# pavlopuzikov.com: full context > Deep-context variant of /llms.txt. Same identity + AI-readable-surface pointers, plus the full project blurbs and research-thread bodies the lean file links out to. Canonical site: https://pavlopuzikov.com. ## Identity - Name: Pavlo Puzikov - Role: Innovation Researcher - Location: Dubai, UAE - Email: pavlopuzikov@gmail.com - LinkedIn: https://www.linkedin.com/in/pavlo-puzikov - GitHub: https://github.com/pavlopuzikov - Letterboxd: https://letterboxd.com/pavlemio/ ## What Pavlo does An innovation and product researcher working across immersive 3D, applied AI, and generative video. The work spans a real-time Gaussian-splat research pipeline, directing generative-video launch films, earning brand visibility inside AI answer engines, and representing brands on-stand at global tech expos. The throughline is taking an emerging technique from first experiment to something a business can put in front of a client. Specialisms (in order of depth): 3D Gaussian Splatting (two companion academic papers in progress), large-language-model fine-tuning + multi-agent orchestration, computational valuation, spatial computing. In parallel he runs marketing and partnerships for Living Homes (luxury smart-home developer), representing both brands on-stand at GITEX Global 2024 + 2025, LEAP 2025, and GITEX Asia 2025. ## Projects ### 01 - Barnes Vantage (3D & Spatial, In development) One map-led surface that folds the scan library, the live 3D viewer, and the listing intelligence into a single tool instead of four. Every point on the map is a real on-site capture; pick one and walk it in the browser. Outcome: Folds the standalone Splat Viewer, the scan library, and the valuation signals into one dark, map-led surface a broker opens instead of four tools. In development: the map, the natural-language search, and the rankings, predictions, and experiments panels run against live scan data, with the same walkable scan engine embedded at the centre. Case study: https://pavlopuzikov.com/work/splat-viewer ### 02 - Generative Video & Creative (Marketing, In production) Directing generative-video films end to end, from brief and storyboard to directing the shots to a delivered edit, working closely with talented videographers and editors, alongside creative direction across BARNES Dubai and Living Homes. The Marina Sozopol launch film ran to 27 storyboarded shots blending real footage, 3D renders, and AI-generated video; the In motion series carries the brand films, with the Maison Margiela Dubai Residences launch as the flagship. Case study: https://pavlopuzikov.com/work/maison-margiela-residences ### 03 - 3DGS Research Pipeline (3D & Spatial, In development) 20 Gaussian-splat engines integrated into one benchmark harness, 6 evaluations run so far. Which one actually survives a real luxury-interior scan, not just the leaderboard. Outcome: A reproducible benchmark harness: around twenty engines wired, 6 evaluations run so far, a scan-difficulty taxonomy and per-engine failure modes, that the academic write-up draws on. The full engines-by-scans matrix and the papers are in progress. A relighting track now runs alongside the benchmark: an interactive viewer with movable light fixtures and real soft cast shadows, plus a delighting and relightable-retrain line still in progress. Case study: https://pavlopuzikov.com/work/3dgs-research ### 04 - Data Atlas (AI / ML, In production) An entity-centric valuation platform and on-demand AVM over 538K entities across 28 sources, with a MAPE and RMSE calibration and backtest harness; an internal snapshot showed ~7% MAPE. Outcome: Turns multi-day analyst cycles into an on-demand AVM brief, ranked opportunities, and an accuracy snapshot the valuation team can audit. Case study: https://pavlopuzikov.com/work/atlas ### 05 - Barnes Dubai LLM (AI / ML, In production) A domain-tuned LLM and agentic broker assistant for Dubai luxury real estate, a QLoRA fine-tune on 15,369 instruction pairs, running locally. Outcome: Live in the Innovation Portal playground. A broker asks `Compare Business Bay vs JVC for buy-to-let yield` and gets a tool-resolved verdict in seconds (medians, gross yields, tenant pool, a one-paragraph recommendation), the brief that used to take an analyst an afternoon. Case study: https://pavlopuzikov.com/work/barnes-ai ### 06 - Academic Papers (Research, In development) Two companion academic papers on applied Gaussian splatting for luxury interiors, one on production-deployment lessons from the benchmark, one (Reflection-Aware 5DGS) on the mirror-and-glass