AI Engineer · Full-Stack · TypeScript
Building production AI systems — not wrappers around API calls
I design and ship AI systems with real architectural decisions behind them — structured generation, multi-phase completions, scoring engines, re-scoring infrastructure, and cost-aware LLM usage. My background is full-stack TypeScript + MERN; my current focus is production AI engineering: how systems fail, how they scale, and how to make them explainable.
Currently deepening: LangChain.js · Mastra · LangGraph · Evals & Observability
Replaces generic career counseling with a data-driven, explainable recommendation system built on a custom scoring engine, two-phase LLM completion, and a full re-scoring infrastructure.
The System
Users complete a 9-dimension behavioral assessment (strengths, learning style, collaboration preference, work style, career values). Each submission is scored against 200+ curated career pathways using a custom weighted matching engine with four band multipliers — strong, supporting, weak, and penalty — producing normalized, comparable scores across both single-select and multi-select dimensions.
Recommendations are returned across a 3-layer hierarchy (domain → field → individual pathway), each with a match percentage, dimension-level score breakdown, and an on-demand AI-generated explanation. Explanations are cached after the first generation — no redundant LLM calls on repeat views.
Re-Scoring Infrastructure
A dedicated re-scoring pipeline handles four trigger types: user assessment updates, pathway profile changes, algorithm version bumps, and manual admin actions. Runs as sequential batch processing with full audit output on every execution — the system is always in a known, traceable state.
Roadmap Generation
After pathway selection, Claude generates a phased action roadmap using strict JSON schema enforcement: steps nested inside phases, estimated hours (not vague day counts), and mandatory search queries on every resource reference — specifically to eliminate hallucinated URLs. Phase count is driven by the user's selected timeline tier.
Two-Phase AI Advisor
The embedded AI advisor runs on a two-phase completion architecture:
- Phase 1 — silent, non-streaming 200-token call that detects whether web search is needed
- Phase 2 — if search is triggered, Tavily fires and injects grounded results; streaming begins immediately after with full web context already loaded
The user sees search happening, then text flows — no generation delay after search completes.
Architecture Decisions
- TypeScript monorepo with shared Zod schemas as the single source of truth — assessment schema changes propagate automatically through models, validators, scoring logic, and API response types
- Locale-independent scoring engine — multilingual support (EN/FA/PS) handled via translation Maps embedded in MongoDB documents, resolved at query time with locale fallback; business logic never touches locale
- Three advisor modes: Deep, Focus, Guided — each with different context budget and tool access
Stack: TypeScript · Node.js · Express.js · Next.js · React · MongoDB · Mongoose · Anthropic Claude API · Tavily Search API · HuggingFace Inference API · Zod · REST API Design · Monorepo Architecture
Full-stack social platform built as a monorepo workspace — engineered and project-managed end to end. Includes a Next.js 16 client, Bun + Express API, shared type contracts, and a reusable UI package. Covers auth, session management, feed, posts, comments, reactions, i18n, and scalable backend layering.
Stack: Next.js 16 · React 19 · Bun · Express · MongoDB · Better Auth · TanStack Query · Zod · TypeScript · next-intl
Role-based internal dashboard for managing employees, projects, tasks, costs, and quotations. Built with clean client/server separation, JWT auth with token refresh, and seeded dev data for fast local setup.
Stack: React · Chakra UI · Bun · Express · MongoDB · Zustand · React Query · TypeScript
AI & LLM
Anthropic Claude API OpenAI API HuggingFace Inference Tavily Vercel AI SDK Prompt Engineering Streaming Completions Structured JSON Generation RAG
Languages & Runtime
TypeScript JavaScript (ES6+) Node.js Bun
Frontend
React Next.js React Native Tailwind CSS TanStack Query Zustand
Backend & Data
Express.js REST API Design MongoDB Mongoose PostgreSQL Prisma Zod
Architecture
Monorepo Architecture Clean Architecture Scoring Engine Design Re-Scoring Infrastructure i18n Architecture
Tooling
Git Docker Vercel Vitest React Testing Library MSW
- Shipping AI features in production (chat, text generation, structured generation, search-augmented completions)
- Studying: LangChain.js · Mastra · LangGraph · multi-agent orchestration · evals
- Goal: AI Engineering Lead — own architecture and system design, not just implementation