Applied AI work for businesses: find the real problem, prototype the risky assumption, measure it, then take it to production. CTO of Mindreader, an AI sales intelligence platform used by teams at AIA, CIMB and Manulife, and of The Modern Shrine. Past work with Rakuten, Google Crowdsource × Explore ML and Omdena.
Site: tonytonggg.com · LinkedIn: tonytongtszhin
| Work | What was measured | Code | Stack |
|---|---|---|---|
| GraphRAG vs hybrid RAG on 13,000 private notes | Graph alone lost to hybrid search (p = 0.04); graph on top of hybrid helped time-based questions | graphrag-vs-hybrid-rag-eval | Python · Neo4j 5.26 · Claude Haiku 5.5 · bge-m3 on Ollama · BM25 + RRF |
| Evals for an AI web-research step | Wrong matches cut from 12 in 60 to 2 in 60, and 0 in 23 on a fresh holdout | llm-research-eval-harness | Python · pytest · JSONL fixtures · Claude Haiku, Sonnet, Opus and Fable |
| Benchmarking The Modern Shrine's AI agent | 36.7% against a 25% random baseline, with a ranked list of fixes | mingli-bench-harness | TypeScript · Vercel AI SDK · Vercel AI Gateway · Zod · YAML agent configs |
| Gated LLM pipeline | Code owns the flow; a cheap classifier gates each step and only unsure items reach the LLM | gated-llm-pipeline | Python · Jev 1.13.0 · Claude Opus 5.5 |
| Agent approval queue | Agents submit drafts; a human approves, edits or rejects before anything is sent | agent-approval-queue | Next.js 16 · TypeScript · Supabase · Postgres RLS · Vitest |
| LLM provider resilience | Fallback chains, credential pools and circuit breakers, with outage runbooks | llm-provider-resilience | Python standard library · OpenAI-compatible APIs · Ruff |
| From a moving brief to 20 years of trusted macro data | 192,971 rows from a client brief that kept moving; monthly data found 29.4 days early, and the health check that now catches it | macro-regime-data-pipeline | Python · SQLite · FRED API · yfinance · CoinGecko API |
| Scheduled AI jobs for cents | Jobs moved off a US$99-a-month service; each run now costs cents | scheduled-claude-jobs | Bash · launchd · Claude Code CLI · ShellCheck |
| An AI quality gate for every article | A gate measured against the author's own verdicts, with a ledger that forces re-scoring after edits | article-quality-gate | TypeScript · TypeSafe SDK · Jev 1.13.0 · gray-matter · Git hooks |
| Sayura voice app | Shipped to the App Store: speech-to-text, a fast LLM with a strong fallback | sayura-app | Expo / React Native · Groq llama-3.1-8b-instant · Azure OpenAI gpt-4o · AssemblyAI · Firebase |
| Persona sales copilot | Mindreader demo for bank relationship managers, on synthetic data | persona-sales-copilot | React · Vite · Tailwind CSS · TypeScript · gpt-4o-mini |
| BERT serving API | A 2021 Mindreader serving stack, reviewed: tokenizer reload cut from 12.3 ms to 0.11 ms per request | bert-tfserving-api | Flask · Gunicorn · TensorFlow Serving 2.15 · Docker · Hugging Face tokenizers |
| Fully private AI for a 50-person firm | Service levels and a phased study before building the app | — | RAGFlow · Ollama · qwen3.5 · bge-m3 · Docker |
| Live UAT with insurance advisers | Generic AI insight turned into a role-specific coaching workflow | — | Live UAT · adviser interviews · AI role-play · Gamma Slides · onboarding workshops |
| Google Crowdsource × Explore ML | Machine learning taught to 900 non-technical learners across Asia | — | TensorFlow Playground · Teachable Machine · Quick, Draw! |
Earlier work: twitter_mbti_ml (2017, predicting personality type from social posts) and mbti_classifier_streamlit (2024), the text-classification thread that led to Mindreader · footprint (Berkeley hackathon, 2017).
Python · TypeScript · Next.js · LLM evals · RAG and GraphRAG (Neo4j) · Claude and OpenAI APIs · MCP · local models