problem. Both drafted; a Tech Communications submission in preparation. Case study: https://pavlopuzikov.com/work/deploying-3dgs ### 07 - Innovation Portal (Product, In production) Internal cockpit for the BARNES Dubai innovation team: project tracking, multilingual RBAC (EN / FR / AR), embedded Apache Superset BI, and a marketing-engineering toolkit (SEO/AEO/GEO, competitor console, Data Atlas, 3D Scans, CRM Copilot, an in-portal Barnes Dubai LLM assistant, and an AI content studio for marketing scripts and storyboards). Outcome: Live internally, the single surface from which the BARNES Dubai team runs 23 innovation projects, 170 milestones, 345 tests, and the AEO/SEO + Barnes Dubai LLM tooling. Case study: https://pavlopuzikov.com/work/innovation-portal ### 08 - House Call (Product, In beta) Invitation-only natal-reading practice. Hand-prepared readings synthesise Western, Chinese BaZi, and Vedic Jyotish, backed by per-user accounts, daily astronomy-engine transits, a falsifiable prediction tracker, synastry, and tarot import. Outcome: Live at housecallastro.com as an invitation-only private beta, payments off. The public launch and pricing follow once the founders' cohort closes. Case study: https://pavlopuzikov.com/work/threadwork ### 09 - AriOS (AI / ML, In beta) A personal knowledge companion that captures what you see, read, and hear into an Obsidian vault, then helps you actually learn it. A Knowledge Trainer turns notes into flashcards, grades your free recall, and schedules spaced-repetition reviews. Local-first; vault sovereignty is the hard constraint. Outcome: Daily driver, now a learning engine too. Capture and the trainer's auto-cards + graded recall are shipped; server-side spaced repetition with Telegram reminders is the active build. Case study: https://pavlopuzikov.com/work/ari ### 10 - Splatlas (3D & Spatial, Live) A navigable solar system where every world is a real 3D Gaussian splat, trained on a GPU from open NASA and USGS imagery and elevation, not a texture on a sphere. Fly through, warp to any planet, read it up close. Outcome: Live at splatlas.space: the Sun to Neptune as navigable splats with real orbits, a date scrubber, and live space-weather, all running in the browser. Case study: https://pavlopuzikov.com/work/splatlas ### 11 - Smart Deal Global (Marketing, Live) A public marketplace for luxury real-estate opportunities, multilingual (EN / AR / FR / RU), with hreflang and structured data tuned for search and answer-engine discovery. I built the first version of the site; a talented colleague has since drastically upgraded it. Outcome: Lead capture via WhatsApp, phone, and email to sales advisors. Case study: https://pavlopuzikov.com/work/smart-deal-global ### 12 - SEO · AEO · GEO Toolkit (Marketing, In production) An audit, scoring, and first-party measurement toolkit for staying visible in AI answers (ChatGPT, Perplexity, Claude) and in Google, scored from your own Search Console data, not a third-party broker's. Outcome: Live in the BARNES Innovation Portal: paste any URL and get a real on-page audit, separate SEO and AEO/GEO readiness scores, a simulated answer-engine visibility check, and an audit-grounded action plan. A companion first-party layer reads Google Search Console and GA4 locally and returns quick wins (queries ranking 5 to 15), topic clusters, content gaps, title and meta CTR fixes, and a weekly report, so one instrument covers both what AI answers cite and what real search traffic does. Case study: https://pavlopuzikov.com/work/ai-search-visibility ## Research threads Three paper-to-product threads. Canonical view: https://pavlopuzikov.com/research ### 3D Gaussian Splatting _Real-time radiance fields applied to luxury real-estate visualisation._ The Kerbl et al. 3DGS paper landed in 2023 and the field exploded into dozens of implementations, each making different trade-offs on training speed, render quality, and runtime fidelity. The benchmark closes the gap between method papers and what actually deploys to a Safari-iOS budget. The shipped artifact, Splat Viewer, is the answer running in production on the BARNES Dubai property-search flow. Steps: - Paper - 3D Gaussian Splatting for Real-Time Radiance Field Rendering (Kerbl, Kopanas, Leimkühler, Drettakis, SIGGRAPH 2023) (https://arxiv.org/abs/2308.04079) - The seminal 3DGS paper. Replaces NeRF's MLP with millions of explicit Gaussians; trains in minutes, renders at 60+ fps. Abstract: The authors introduce three elements that together achieve state-of-the-art visual quality while maintaining competitive