_@baodev

I build backend and AI systems for real operational use — explicit state, bounded decisions, and failure-aware automation.

Most of my work sits at the intersection of backend systems, AI workflows, and operational reliability. I care about what happens outside the happy path: retries, restarts, stale state, partial failure, authorization boundaries, and whether a system can explain why it made a decision.

My main commercial project: starting as a suggestion-first copilot for Shopee sellers, with a staged path toward controlled auto-send once real-shop evidence supports it.

• MVP: suggestion-first — captures the right conversation context, retrieves shop-approved evidence, generates a bounded suggestion, and lets the seller accept, edit, reject, or handle the reply manually
• Direction: controlled auto-send — accept/edit/reject evidence is used to learn which intents can safely graduate toward seller-enabled automation, with per-intent thresholds, canary rollout, kill-switch, and rollback rather than unrestricted model authority
• Grounded and fail-safe — answers from approved facts, FAQ, policy, product specs, and tone; weak evidence, wrong-shop state, stale context, or insufficient authority stops the path instead of encouraging a guess
• Clear trust boundaries — the MV3 extension owns the Shopee browser session, local state, seller-visible UI, and final runtime guard; the Fastify backend owns signed LLM/RAG, knowledge processing, license/config, and never needs marketplace session material

_{TypeScript · Chrome MV3 · Fastify · IndexedDB/Dexie · LLM + RAG}  ·  Active private development  ·  🌐 shoracle.com  ·  🔒 Private source

An installable open-source Agent Skill for marketing work that needs stronger decision discipline: preserve what is already resolved, identify the open decision, load only relevant knowledge, and keep outputs inside the available evidence.

• Decision-first runtime — classifies the current job, freezes approved state, names the unresolved decision, and avoids reopening strategy just because the final artifact is copy, email, a platform post, or a commerce task
• Addressable JIT knowledge — logical knowledge IDs route agents to the smallest useful handbook, platform, commerce, or adaptation section; the current index validates at 252 routes / 225 evidence sources
• Evidence and claim discipline — keeps observation, interpretation, attribution, causality, commercial state, platform signals, local evidence, and product truth distinct instead of silently upgrading weak evidence into stronger claims
• Portable and locally extensible — host-aware guidance covers Claude Code, ChatGPT, Cursor, and Codex, while scoped adaptation units specialize existing decisions without turning the repository into country packs or a cultural encyclopedia
• Research with explicit limits — adversarial reviews, routing checks, and behavioral pilots report what they actually establish; the project records unresolved or non-advantage results instead of treating passing tests as universal proof

_{Agent Skill · Python tooling · JIT Knowledge Routing · Marketing Research · Localization · Evaluation}  ·  v1.2.0 · Building in public  ·  🌐 MIT-licensed open source

A completed full-stack Vietnamese reading and audio platform built around CDN-backed content delivery, realtime features, background processing, and mobile-first UX.

• Static chapter/audio delivery through object storage + CDN while the API serves metadata and application state
• Redis-backed background jobs for batched counters, statistics, reconciliation, cleanup, rewards, and ranking workloads
• Realtime chat/notifications, PWA support, JWT + Google OAuth, rate limiting, and production deployment tooling
• Separate media/content paths for image, JSON, and audio-heavy workloads instead of pushing everything through the application server

_{Vue 3 · TypeScript · Node.js · MySQL · Redis · Cloudflare R2 · Cloudinary}  ·  Completed system  ·  🌐 Public source

An ongoing research track on state integrity in long-lived AI agents: provenance, authority, temporal correctness, durable execution, and the gap between rich runtime traces and governance-relevant interfaces.

• Separates established foundations from agent-specific hypotheses instead of assuming novelty
• Uses falsification-first gates: weak ideas are narrowed or abandoned rather than rescued after the fact
• Current benchmark work studies whether an already-produced action can still be bound to the exact committed external effect occurrence after crash/recovery boundaries
• Freezes claims, implementation, held-out evaluation, and adjudication rules before observing authoritative results

_{Python · Agent Runtimes · Durable Execution · Provenance · Benchmarking}  ·  Research in progress  ·  🔒 Private until release

A paused product/R&D track for finding products with promising source-market signals on 1688 that still face limited competition on Shopee Vietnam.

• Engine pipeline: source crawling → deduplication → translation → Shopee image/market matching → supply enrichment → evidence-backed scoring and decisions
• Opportunity scoring combines demand, visual gap, estimated margin, local-market gap, competition pressure, data confidence, and explicit risk signals
• Separate SaaS consumes a versioned engine contract and provides auth, subscriptions, billing, opportunity exploration, watch/reserve workflows, and market radar

_{Python · Vision / pHash · Next.js · TypeScript · PostgreSQL · payOS}  ·  Paused  ·  🔒 Private source

These are smaller or more inspectable public artifacts that show additional engineering work beyond the primary projects above.

• BAO.OS / Portfolio — interactive portfolio with command routing, multiple RAG modes/personas, Supabase pgvector retrieval, local deterministic embeddings, an OpenAI-compatible API surface, and server-side redaction boundaries; live LLM/RAG is currently disabled on the public deployment.
• CD1-2 — Linux security monitoring lab combining Suricata, Wazuh, deterministic correlation/risk scoring, local IsolationForest signals, evidence-aware explanations, and controlled response policy.
• Audio Ingest — queue-based YouTube audio ingestion pipeline using Redis workers, FFmpeg, object storage, recovery paths, and database synchronization.

Area	Stack
Languages & Runtime
Backend & Security
Data & Processing
AI & Agents
Infra & Storage

_{Open to backend / AI systems roles where reliability actually matters  ·  reach out}