robot that takes correction from your brain, mid-task, with no retraining

I trained a VLA policy in simulation, transferred it to a real arm, then wired a person's brain into the control loop. When the robot is about to get something wrong, EEG-derived signals become natural-language hints that nudge it back on track while it is still moving. The model never sees a gradient update.

Ran a pi0.5 DROID VLA policy on a Franka Panda in NVIDIA Isaac Sim, then went sim-to-real onto a physical LeRobot SO-101 arm, with VLA data collection and fine-tuning on real hardware.
Designed the tri-node Actor / Reasoner / Interpreter architecture: Cosmos-Reason1 grounds the scene; EEG-derived workload, agreement, error, and surprise signals drive the correction.
Held a 15 Hz control and 10 Hz reasoning loop, stable enough to demo live to investors. Seeds two patent filings: HAL and GCRM.
First-author paper at NAT 2026; presenting author of the companion poster at IK 2026.

isaac sim · pi0.5 droid · franka panda · lerobot so-101 · cosmos-reason1 · eeg / lsl · zeromq · sim-to-real

The real closed-loop dashboard from the NAT 2026 paper: camera feeds, robot reasoning log, and four live neural-state traces — the actual readout: reasoning log + 4 neural channels · NAT 2026

System architecture figure from the NAT 2026 paper — tri-node architecture · actor / reasoner / interpreter

CH.02.2 LINKEDCHAT IN PRODUCTION · 4 EU LABS

3-5 weeks → under 18 hours Built solo Feeds 2 papers

Cut EEG paradigm rollout from weeks to hours, now running in four EU labs

EEG experiment setup was the bottleneck for every study the lab wanted to run. I built the full-stack platform end to end, solo: experiment orchestration, millisecond stimulus timing, data capture, and the researcher-facing tooling. A new paradigm went from a multi-week engineering project to an afternoon of configuration.

Paradigm rollout time dropped from 3-5 weeks to under 18 hours.
Adopted as the backbone for spoken-dialogue BCI research across four European labs, so results are comparable instead of bespoke per site.
FastAPI and WebSocket backend, React frontend, LSL markers for millisecond stimulus-onset timing, Docker on GCP, Playwright end-to-end tests.
Feeds the data behind the ACM CUI 2026 paper and the arXiv 2026 workload-decoding paper below.

fastapi · websockets · react · lsl · docker · gcp · playwright · built with an agentic harness

PARADIGM PIPELINE · FIG. 04

one platform, four sites, same protocol

Workload decoding results: mean workload rising across contest rounds with confidence intervals — workload decoded from EEG during LLM dialogue · arXiv 2026

CH.02.3 ENTERPRISE XR DEPLOYED · EMEA + BEYOND

Led a dev team Discovery to rollout 3+ continents

Led a dev team to ship a multiplayer Quest 3 platform across a global medtech

At the Siemens Healthineers OpenIT Studio I led a team of developers and grew a 3D proof of concept into a production multiplayer Meta Quest 3 training environment with LLM-driven NPCs. Employees on different continents join the same shared session. Cybersecurity training was the first proof point.

Prototype to product to multi-region rollout, inside a regulated multinational.
Multiplayer via Netcode for GameObjects; LLM-powered NPCs for dynamic in-scene dialogue.
Alongside the XR work, shipped two production internal tools on Microsoft Power Platform, replacing manual multi-country Excel workflows.
Forward-deployed in practice: ran the discovery and design-thinking workshops with department stakeholders myself, then supported the multi-region rollout.
Grew the engagement from intern to bachelor thesis to working student.

unity · c# · netcode for gameobjects · meta quest 3 · llm npcs · power platform

SESSION DIAGRAM · FIG. 06

in-world imagery pending approval, diagram instead

CH.02.4 SOCIALTREE PRE-BETA · 188+ TESTS

188+ tests passing On-device RAG Solo build

An iOS app where the database is truth and the AI only proposes

SocialTree remembers the people in your life so you do not have to. Built solo in Swift 6 with an architecture I trust: the database is the system of record, the model proposes changes, and nothing the AI suggests is committed without passing the confirm gate. Every AI-generated fact carries provenance.

Database-as-truth, AI-proposes architecture: the model is helpful without ever being authoritative.
On-device RAG with 512-dim sentence embeddings: private, low-latency recall with zero cloud calls.
Retrieve-reason-propose anti-hallucination pattern and per-feature model routing.
188+ passing tests, because a memory app that loses memories is worthless.

swift 6 · swiftui · swiftdata · on-device embeddings · gemini · built with claude code + mcp

TRUST ARCHITECTURE · FIG. 07

nothing touches truth without your sign-off

CH.02.5 REDACTABLE OPEN SOURCE · APACHE-2.0

Open source MCP server + hooks Evals gated in CI

PII redaction as a proof problem, not a model problem

Redactable is my open-source de-identification engine. Structured identifiers are caught deterministically, with checksum proofs and zero hallucination risk; open-source models (GLiNER, Gemma) catch the contextual ones, names and places, and can run fully in the browser, so nothing leaves the machine; and every engine is scored on a labeled corpus with a CI regression gate. It sits between you and the model, so no provider ever sees real PII.

