Work

  1. Aug '26

    Rkive AI · camera effects engine

    • A face-reshape and retouch engine

      Retouching driven by facial landmarks on the live frame — skin, makeup and facial geometry as three families of continuous controls rather than presets, so the adjustment tracks the face as it moves instead of sitting on top of it.

    • One effects pipeline, written twice

      The same effects stack expressed against each platform's own GPU path — Metal compute kernels on iOS, OpenGL on Android — fused into a single pass off the main thread, because a camera preview only has the gap between two frames to spend and a pass per effect spends it several times over.

  2. Jun '26

    Rkive AI · Media compositing

    • Compositing in the browser

      The web compositor draws every manual edit — titles, subtitles, overlays — live over the footage, so a change previews the instant you make it instead of waiting on a re-encode.

    • Compositing on iOS

      A native iOS compositor, on Skia, that matches the web one frame-for-frame and holds audio-video sync through playback — the live preview runs on the phone's own hardware.

    • Compositing on Android

      The same Skia compositor on Android, so the live preview is identical across web, iOS and Android rather than three approximations of one another.

    • Compositing-vs-encoding routing

      A layer that keeps an edit in the live composition for as long as it can and only re-encodes when it must — instant previews while editing, and far less compute than baking every change.

  3. Jun '26

    albertoluengomartin.com · this site

    • The ASCII transition layer (this site)

      Everything here is drawn as characters on a single canvas that outlives every page change — so navigating isn't a cut between pages but one field of characters rearranging itself from one into the next.

  4. Jun '26

    Rkive AI · unified agent system

    • Cross-surface search

      A single search over files, messages, conversations, attachments and the web that follows the relationships between things rather than just matching words — and cites the exact passage it drew from.

    • A versioned memory layer

      A memory for AI assistants where facts aren't overwritten but versioned — carrying confidence, strengthening with reinforcement, and superseding earlier beliefs — so the assistant's model of you has a history, not just a latest state.

    • A proactive background daemon

      A background process that, on its own schedule, synthesizes a handful of concrete proposals from everything it knows about your work — the first half of a computer that keeps thinking about you when you're away.

    • Conversation routing and compression

      Sends each new message to the right existing conversation on its own, spins up continuations as threads near their limits, and lets you pull just the relevant slice of a long history into a new prompt.

    • A unified attachment model

      Collapsed two tangled upload paths into a single idea — an attachment belongs to a conversation and can move between them — with deduplication and a caption fallback so even text-only models can read an image.

  5. May '26

    ARC Prize · ARC-AGI-3

    • A swappable reasoning oracle

      One reasoning contract — query the game's structured state, get back reasoning and the next moves — with the reasoner swappable behind it: a scripted teacher, a live model in the loop, or a small distilled model trained on the teacher's traces to fit the competition's size limit.

    • A continuously-steered search engine

      A search engine whose heuristics steer it continuously rather than as on/off rules — folding many signals about the current state into a soft directive, so exploration flows toward what looks promising instead of hard-gating.

    • Autonomous state abstraction

      Learns which parts of a changing environment move on their own and abstracts them out of the state — turning an intractable search space into a small one.

    • A depth-biased random-walk fallback

      When systematic search stalls, fall back to a guided random walk biased toward less-visited, deeper states — recovering nearly all the wasted budget and reaching far more of the space.

  6. May '26

    Rkive AI · Atlas

    • A 3D embedding explorer

      Turns high-dimensional embeddings into a space you can move through — clusters as nested glass containers with typed relationships between them — so a whole body of knowledge can be explored by navigating it. (The Work view you're reading is a flat cousin.)

    • Your data, as an atlas view

      The same explorer turned on your own material — files, conversations, contacts — as a built-in view you can open any time, always current and private to you, not something you have to generate.

    • A prompt-to-atlas pipeline

      An agent pipeline that takes a single prompt and produces a whole published atlas — gathering and describing the material, embedding and relating it, generating imagery, clustering it — with a human review gate before anything ships.

  7. May '26

    Independent research · perfumery

    • A perfumery knowledge graph

      Modeled the world of perfumery as a rich graph — materials, accords, families and the relationships among them, with learned embeddings — the structure the embedding explorer was first built on.

    • A material-level odor dataset

      Built the first material-level dataset of odorants — pulling perfumery, molecular and culinary sources into one resource for studying how things actually smell.

    • Odor representation learning

      Representational research on odor — testing what really predicts how materials smell together, and finding that how they're used outperforms their chemical structure.

  8. Apr '26

    Rkive AI · Superapp

    • Studio, comms, and ops in one app

      Folded Studio, communications, finance-and-operations analytics and file editing into a single application — one surface the unified AI works across, instead of four disconnected tools.

    • Finance and operations analytics

      Brought financial and operational analytics into the same app — the numbers that run a business sitting next to the tools that do its work.

    • An in-app file and code editor

      An editor for files and code built into the assistant, so it can see and change real work directly rather than only describing what to do.

  9. Apr '26

    Rkive AI · AIR

    • In-context search steering

      An agent that steers a large automated-ML search as it runs — reading the experiment history and reallocating compute toward what's working, entirely in context, with no fine-tuning.

