Daichi Zaha — AI Engineering Portfolio

How I develop with Claude Code

The risk of a coding agent isn't a dramatic crash — it's the quiet failure: the job that silently stopped, the number nobody re-checked, the code nobody reads anymore. My workflow is built to make those loud. The core move is separating the hand that writes from the eye that judges.

01 · FRAME

Grill before build

A requirements interview — one question at a time — until inputs, outputs and out-of-scope are written down. Verify the task isn't already solved.

02 · GENERATE

Agent implements

Subagents build against a spec, test-first. Seams agreed up front so tests land on the critical paths.

03 · JUDGE

Adversarial verify

A separate read-only agent treats the work as assume-broken and tries to disprove it. Not the author grading its own homework.

04 · SHIP

Human checkpoint

Never auto-merge. A pre-deploy hook runs lint / types / tests. I make the call, then the change goes out.

Separate generator from judge. Code-reviewer and an "assume broken" verifier are distinct agents from the one that wrote the code.

Write the limits before an unattended run. Per-run and retry ceilings are declared up front, not discovered mid-thrash.

"Zero" is a claim, not a fact. A self-audit skill checks whether a zero is really zero or a failed fetch, before anything ships.

Watch for silent death. A watchdog scans every scheduled job each morning and surfaces the ones that quietly died.

$ launchd-watchdog  # every morning, read-only
OK  14  /  WARN 2  /  scanning scheduled automations
— a job that stopped firing weeks ago is no longer invisible
    

Systems built — in detail

Company names are withheld; the systems and outcomes are not. These are distinct systems — not features of one app. Several run on a shared platform I architect; others are standalone products. AI-forward systems are marked.

AI systems— the AI-forward work, highlighted

Diffusion image-editing tool

AI core

Problem

A business image-retouch task needed consistent, reproducible edits at volume — too slow and uneven by hand.

Approach

Diffusion-based editing (PyTorch, Diffusers, Transformers, Gradio, Florence-2) with GGUF quantization (Q2–Q8) for 12GB-VRAM machines, Lightning LoRA 4-step generation, JP→ZH/EN prompt auto-translation, and seed / temperature control for reproducibility.

Outcome

2m30s → 41s / image (70%)low- & high-VRAM pipelines

Why it matters

Same diffusion / VLM muscle that got me recruited onto the physical-AI research program below.

AI assistant engine with guardrails

AI core

Problem

Department-facing AI chat had to answer from real data without inventing figures — a hard requirement where a hallucinated number is a business incident.

Approach

A shared FastAPI + SSE chat engine with runtime-enforced guardrails: model output is checked against real values and fabricated numbers are blocked outright, verified with prompt-injection tests.

Outcome

pytest 27/27E2E 7/7injection blocked

Claude Code

Adversarial verification (assume-broken) caught guard-enforcement gaps before they shipped.

Flight-booking AI & actuals analytics

AI core

Problem

Travel arrangement was manual, and real spend wasn't visible to management in any structured form.

Approach

An AI-assisted booking flow plus an actuals pipeline that ingests real application history into BigQuery for management dashboards.

Outcome

5,756 records~¥47.3M tracked

Claude Code

Built and verified data-first — every figure traced back to source rows.

Internal knowledge RAG & assistant

AI core

Problem

Company knowledge was scattered; people couldn't find prior decisions or documents quickly.

Approach

Vector-search RAG over the document base with an AI chat and a feedback loop, kept lean on cost.

Outcome

~¥5,400 / monthfeedback loop

Claude Code

The same RAG pattern I later published as an open-source starter (below).

A multi-department business platform I architect

A shared platform hosting many distinct department systems on one FastAPI + BigQuery + React foundation. Each system below is its own product with its own users; the platform is the base they run on, the auth / permissions layer, and the deployment pipeline.

Systems on the platform

Sales & SEO analytics

Production

Problem

Ranking, ad-spend and sales signals for 29–31 stores were scattered; sales had no single operational dashboard, and rank-tracking config gaps went unnoticed.

Approach

FastAPI + SQLAlchemy on BigQuery (60+ tables), React 19 front end, GA4 / Search Console / ValueSERP ingestion on Cloud Run.

Outcome

172 keywords · 107+ endpoints16,807-line API → 13 routers

Claude Code

Drove the large refactor with a review subagent on every diff; hooks enforced the quality gate.