Deterministic-first detection: Luhn and MOD-97 checksums make catches provably correct. Perfect recall on emails, URLs, and IBANs against a third-party benchmark.
Reversible tokenization: PII leaves as [NAME_1], comes back restored. The model keeps the thread without ever holding the data.
Ships every way a harness can consume it: MCP server, transparent scrub proxy, Claude Code hooks, a GitHub Action, and a WebGPU build that runs entirely in the browser.
An eval harness with per-entity precision and recall, wired into CI: a recall regression cannot merge.

python · gliner · gemma · mcp · webgpu · hipaa safe-harbor policies · github.com/praneelbhatia/redactable

SCRUB PROXY · FIG. 08LIVE

live scrub session · open-source models, in-browser · 0% pii retention

APPENDIX

Components bin

SMALLER BUILDS, SAME STANDARD

Electronic chess board

Custom Hall-effect PCBs for analogue piece identification. Real boards, real sensors, and a real IEEE paper.

HARDWARE · PCB · IEEE ICHORA 2025

Realtime voice agents

Low-latency voice pipelines that listen, reason, and respond in the same breath. The hard part is the streaming latency budget.

VOICE · LLM · WEBRTC

Retro Gameboy PCB

Schematic in Eagle, board self-etched and milled, every SMD and through-hole joint hand-soldered.

HARDWARE · SOLDERING · GITHUB

Animatronic hand

OpenCV tracks your hand; a mechanical one mirrors it in real time. Presented at a university fair.

ROBOTICS · OPENCV

Automation backbone

Self-hosted n8n on GCP running a 14-person startup's internal ops: daily meeting digests adopted company-wide, one-click leave approvals.

N8N · GCP · OPS

AI grading assistant

Built while teaching C and MATLAB: an OpenAI-driven grading workflow with roughly 85 percent efficiency gain for the human in the loop.

LLM · TEACHING · TOOLING

Gaze-to-object pipeline

Syncs Tobii eye-tracking with screen recordings on LSL time, then labels which on-screen object the viewer fixates, frame by frame, via YOLO.

EYE TRACKING · YOLO · LSL · GITHUB

SharedPlaylist

A self-hostable agent that keeps one playlist bidirectionally in sync across Spotify and Apple Music: ISRC-first matching with a fuzzy fallback.

AGENT · TYPESCRIPT · OSS · GITHUB

Bloxwarz

Co-op multiplayer puzzle game: authoritative Colyseus server, Three.js isometric levels, and a BFS solver that proves every co-op level is beatable.

MULTIPLAYER · THREE.JS · GITHUB

CH.03

The harness

HOW THE WORK GETS MADE

I run an agentic engineering harness, not a chat window. Claude Code with 34+ custom skills, 20+ MCP servers, and sub-agents fanned out on isolated git worktrees, so exploration never pollutes the main context. Every change moves through the same four gates, and verification is the loop, not an afterthought. It is how I ship fast without shipping sloppy, and how I got a 14-person startup to work the same way.

GATE 01

Spec

Write the spec before any code. What the change is, what done looks like, what is explicitly out of scope. If I cannot write it down, it is not ready to build.

GATE 02

Plan

Turn the spec into a phased plan. Order the work, name the risks, decide what gets a sub-agent and what stays on the main thread. The plan is reviewable on its own.

GATE 03

Execute

Phased execution against the plan, with sub-agents on separate git worktrees so parallel work never collides. Deterministic checks decide when a phase is done: tests, type-checks, build exit codes.

GATE 04

Sign off

Two-stage review, then sign-off with test counts and commit hashes. A fresh agent grades the diff against the spec, so the agent that did the work is never the one grading it. Nothing merges on one pass.

The win was not that I worked faster. It was that everyone did.

I took Claude Code from a personal tool to the company standard at a 14-person startup. Inside about two weeks, non-engineers were shipping working prototypes and collecting live participant data on their own, work that previously queued for dev capacity. I also wrote the guardrails that made it safe: the agentic coding-tools policy, the AI code-generation rules, developer onboarding, and the information-classification framework.

The harness compounds. Every mistake an agent makes becomes a rule or a skill in versioned memory, so no agent makes the same mistake twice.

This site was built the same way: spec, plan, phased execution, two-stage review, sign-off. Hand-coded SVG, no framework. View source.

T-SHAPED

Passive BCI, VLA robot learning, and LLM systems are normally three different jobs. I do all three, and I make them talk to each other.

ZERO TO ONE

The first working version is my favourite thing to ship. I bring fresh ideas to every team I join.

RANGE

Dubai-raised, Germany-trained, Berlin-based. PCBs to robot policies to production LLM systems.

