VS
Vache prompts. Claude codes.How it works
Blog

Writing from the system.
Shipping the work.

Build logs, technical deep dives, and research from shipping production software with AI-directed systems.

Latest Build
Research Brief

A Neural Router for My Knowledge Vault

I built a 15-agent routing system with a 7-layer scoring formula, a self-improving optimizer, and a hygiene scanner. The real insight wasn't any single layer — it was that sophisticated infrastructure without a discoverability layer rots on the shelf.

neural-routingknowledge-vaultautoresearchai-infrastructureclaude-code
Topic Hubs

Explore by Topic

Engineering & Building

Product development, AI collaboration, and building software with Claude.

Research & Analysis

Deep dives into science, economics, systems thinking, and cross-domain synthesis.

AI Infrastructure & Optimization

Local models, GPU optimization, ML training pipelines, and infrastructure automation.

Buyer Journey Stage
Series
All Posts
Research Brief

I Ran Autoresearch on My Own Website

Karpathy's autoresearch optimizes train.py against validation loss. What happens when you point the same pattern at a production website? 9 iterations, a 68% bundle reduction, 120 new internal links, and a real-time Convex backend — all discovered autonomously.

autoresearchperformanceseoconvexweb-optimization
Research Brief

The Science of Spaced Repetition: Why Spacing Your Learning Matters More Than You Think

Discover why spaced repetition is the most empirically supported learning technique in cognitive psychology, how it works in your brain, and how the vachsark learning engine uses it to maximize your retention.

learningspaced-repetitioncognitive-sciencememorystudy-technique
Build Note

How Psychology Explains Sales Pipeline Performance

Why variable-ratio reinforcement outperforms fixed bonuses in B2B sales. Connect operant conditioning research to pipeline analytics and buyer engagement.

salespsychologyb2b-strategybehavioral-economics
Research Brief

The Forensic Principle: Why Accountability Requires Irreversible Evidence

Forensic evidence trails solve a core problem: proving someone violated a rule without requiring trust in any single authority. The same principle appears in Byzantine consensus, legal systems, and corporate governance.

securitysystemsaccountabilitygovernancecryptography
Research Brief

Attention and Consciousness: Two Systems, Not One

Why you can see things without attending to them—and attend to things without seeing them. The neuroscience of why attention and consciousness are dissociable.

neuroscienceconsciousnessattentioncognition
Research Brief

Why Systems Collapse Suddenly: The Mathematics of Catastrophe

Catastrophe theory explains why lakes flip from clear to turbid overnight, marriages crumble after years of stability, and markets crash without warning. René Thom proved there are exactly seven types of threshold effects.

mathematicsdynamical-systemsphase-transitionstipping-points
Research Brief

Why Water and Iron Have the Same Critical Exponents

How physics discovered that wildly different systems near their breaking point obey identical mathematical laws. A story about universality, renormalization, and why microscopic details don't matter.

physicscomplexityemergenceuniversality
Research Brief

How the Brain Forgets Fear: The Prefrontal Circuit That Extinction Therapy Targets

Your brain has two competing systems for fear: one to express it, one to suppress it. Understanding the switch between them explains why exposure therapy works—and why it sometimes fails.

neuroscienceanxietypsychologymental-health
Research Brief

The Mechanics of Information Retrieval: Ranking Relevance at Scale

Information retrieval solves the problem of ranking millions of documents by relevance without exhaustive scoring, using layered architectures of lexical matching, semantic embeddings, and probabilistic ranking functions.

information-retrievalsearch-enginenlpbm25vector-search
Research Brief

Evaluating Language Models for Low-Resource Languages: Beyond BLEU

Why BLEU fails for low-resource NLP and which metrics actually work when you have one reference translation and a tiny annotation budget.