training times and high-quality real-time (≥30 fps) novel-view synthesis at 1080p. Scenes are represented as 3D Gaussians initialised from sparse SfM points; an interleaved optimisation + adaptive-density-control step shapes the radiance field; a fast tile-based rasteriser keeps the render path GPU-resident. Demonstrated on several benchmarks against the prior NeRF state of the art. - Benchmark - 20-engine 3DGS pipeline benchmark (6 evaluations so far) (https://pavlopuzikov.com/work/3dgs-research) - Around twenty Gaussian-splat engines wired into one harness across studio scans, hard scans, and city-scale capture, the first evaluations run. PSNR, SSIM, LPIPS, plus a controlled raw-vs-cleaned comparison the Operate-the-Result slider lets you wipe through. - Shipped - Splat Viewer in the BARNES property-search flow (https://pavlopuzikov.com/work/splat-viewer) - Web-native viewer with tap-to-move navigation, LOD streaming, and golden-hour relighting. First georeferenced building (Onyx Tower) live for brokers. Per-thread anchor: https://pavlopuzikov.com/research#gaussian-splatting ### Mirrors and glass capture _The open problem Gaussian Splatting still hasn't solved for high-end interiors._ Splatting reconstructs surfaces from photographs. Mirrors and floor-to-ceiling glass break that assumption, the surface is a reflection, not a thing, and current pipelines either ghost the reflected scene into the wrong volume or hollow the geometry behind the glass. For a luxury-real-estate use case (where mirrors and glass walls are everywhere), this is the bottleneck. The Living Homes industrial POV is that projector showrooms are the wrong proxy and that the problem has to be handled when the scene is captured, not corrected in the renderer afterward. Steps: - Reference - MirrorGaussian: reflective surface reconstruction (Liu et al. 2024) (https://arxiv.org/abs/2405.11921) - A leading attempt at handling mirrors in 3DGS by jointly modelling the real and reflected geometry. Promising on benchmarks; not yet plug-and-play for production capture. Abstract: The key insight is that a mirrored scene is symmetric to the real scene with respect to the mirror plane. Building on this, the authors introduce a dual-rendering strategy that models the real-world and the mirror-world simultaneously, with an intersection-aware mirror Gaussian formulation that handles partial reflections. Demonstrated on real-world scenes containing mirrors with state-of-the-art reconstruction quality vs. prior 3DGS variants. - In progress - Capture-side mitigation pipeline (https://pavlopuzikov.com/work/3dgs-research) - Inside the benchmark, the per-engine failure analysis names mirrors / glass / chrome as the dominant artefact class on the Hard tier. Mitigation candidates are at the photogrammetry layer (mask + inpaint pre-COLMAP, dual-pass capture, controlled-lighting subsets) rather than the rasteriser. Per-thread anchor: https://pavlopuzikov.com/research#mirrors-glass-capture ### Domain LLM fine-tune _Parameter-efficient adaptation of a base model to Dubai luxury real-estate._ The QLoRA paper made it tractable to adapt a 7-13B base model on a single GPU. Barnes Dubai LLM v8 is the production answer to 'what does a domain-tuned model buy you versus base + RAG?' for a specific vertical (Dubai luxury real-estate, mixed-language, terminology-heavy, low-frequency entities). Trained on 15,369 instruction pairs sourced from Dubai Land Department records, UK Land Registry, France DVF, Singapore URA, and World Bank data. Steps: - Paper - QLoRA: Efficient Finetuning of Quantized LLMs (Dettmers, Pagnoni, Holtzman, Zettlemoyer, NeurIPS 2023) (https://arxiv.org/abs/2305.14314) - 4-bit NormalFloat + double quantization + paged optimisers, the toolkit that brought 33B+ fine-tunes onto a single workstation GPU. Abstract: QLoRA is an efficient finetuning approach that reduces memory usage enough to finetune a 65B-parameter model on a single 48GB GPU while preserving full 16-bit-finetune task performance. QLoRA backpropagates gradients through a frozen 4-bit quantized pretrained language model into Low-Rank Adapters (LoRA), and introduces three contributions: a new 4-bit NormalFloat data type, double quantization, and paged optimisers to manage memory spikes. - Implementation - 15,369 instruction pairs, multi-source Dubai corpus (https://pavlopuzikov.com/work/barnes-ai) - Instruction set sourced from DLD transactions + buildings + brokers + projects + developers, plus UK/FR/SG comparables. Tool-calling agents (getCommunityInventory, comp-search, valuation