CH.04

Experience

ONE THROUGH-LINE: HUMANS IN THE LOOP

STILL RECORDING · BERLIN

05/2026 - NOW

Lead Applied AI Engineer ZanderLabs · Berlin

Leading LLM Paradigm 1, the neuroadaptive LLM line, at the €30M-funded neuroadaptive-technology company founded by Prof. Thorsten O. Zander, who founded the field of passive BCI. The Auryal IP and team transitioned in with me in 2026.

2025 - 04/2026

Founding AI Engineer Auryal GmbH · Berlin

Founding engineer at a 14-person SPRIND-funded startup, working directly with the CEO. Built the flagship neuroadaptive robotics system, specced the €300k GPU cluster behind a state grant, owned the in-room investor robot demo, authored company policy, and held three company OKRs. NVIDIA Inception member; co-author on an NVIDIA research proposal.

2024 - 2025

Intern → Thesis → Working Student Siemens Healthineers · Erlangen

Led a developer team on the multiplayer Quest 3 XR platform with LLM NPCs, deployed across EMEA and beyond. Shipped two production internal tools on Microsoft Power Platform at the OpenIT Studio.

2022 - 2025

Teaching + Research Assistant Hochschule Rhein-Waal · Kleve

Taught undergraduate C and MATLAB and built an AI grading assistant with roughly 85 percent efficiency gain. Research assistant on the Hall-effect sensor hardware that became an IEEE publication.

09/2021 - 03/2025

B.Sc. Mechatronics Systems Engineering Hochschule Rhein-Waal

Grade 1.7, Deutschlandstipendium scholar. University Senate student representative and robotics club mentor. The hardware grounding under everything I build, from PCBs to robot arms.

CH.05

Publications & patents

RESEARCH THAT SHIPPED

REFTITLEVENUEROLE

P.01 Neuroadaptive Architecture Integrating Synthetic Neural Feedback and Reasoning for Robotic ControlBhatia et al. · Auryal x BTU Cottbus-Senftenberg · companion poster at IK 2026, presenting author NAT 2026 FIRST AUTHOR

P.02 From Interlocutor to Cognitive Extension: Brain-Computer Interface as a Step Beyond Dialogue in Conversational AIMihic Zidar, Bhatia, Wicke ACM CUI 2026 · Bremen CO-AUTHOR

P.03 Decoding Workload and Agreement From EEG During Spoken Dialogue With Conversational AIMihic Zidar, Wicke, Bhatia, Lutz, Klug, Zander · under review arXiv 2026 CO-AUTHOR

P.04 Design of Electronic Chess Board Using Analogue Hall-effect Sensors for Piece IdentificationCheong, Bhatia, Krauledat, Hartanto IEEE ICHORA 2025 CO-AUTHOR

Contributor to two neuroadaptive-AI patent filings: HAL (Human Alignment Layer) and GCRM (Generative Cognitive Reward Model, filed at the European Patent Office).

OPEN LOGBOOK →

CH.06

Capabilities

T-SHAPED: BROAD, THEN DEEP

Agentic engineering

Agent harness: Claude Code, MCP servers, sub-agents, skills
Context engineering: worktree isolation, compounding skill memory
Evals gated in CI: a recall regression cannot merge
Spec to sign-off: a fresh agent grades every diff

LLM systems

Serving: vLLM, TGI, Ollama · structured outputs, tool calling
Model routing per task and per feature, cloud to on-device
Voice and realtime agents: streaming latency budgets, WebRTC
Guardrails: open-source PII redaction, policy authoring

Robotics & world models

VLA policies: pi0.5, GR00T · fine-tuned on real-robot data
Sim-to-real: Isaac Sim Franka to LeRobot SO-101
World-model grounding with Cosmos-Reason1
Closed-loop passive BCI at 15 Hz · LSL pipelines

Forward-deployed

Discovery workshops to multi-region enterprise rollout
Demos that close: investor robot demos, CEO-level stakeholders
Full-stack delivery: Swift 6, React, FastAPI, GCP, Quest 3
Standards a 14-person company adopted: onboarding, policy, OKRs

SELECT

What brings you here?

OPEN A FOCUSED VIEW

LLM SYSTEMS · EVALS · AGENT HARNESS

Agentic Engineering

Context engineering, evals gated in CI, multi-agent orchestration, model routing, forward deployment.

OPEN FOCUSED VIEW → VLA · BCI · NEUROADAPTIVE SYSTEMS

Autonomous Robotics

World models, VLA architectures, passive BCI, neuroadaptive control loops, SO-101 / LeRobot.

OPEN FOCUSED VIEW →

CH.07

Contact

OUTPUT TERMINAL

YOU · THE OTHER END OF THE LOOP

Let's close the loop.

Hiring, founding, or collaborating on BCI, robotics, or LLM systems? I read every message.

praneel.bhatia@gmail.com

LINKEDIN GITHUB SCHOLAR DOWNLOAD CV [PDF]

BERLIN, GERMANY · --:--:--

EU BLUE CARD · UAE GOLDEN VISA · OPEN TO RELOCATE · NO SPONSORSHIP NEEDED IN EU OR UAE