NLPevaluationlow-resourcemachine translationmetrics
Research Brief

Why Mechanism Design Can't Escape Social Choice Impossibility

Arrow's impossibility theorem doesn't just constrain voting outcomes — it constrains what mechanisms can implement. The Gibbard-Satterthwaite theorem proves the same problem reappears in mechanism design. Designers must choose which fundamental axiom to sacrifice.

economicsmechanism-designsocial-choiceimpossibilitygame-theory
Research Brief

Fixed Effects vs. Random Effects: When to Use What in Panel Econometrics

A practical guide to choosing between FE, RE, and Hausman-Taylor estimators in panel data analysis. Covers the Hausman test pitfalls and when each method is appropriate.

econometricscausal-inferencepanel-datamethodology
Research Brief

Path Dependence: Why Historical Accidents Shape Markets

How small contingencies compound into lock-in, from QWERTY keyboards to institutional inequality

economicstechnologyinstitutionsgame-theory
Research Brief

Why Different Things Behave Alike: Phase Transitions and Computational Complexity

How water freezing, hard math problems, and AI learning follow the same universal laws—and what that tells us about information.

complexityphysicsinformation theorymachine learning
Research Brief

The Quantum Brain: How Dopamine Encodes Superposition

A surprising structural parallel between quantum superposition and how dopamine neurons encode distributional value representations—and what it tells us about decision-making under uncertainty.

neurosciencequantum-mechanicsaireinforcement-learning
Research Brief

How Individual Incentives Create Population Equilibrium

The surprising mechanism where individual payoff differences drive group-level stability in replicator equations

game theoryevolutionary dynamicssystems
Research Brief

One Math, Four Domains: How Stochastic Processes Unify Finance, Systems, Physics, and Biology

Why stock market crashes, neural firing, gas diffusion, and server latency all follow the same probability distributions

mathstochastic-processesfinancesystems-engineering
Research Brief

Why Chronic Stress Breaks Threat Learning

Your brain learns threats through a dopamine-dependent circuit. Chronic stress floods it with cortisol and depletes dopamine, leaving you unable to update danger predictions. Here's the neuroscience—and what to do about it.

neurosciencestresslearningmental-health
Research Brief

Why Chronic Stress Breaks Your Decisions

Stress doesn't just make you tired — it physically shrinks the brain regions that make good decisions. Here's how, and what you can actually do about it.

neurosciencestressdecision-makingproductivity
Case Study

ClaudeSearch Day 2: 3.76 → 4.35, 175 Improvements, Self-Healing Validated

The Day 1 post described the vault-as-model metaphor and the three execution modes. This is the Day 2 update: what the quality data looks like after two days, what the self-healing loop actually fixed, and what shipping 175 improvements in 48 hours revealed about autonomous infrastructure.

autoresearchautomationlocal-modelsai-infrastructureopen-source
Research Brief

Overnight Results: From 3.76 to 4.35

What happened while I slept: 449 knowledge modifications, 11 new research notes, a vault quality score jump from 3.76 to 4.35, and auto-implement deploying real features. The knowledge flywheel is closing the loop.

vaultautoresearchautomationlocal-modelsknowledge-engine
Research Brief

Building RepRewards: The Engineering Behind a Gamified Fitness Tracker

Grace day streak mechanics, an XP economy built on habit formation research, buddy systems with shared streaks, and what it took to find and fix 24 bugs — including a 365-sequential-query N+1.

convexreactgamificationfitnesstypescript
Research Brief

vault-search: Building Hybrid Retrieval That Actually Works on Local Hardware

How BM25 + embeddings + RRF fusion, HyDE expansion, typed sub-queries, and a knowledge graph work together in a zero-API-cost local search tool — and how we cut query time from 9.9s to 0.5s.

searchretrievallocal-aiollamaopen-sourcenlp
Research Brief

Vault Autoresearch: A Personal AI Learns From Itself

What if your vault infrastructure improved itself overnight? Inspired by Karpathy's autoresearch, we built a system that treats the vault as a machine learning model, quality scores as loss, and improvement cycles as training steps. 30+ bottlenecks fixed, 34 GPU experiments, 12 code fixes deployed — all in one session.

autoresearchlocal-modelsvaultautomationllm-optimizationfine-tuning
Research Brief

How a 4B Model Beat a 7B: Fine-Tuning Bake-off on 16GB VRAM

Qwen3-4B scored dead last in our base model bake-off. Then it beat our production Qwen2.5-7B after fine-tuning — 4.8/5 vs 3.0/5 on format compliance. The secret: VRAM headroom matters more than parameter count.