lookups) wrap the base capability. - Shipped - Barnes Dubai LLM v8 in production (https://pavlopuzikov.com/work/barnes-ai) - FastAPI service the brokerage team dogfoods in a playground before WhatsApp distribution to brokers. The customer-facing answer to 'can a domain LLM beat base + RAG for our vertical' is currently 'yes, on terminology and named-entity recall'. Per-thread anchor: https://pavlopuzikov.com/research#qlora-domain-llm ### Spatial asset intelligence _Turning one handheld building scan into a per-object, queryable asset registry._ A single walkthrough scan should do more than render a 3D model. It should locate every fire extinguisher, smoke detector and AC unit as a clickable object. The hard part is not detection, it is the camera trajectory: the device exports only ~23 seconds of poses out of a six-minute scan, so the rest is re-derived from raw LiDAR and IMU. A KISS-ICP spike is metric on the stable opening segment but loses tracking over the full walk, so IMU-fused LiDAR-inertial odometry is next. Steps: - Reference - KISS-ICP: In Defense of Point-to-Point ICP (Vizzo, Guadagnino, Mersch, Wiesmann, Behley, Stachniss, IEEE RA-L 2023) (https://arxiv.org/abs/2209.15397) - Point-to-point ICP against a sparse voxel map, with no IMU and no learned components. The dependency-light baseline for recovering a sensor trajectory from raw LiDAR sweeps. Abstract: The system performs LiDAR odometry via point-to-point ICP combined with adaptive thresholding for correspondence matching, a robust kernel, motion compensation from a constant-velocity model, and point subsampling, working across sensors and mounting positions at sensor frame-rate without dataset-specific parameter tuning. The authors argue much of the complexity in modern odometry pipelines is unnecessary and a carefully engineered classic approach stays competitive. - Reference - FAST-LIO2: Fast Direct LiDAR-Inertial Odometry (Xu, Cai, He, Lin, Zhang, IEEE T-RO 2022) (https://arxiv.org/abs/2107.06829) - Tightly-coupled LiDAR + IMU via an iterated Kalman filter on an incremental kd-tree. The robustness step for the fast-rotation and featureless moments where LiDAR-only odometry drops. Abstract: FAST-LIO2 is a direct LiDAR-inertial odometry framework that registers raw points to the map without feature extraction, tightly fusing inertial measurements through an iterated extended Kalman filter, and maintains the map with an incremental k-d tree (ikd-Tree) supporting incremental insertion and dynamic re-balancing. It demonstrates high accuracy and robustness at high odometry rates across structured and unstructured environments and varied LiDARs. - In progress - Trajectory recovery on the Onyx Tower reference scan (https://pavlopuzikov.com/research#spatial-asset-intelligence) - On the reference scan a KISS-ICP spike recovers a metric path on the stable opening segment but loses tracking over the full six-minute walk; IMU-fused LiDAR-inertial odometry is the next step. The payoff is lifting per-frame object masks to per-object 3D across the whole scan, beyond the roughly 6 percent of assets currently anchored by direct camera evidence. Per-thread anchor: https://pavlopuzikov.com/research#spatial-asset-intelligence ## Marketing-engineering specialisms - SEO / AEO / GEO: answer-engine-optimisation tooling inside the BARNES Innovation Portal: keyword + prompt research, generative-engine visibility scoring, AI-answer-engine tracking across ChatGPT / Claude / Perplexity / Gemini, competitor research console. - 3D Gaussian Splatting in production: end-to-end pipeline ownership (on-location LiDAR + photogrammetry capture, training stack orchestration, mobile-grade WebGL delivery, golden-hour relighting). - Domain LLM fine-tuning: instruction-pair dataset construction, QLoRA on local hardware (Ollama), evaluation harness, multi-agent inference wrappers. - Computational valuation: hex-grid geospatial modelling, ML-driven price forecasting, entity-resolution for real-estate transaction data. ## Press / fairs GITEX Global 2024, GITEX Global 2025, LEAP 2025, GITEX Asia 2025: on-stand representation for both BARNES Dubai and Living Homes. ## Sibling AI-readable surfaces - https://pavlopuzikov.com/llms.txt: lean entry point. - https://pavlopuzikov.com/resume.json: JSON Resume v1.0.0 (structured CV). - https://pavlopuzikov.com/research-index.md: flat markdown index of /research with per-thread anchors. ## Cite Canonical: https://pavlopuzikov.com. Per-project: https://pavlopuzikov.com/work/[slug]. Per-research-thread: https://pavlopuzikov.com/research#[thread-id].