Data platform & direct-DB migration

Production

Problem

Fragile scraping fed the analytics layer; a five-day silent ETL failure once went unnoticed until wrong numbers surfaced.

Approach

Retired scraping for a direct source-DB connection (cp932-aware), split collection into 9 parallel Cloud Run Jobs, added a 5-axis health check.

Outcome

116 tables / 1,388 colsfull refresh ~30 min9h TZ skew fixed

Claude Code

The silent-failure lesson here became the watchdog in my OSS kit.

Recruiting ATS

A hiring-management system that parses application emails with LLM extraction and auto-registers candidates — turning an inbox into structured hiring data.

Tech: dedicated worktree · 6 ats_* tables + 710-row master seed · LLM email extraction
Status: production, zero-loss verified against live data

Recruiting analytics

Reconciles interview schedules with a separate data source via hashed-name matching to analyze inflow, interviews, hires and re-engagement.

Tech: hashed-name reconciliation · BigQuery
Outcome: 74.8% auto-linking · shipped to 12 users

Inbound reception board

A multilingual front-desk board for foreign-visitor reception, with a confirm-before-submit step and searchable reception history.

Tech: 5 languages · confirm popup · 48-tab permissions
Scale: 28 stores in production

IT cost management

Auto-ingests cloud spend and PDF invoices into one place, subscriptions included, for a single view of IT cost.

Tech: GCP cost auto-ingest + PDF parsing · subscription tracking
Status: production (8 subscriptions managed)

Ad-cost management

Lets a department self-upload ad-spend data and manage it without engineering involvement — fully self-running once shipped.

Tech: self-upload workflow
Status: complete, self-operating

Production-dept system

Handles the production team's assessment, ad-effect measurement and asset management.

Tech: platform module
Status: in operation

SNS analytics aggregation

AI-adjacent

Collects and aggregates cross-platform social-video metrics — including accounts that block cloud-IP scrapers — and makes collection logs visible.

Tech: Apify · residential-IP yt-dlp fallback · launchd · GCS logs · chat alerts
Note: a candidate for standalone SaaS

In-app feedback

Lets users file bug reports directly from the screen they're on, routed to the right place with notifications.

Tech: in-screen capture · notification routing
Status: production

Opportunity & reservation analytics

Opportunity-loss analysis and a reservation board running on the shared platform.

Tech: platform modules on FastAPI + BigQuery
Status: in operation

Standalone products— separate repos, not on the platform

Media-sync automation (Playwright RPA)

Production · standalone

Problem

Store rosters had to be pushed to 18 external media platforms by hand across 23 stores — slow and error-prone, and a wrong write is destructive.

Approach

Parallel Playwright automation with a dry-run → production gate, read-back verification before any write is trusted, Fernet-encrypted credentials, and a DB-keepalive watchdog.

Outcome

name-match 88.95% → 91.0%mapping 97–99%dupes −36%, reversibly

Claude Code

Its safe-write pattern is the basis of an open-source RPA kit I'm extracting.

Engagement bot — cloud migration

Moved a store-engagement bot to a serverless cloud setup and fixed a paging failure caused by an upstream URL change.

Tech: Cloud Run Jobs · GCS FUSE · Firestore · Google-login dashboard
Outcome: 23 stores · under ¥1,000/mo · paging 52 → 780

Internal video auto-editor

AI-adjacent

Automates editing of internal videos — transcription, video processing and LLM summarization.

Tech: Whisper · FFmpeg · LLM
Status: tooling (no published metrics)

Photoshop-automation extension

A Chrome extension that automates image-editing-software operations for the internal team.

Tech: Chrome extension · Photoshop scripting
Status: internal tool

Expense-management hub

A hub for managing IT expenses and costs, connected to the platform's cost data.

Tech: expense hub
Status: internal tool

Research & contract

Agricultural physical AI — robot imitation learning

AI research

The project

A 10-person, grant-funded program teaching a two-armed agricultural robot (grip + cut end-effector) generalizable behavior via a VLM + Diffusion Policy hybrid over a 224-dim action space (14×16 chunk), trained on A100 GPUs. Target: an international journal paper.

My role

Brought in for diffusion-model & VLM implementation know-how from production image work (Diffusers / PyTorch / GGUF / LoRA · Florence-2); literature review; contributing to a study being prepared for journal submission.