    • A Gaussian-process surrogate model

      A Gaussian-process surrogate that predicts an experiment's outcome from its configuration, so the search explores promising directions by sampling the model instead of paying to train every candidate.

    • A reproducible-experiment grammar

      A framework where a researcher declares the intent of an experiment — what's being compared and why — and the system turns that into a fully reproducible run, so a result can never drift from its definition.

    • Contract-based experiment protocols

      A small set of formal experiment types — from open-ended discovery to confirmation and qualification — each a contract that fixes what must be measured and who is responsible for what, so a study means the same thing every time it runs.

  10. Mar '26

    Rkive AI · Golden Glass UI

    • A parametric glass material

      Reverse-engineered how light moves through real glass into a layered, declarative material — bezel, shadow, diffraction, sheen — that renders identically on the web and in native apps.

    • A position-driven scroll engine

      A scroll engine where every surface responds to its position in the viewport — angle, lens, shadow and type each easing on their own — so a page reads as one continuous material rather than stacked elements.

  11. Mar '26

    Rkive AI · edge media processing

    • On-device media, iOS

      A native iOS pipeline that runs edits and effects on the phone's own hardware — no upload, no round-trip.

    • On-device media, Android

      The same local capture-to-render pipeline on Android, tuned to the platform's own media stack.

    • Media processing in the browser

      A web pipeline that does in-page what the native apps do on-device, so the browser isn't a second-class citizen.

    • Device-or-cloud routing

      A router that decides per task whether an edit runs on the device or a server GPU — by size, battery and capability — so the heavy work lands where it's cheapest.

  12. Mar '26

    Rkive AI · adaptive anti-jailbreak

    • Adaptive jailbreak detection

      Catches prompt-injection by comparing a model's output against its live system prompt rather than a fixed blocklist, so it keeps working as the prompt evolves — layered with a behavioral perimeter and a semantic check.

    • System-prompt obfuscation

      Continuously rewrites the live system prompt so a jailbreak that worked once can't be replayed, pairing with detection so an attacker has to rediscover both the prompt and the exploit each time.

  13. Feb '26

    OpenAI · Parameter Golf

    • Gradient-guided mixed-precision quantization

      Gives each tensor its own bit-width by how sensitive the model's loss is to it, shrinking a network with little accuracy cost. From work on parameter-golf, where the aim is the strongest model inside a tiny budget.

    • Backout connections

      A training idea of his own: learned intermediate exits that let gradients leave the network before its final output, widening where a model can learn from.

    • Spectral warmdown

      A learning-rate schedule that gauges how converged each layer is and spends the last of the training budget where it still moves the needle.

    • flash-dispatch

      One call that routes attention to whichever fast kernel a given machine supports, and transparently upgrades existing code onto the faster paths — so models run at the best available speed without rewrites.

  14. Feb '26

    Rkive AI · in-app camera

    • A hybrid native camera

      A camera built as native iOS and Android modules bridged into one React Native surface — native capture and preview performance behind a single cross-platform interface.

    • Real-time segmentation and effects

      On-device segmentation that separates subject from background as you film, feeding a layered effects system that composites live rather than in post.

    • An in-camera teleprompter

      A teleprompter that runs the AI editor's own timed script — highlighting it word-by-word or line-by-line in sync, with per-line cues for rhythm and pauses — so the take you record already fits the cut that's planned.

    • A Golden Glass camera interface

      A capture interface built from scratch in Rkive's Golden Glass material — controls, modals and framing rendered as layered glass over the live camera, not a stock system camera sheet.

  15. Dec '25

    Rkive AI · multi-layer anti-jailbreak

    • Semantic-overlap detection

      Flags a prompt injection by how closely an input's meaning tracks the protected system prompt — the meaning-level layer of the defense.

    • Syntactic-overlap detection

      Catches the cruder attacks by surface overlap — shared structure and phrasing with the system prompt — a fast first pass before the semantic check.

    • Prompt sanitization

      Strips and neutralizes known injection patterns from user input before it ever reaches the model.

  16. Nov '25

    Rkive AI · server media processing

    • Server-side image processing

      A GPU image service for the work too heavy for a phone — the expensive transforms run server-side at full quality, then come back.

    • Server-side video processing

      A GPU video pipeline for render-heavy edits — decode, transform and re-encode on server hardware when a timeline outgrows the device.

  17. Oct '25

    Rkive AI · media agent system

    • An autonomous-creation planner

      Turns one intent into very different kinds of output — a single post, a multi-part portfolio, a background maintenance task — choosing the form from what is being asked rather than from a menu.

    • An autonomous video editor

      The first system to treat a finished edit as something an agent produces rather than a person — taking footage and a goal to a cut, the seed the later agent systems grew from.

  18. Mar '25

    Manfacter AI · additive manufacturing

    • A deep model distilled into a formula

      Estimated cost and time and generated quotes across every material, machine setting and printing technology — by training a deep model, then distilling it into one parametric function whose coefficients could be read and tuned by hand. The accuracy of deep learning, kept interpretable.

    • The system and AI infrastructure

      The systems and ML infrastructure underneath it — data, training and serving wired together so the models could run as part of a real manufacturing pipeline.

    • Image and text to a printable 3D model

      A generative model that turns a reference image and a written description into a 3D model ready for the printer.