fine-tuningqloraqwen3rdna4hqqlocal-modelsbakeoff
Research Brief

What Arrow's Impossibility Theorem Teaches Us About Distributed Systems

Consensus is hard. Whether you're voting for a president or electing a leader in a database cluster, the mathematical constraints are identical.

distributed-systemseconomicsmath
Research Brief

Bank Runs and Gossip Protocols: Why Cascading Failures Look the Same Everywhere

A bank run and a network partition cascade follow identical dynamics: local information, confidence drop, withdrawal, amplification. The same differential equation governs both — the only difference is whether the absorbing state is truth or catastrophe.

distributed-systemseconomicsepidemicscascading-failuresnetwork-effects
Research Brief

Building a Knowledge Vault: How I Use Claude to Research 80 Topics in a Day

A deep dive into the research pipeline behind 10,800+ knowledge notes: multi-agent orchestration, quality gates, synthesis generation, and autonomous maintenance on a local GPU.

knowledge-managementmulti-agentautomationobsidianresearch
Research Brief

MagicLog: Building a Construction RFC/Submittal Tracker in 5 Days

A tool for my dad (a project manager) to track RFCs, submittals, and change orders across construction projects. Next.js + Convex, with Excel import, PDF upload, audit trails, and real-time notifications. Built in 5 days of prompting Claude. Not yet in real use.

magiclognext.jsconvexconstructionfull-stackcase-study
Research Brief

Poverty Traps Are Bistable Systems: A Mathematical Lens on Development Economics

Development economics and dynamical systems theory describe the same phenomenon with different notation. A poverty trap is a stable equilibrium in a double-well potential. The policy implication: you need enough energy to cross the barrier, not just more energy.

economicsdynamical-systemsdevelopmentmathematicspolicy
Research Brief

7B QLoRA Fine-Tuning on AMD RDNA4: The HQQ Path

bitsandbytes doesn't work on the RX 9070 XT. Here's how we trained five 7B models using HQQ's pure PyTorch quantization — and the six lessons (three patches, three gotchas) that made it possible.

gpuqlorardna4amdfine-tuninghqqmachine-learning
Research Brief

TCP Congestion Control and Exchange Rate Overshooting: Dornbusch Meets Van Jacobson

Both TCP and currency markets overshoot because of the same mechanism: fast variables adjusting before slow ones can respond. Same math, same oscillation, same solution — damping through better models.

tcpeconomicscontrol-theoryovershootingnetworking
Case Study

Fine-Tuning Local Models: 10% to 100% Format Compliance in 5 Minutes

How I trained custom Qwen2.5-1.5B models on an RX 9070 XT — 9.2x faster, 4.6x less VRAM, and $0 cost. The secret: fine-tune format, not reasoning.

fine-tuninglocal-gpuollamamachine-learning
Research Brief

The Knowledge Engine: 10,800+ Notes Built by AI Agents

An autonomous knowledge system spanning 7+ disciplines — seeded, judged, deepened, and maintained by a pipeline of AI agents running 24/7 on local and cloud models.

knowledge-engineautomationzettelkastenmulti-agent
Research Brief

Building a Multi-Agent Research Pipeline — Benchmarked on DRB

A 6-wave multi-agent research system scoring 0.5166 on RACE (n=50, 0.5 = reference baseline) — competitive with NVIDIA AIQ and Gemini Deep Research, with honest notes on citation quality gaps.

researchmulti-agentautomationbenchmarks
Research Brief

Actor-Critic Methods: The Mechanics of Variance Reduction in RL

Actor-critic methods solve the high-variance problem of policy gradients by using a critic network to estimate value, enabling stable training for continuous control and LLM alignment.

reinforcement-learningactor-criticpporlhfmachine-learning
Research Brief

Linesheet Ships Eight Features in One Commit: From Buyer Segments to PO PDFs

A single 2,839-line commit adds ships-now filtering, sustainability metadata, order confirmation emails, multi-currency display, season archives, buyer segments, PO PDF generation, and enhanced analytics to the Linesheet app.

linesheetconvextypescriptreact-pdffeature
Build Note

Building a Knowledge Engine with Spaced Repetition and Client-Side Encryption

A 1,329-line React component that turns 10,800+ atomic knowledge notes into a browsable, quizzable learning tool — with FSRS-5 adaptive spaced repetition, AES-GCM encrypted private notes, and discipline-level progress tracking.