Division of work

I own the diffusion / VLM implementation and the paper; imitation-learning specifics, robot control and mechanical / sensor design sit with the core robotics specialists.

Why it matters

Direct evidence my generative-AI practice reaches into embodied, physical AI.

Addiction-risk classification model

Approach

Feature engineering over multi-indicator behavioral data, pseudo-labeling, XGBoost / LightGBM, SHAP interpretation, statistical testing. Dockerized inference, spec grounded in academic literature.

Measured & reported to client (self-built 3-axis harness, ~475k rows)

rank ρ 0.94 (median)SHAP top-5 = 99% of attribution

Verification-first

Built the evaluation harness myself and validated the metrics' honesty — including the gap between predicted-label and true-label precision — rather than citing the flattering self-consistent figure.

Prior role

Google-ads review AI — data pipeline & RAG

Adecco · 3y4m

Problem

Ad-review throughput and consistency across a multinational team; a review AI needed clean training data and retrievable precedent.

Approach

End-to-end on GCP: annotation design, cleaning / categorization, training-set packaging, and a review RAG (vectorized precedent, Gemini / Cloud Run).

Outcome

60 min → 5 min (12×)~400 hrs/month saved

Foundation

The GCP data-to-training-to-RAG pipeline that underpins my current work.

Open source

The strongest answer to "can you actually use Claude Code?" is public code. Every repo below was genericized from real engineering work — the techniques are mine, the business logic and credentials are stripped out entirely.

claude-code-ops-kit

Operating machinery for running AI coding agents semi-autonomously without silent rot: a launchd watchdog, discipline skills (grill / self-audit / systematic-debugging), an assume-broken verifier subagent, and the Loop Engineering write-up.

MIT · github.com/daichi-0818/claude-code-ops-kit

safe-rpa-kit

Safe-by-default browser automation: dry-run default, a fail-closed write gate, and read-back verification — so RPA writes are impossible by accident, off by default, and not trusted until re-read.

MIT · github.com/daichi-0818/safe-rpa-kit

silent-failure-checks

Lightweight declarative data-quality checks specialized for silent failures: empty-write-disguised-as-fresh, regression-vs-chronic, and two-source reconciliation.

MIT · github.com/daichi-0818/silent-failure-checks

fastapi-llm-rag-starter

Production-ready FastAPI + RAG starter with pluggable LLM providers (Claude / Gemini), ChromaDB, and offline-runnable tests. CI green.

MIT · CI green

idempotent-backfill-pattern

Daily sync and weekly backfill that converge on one dedup-by-latest upsert, so re-running any range is safe.

MIT · github.com/daichi-0818/idempotent-backfill-pattern

fernet-versioned-secrets

Versioned Fernet secret encryption (enc:v1: prefix) with backward-compatible plaintext passthrough for gradual, low-risk rollout.

MIT · github.com/daichi-0818/fernet-versioned-secrets

cloud-run-jobs-etl-scaffold

One Docker image, N Cloud Run Jobs dispatched by an env var — a scaffold for parallel ETL with Scheduler wiring and the env-reset gotchas documented.

MIT · github.com/daichi-0818/cloud-run-jobs-etl-scaffold

phash-dedup

Perceptual-hash near-duplicate image clustering: DCT pHash + Hamming threshold + Union-Find to collapse duplicates to one representative.

MIT · github.com/daichi-0818/phash-dedup

Stack

Agent dev

Claude Code Max — custom skills, role-separated subagents, PreToolUse / PostToolUse hooks, MCP, parallel agents · Codex · Cursor

Language

Python (FastAPI, asyncio, SQLAlchemy 2.0 + Alembic, pydantic v2) · TypeScript · Node (ESM)

Cloud

GCP — Cloud Run, Run Jobs, Scheduler, BigQuery, GCS, Secret Manager, Firestore, Cloud Build, Vertex

ML / DL

PyTorch, Diffusers, Transformers · XGBoost, LightGBM, SHAP · data prep → training → evaluation on GCP

AI / LLM

Claude / GPT / Gemini APIs — Batch, Structured Outputs, Prompt Cache · RAG, ChromaDB, embeddings, HITL

Automation

Playwright (parallel, 18 platforms), Apify, BeautifulSoup

Frontend

React 18/19, Vite, TailwindCSS, React Query, Recharts

Quality

Docker, GitHub Actions, pytest / Vitest / Playwright Test, pre-commit, git worktree