Writing

Notes and deep-dives — bridging theory and the systems I build.

  1. · 11 min read

    RAG That Renders the Diagram: Outline-Aware Chunking and Image-Grounded Answers

    How I built a hallucination-resistant RAG over a large technical-document corpus: chunk along the author's heading outline, embed with that outline as context, and treat images as first-class chunks with a deterministic verifier.

    EngineeringRAGRetrievalLLMSearchDocument AI
    Read
  2. · 10 min read

    The real cost of best-arm identification isn't the arms — it's the clusters

    When arms come with a similarity graph and rewards vary smoothly over it, best-arm identification's binding cost drops from the number of arms to the number of clusters.

    multi-armed banditspure explorationbest-arm identificationgraph signal processinglearning theory
    Read
  3. · 12 min read

    The Branch Is the Variant: Zero-Copy, Privacy-Safe Test Data on a Versioned Lake

    How a versioned data lake made the branch itself the test-data variant — zero-copy generation, blobstore-native metadata, and in-database hooks as the output channel.

    EngineeringData InfrastructurelakeFSSystem DesignKubernetesPrivacy
    Read
  4. · 10 min read

    Why You Can Start Gradient Descent Anywhere — Even on an NP-Hard Recovery Problem

    With order-n complex Gaussian quadratic measurements, recovering a vector up to phase becomes both identifiable and solvable by gradient descent from any starting point — two faces of one geometric condition.

    phase retrievalnonconvex optimizationquadratic feasibilitysample complexityoptimization landscape
    Read
  5. · 11 min read

    One Coin Flip That Spans Fast Search and Cheap Search

    Two robotic-search strategies optimize opposite costs; randomly interleaving them with one tunable coin lets you slide between "find targets now" and "spend the least energy."

    multi-armed-banditsmulti-agent-searchadaptive-sensingroboticsthresholding-bandit
    Read
  6. · 10 min read

    Generate the Spectrum, Don't Store It: Physics as a Trainable Prior for Unmixing

    Why a 100-year-old optical-physics model, made differentiable, beats data-hungry neural nets at hyperspectral unmixing exactly when labeled data is scarce.

    hyperspectral unmixingdifferentiable programmingremote sensinginverse renderingphysics-based MLspectroscopy
    Read
  7. · 9 min read

    Learn the Calendar Once, Not the Leader Every Week

    Why amnesiac non-stationary bandits keep paying to rediscover a periodic optimum, and how a number-theoretic period estimator lets the policy learn the cycle once instead.

    Multi-Armed BanditsNon-Stationary LearningRamanujan Periodicity TransformSignal ProcessingRegret Minimization
    Read
  8. · 10 min read

    Design the Algorithm First, Then Reverse-Engineer the Utility It Secretly Maximizes

    A trick for games whose natural Nash equilibrium is intractable: hand-build the adaptive rule with the right selfish reflexes, then discover the well-behaved utility it implicitly optimizes.

    game theorylearning in gamesNash equilibriumdistributed algorithmscongestionwireless
    Read
  9. · 6 min read

    Projected Gradient Descent with Skipping

    Exploring an interesting optimization technique that skips projection steps to potentially improve convergence in constrained convex optimization.

    optimizationgradient-descentresearchmathematics
    Read