learningreactspaced-repetitionencryption
Research Brief

Polishing Inventory Intelligence: Shared Badges, Single-Pass Queries, and Validation

How Linesheet's Inventory Intelligence feature got refined — extracting a shared badge system across 3 components, replacing 8 separate filter calls with a single-pass accumulator, and adding server-side range validation to analytics thresholds.

linesheetinventoryconvextypescriptrefactor
Research Brief

Building a Fashion Trend Intelligence Pipeline for $3/Month

How I built an automated fashion trend research system that monitors 6 RSS feeds, extracts signals with a local model, and produces weekly slide deck reports with real article images — all for under $3/month.

automationlocal-gpuollamafashionmarp
Research Brief

Private AI for Legal Work

Run AI models locally for confidential legal work. No cloud, no API keys — everything stays on your machine.

ailocal-gpuollamaprivacylegal
Research Brief

Linesheet Security Hardening: Securing Convex Functions

An audit found 17 vulnerabilities in Linesheet's Convex backend. Any authenticated user could access other tenants' data. The fix: require explicit auth checks in every function and validate tenant ownership. How a one-line security pattern eliminated a critical attack surface.

convexsecuritylinesheettypescriptrefactor
Research Brief

Local Model Shootout: Finding the Right LLM for Every Task

Benchmarking 6 local models on an RX 9070 XT to build a three-model strategy — speed, code quality, and reasoning each get their own specialist.

ailocal-llmollamabenchmarksgemma3qwen3
Research Brief

What $1.09 Trillion in Medicaid Spending Actually Looks Like

HHS quietly released 227 million rows of provider-level Medicaid data. I built a platform to explore it. Here's what the numbers reveal about where the money goes.

datamedicaidpolicypythonnext.jsduckdb
Research Brief

Vulkan Beats ROCm: +20% LLM Inference on RDNA 4

How building llama.cpp from source revealed that RADV Vulkan outperforms ROCm HIP on the RX 9070 XT — and the pipeline parallelization that delivered a 44x speedup on top.

gpuvulkanrocmllama.cppoptimizationrdna4
Research Brief

The Supreme Court Just Struck Down Trump's Tariffs. Here's What Actually Happens Next.

A fact-based breakdown of Learning Resources v. Trump: the 6-3 ruling, the $175 billion refund question, the replacement tariff Trump signed hours later, and what it all means for businesses and consumers.

tariffssupreme-courttradeeconomics
Case Study

Building blog.vachsark.com in One Session

How I shipped a full MDX blog with custom typography, auto-drafting heartbeat tasks, and Cloudflare deployment — and the silent config bug that cost four debugging cycles.

reactmdxcloudflarevitedebugging
Research Brief

Shipping AI Image Generation While Cutting 2MB from the Bundle

How Linesheet added a full Gemini-powered AI generation pipeline, shaved ~2MB from initial page loads, and fixed a subtle Convex caching bug — all in one sprint.

next.jsconvexgeminiperformanceai
Research Brief

Reweave: Teaching Your Vault to Link Itself

Building a system that discovers semantic connections between notes and auto-generates wikilinks — turning a flat file vault into a self-wiring knowledge graph.

obsidiansemantic-searchknowledge-graphautomation
Research Brief

Zero-Cost Automation: 16 Tasks on a Local GPU

How I built a heartbeat system that runs 16 automated maintenance tasks — dependency monitoring, vault health checks, ecosystem scanning — all on local models at zero API cost.

automationlocal-gpuollamadevops
Research Brief

Linesheet Bundle Diet: Cutting 2MB from Initial Load

How lazy-loading Clerk, react-pdf, and Sentry Replay shaved ~2MB from a Next.js 15 production app — and fixed a subtle Convex caching bug along the way.

next.jsperformancereact-pdfclerkconvex
Research Brief

Local AI Optimization: From 148ms to 85ms

How I profiled and optimized semantic search across 1,620 vault files — numpy vectorization, float32 storage, and model benchmarking on a local GPU.

aioptimizationembeddingslocal-gpu
© 2026 Vache Sarkissian