Technical notes, with the evidence boundaries visible.
Concise architecture notes from projects I have built, plus dated research notes on AI systems and models. Reading times are calculated from the actual text.
TokensCache: Testing a 35.5% Token Reduction in a Controlled Agent Benchmark
How exact and semantic caching were evaluated with 92 repository tests and a scoped mock coding-agent A/B comparison.
Reliability Guard: 24 Scenarios for Catching Unsupported Agent Claims
A verification harness for fabricated completion, missing evidence, stale state, and contradictory coding-agent output.
TruthNet: Building a Four-Agent Fact Checker in Seven Hours
How scope control, adversarial agent roles, a streaming backend, and stage-demo constraints shaped the hackathon build.
Interviewlary: Designing Role-Specific Voice Interview Practice
A product and architecture note on voice sessions, role-aware questions, structured scoring, and actionable feedback.
RoutineAI: Building a Student Planning App with Next.js and Supabase
Deterministic scheduling, AI-assisted planning, reminders, focus tools, and the 106-test verification scope behind the student planner.
Web Dev Skills: Structuring Repeatable AI-Assisted Delivery
How reusable skills, lifecycle prompts, and phased checks make AI-assisted web work easier to review and repeat.
KaisLearnAI: Protecting Learning Data with Supabase RLS
A bilingual learning-platform pattern for role-aware access, protected progress data, and database-enforced tenant boundaries.
StudyK: Connecting a React Admissions Interface to Edge Functions
A practical look at typed form intake, serverless workflows, and automated follow-up for an education consultancy product.
Damascus Trip: Designing a Bilingual Guide to Historical Places
A bilingual web guide combining curated place data, search and filtering, favorites, visited state, and right-to-left typography.
Kaiss Tech: Structuring Short, Reusable AI Engineering Guides
Lessons from turning agent workflows, prompting patterns, and tool integrations into concise educational material.
Lary OS: Designing a Founder Operations Control Plane
A planning pattern for executive briefs, vendor monitoring, workflow review, and scheduled operational automation.
MCOS: Planning Evidence-Backed Content Search with pgvector
A system design for PostgreSQL vector search, research synthesis, source traceability, and human approval gates.
Opti: Architecture for a Local Voice Assistant
A private implementation pattern combining wake-word detection, speech recognition, local speech output, and tool execution.
Parallel System: Isolating Coding Agents with Git Worktrees
A specification for parallel agent work using separate checkouts, explicit ownership, and independent validation gates.
How to Evaluate Long-Horizon Agent Claims in Frontier Models
A source-checking framework for context length, multi-step tool use, recovery behavior, and long-running agent stability.
GPT-5.6 Sol: Separating Peak Token Speed from End-to-End Latency
How to read the official peak-speed claim and measure first-token, tool-call, reasoning, and total task latency separately.
How Dynamic Reasoning Changes Fast-Model Evaluation
A practical framework for comparing reasoning budgets, response speed, task quality, and total completion cost.
How to Vet Anonymous Reasoning-Model Claims
A cautious approach to model identity, tokenizer clues, context-window claims, retrieval tests, and source confidence.
What Reinforcement-Learning Claims Mean for Reasoning Models
A research note on separating training-method claims, observed reasoning behavior, distillation, and reproducible evidence.
A Practical Frontier-Model Evaluation Matrix for 2026
A durable comparison framework based on task quality, reliability, latency, total cost, and operational fit—not a static leaderboard.
A Four-Scope Memory Design for Multi-Agent Systems
How user, agent, session, and organization memory boundaries can reduce leakage and make retrieval behavior easier to reason about.
A Practical Guide to Evaluating Test-Time Compute
How to compare inference-time search and verification without assuming that more reasoning tokens always improve task outcomes.
Evaluating Synthetic-Data Feedback Loops
A review framework for generated traces, automated verification, error filtering, and the limits of self-correction.
Designing Safe Self-Improvement Loops for Software Agents
How tests, scoped changes, review gates, and rollback boundaries can constrain automated maintenance and refactoring.
Model Context Protocol: Designing Safer Tool Connections for AI Agents
A practical introduction to tool schemas, transport boundaries, permission checks, and failure handling in MCP integrations.
Test-Driven Guardrails for AI-Generated Code
Why generated code needs deterministic tests, type checks, reviewable diffs, and explicit evidence before integration.
Where Voice-AI Latency Actually Comes From
A breakdown of audio capture, speech recognition, model response, tool calls, speech synthesis, and perceived turn-taking delay.
Sub-Agents or One Prompt? Choosing the Right Orchestration Pattern
A balanced comparison of specialization, coordination overhead, shared context, review cost, and failure isolation.
Deterministic Validation for Structured LLM Outputs
How schemas, state machines, retries, and explicit rejection paths keep probabilistic output inside application boundaries.
The Cost of Oversized Context Windows
A practical look at retrieval quality, token spend, latency, and when selective context beats sending the entire workspace.