Vahid Faraji
Vahid Faraji
Senior Applied AI Specialist โ€” Agentic Systems & Context Engineering

Senior Applied AI who designs agentic systems, context engineering workflows, and evaluation-driven LLM automation for enterprise use cases. 8+ years in analytics and data, expert on multi-agent architecture, retrieval pipelines, and cost-aware AI tooling that turn unstructured business tasks into reliable production systems. Hands on experience on enterprise data, agentic develop and cloud infra.

Creator of Tokalator: open-source infrastructure for context optimization and cost-aware agent execution across 15 models. Google-certified Generative AI Leader.

Agentic Systems Design Multi-Agent Design Product Engineer Context Engineering Token Economics Enterprise Data RAG Pipelines DB Vector MCP and Tool Call
agent_arch
Multi-Agent Orchestration
Designed hierarchical agent frameworks (planner โ†’ specialist โ†’ reviewer โ†’ tool) deployed at 20M+ user scale
eval_loop
Evaluation-Driven AI
95% LLM accuracy, 86.2% rule-based segment accuracy, 0.94 avg confidence โ€” measured on 14K+ positions
token_budget
Context & Cost Engineering
Built tokalator.wiki โ€” open-source context optimization toolkit, VS Code extension & caching ROI calculators
prod_impact
Production Impact
30K+ job titles normalized, 72.5% faster classification, 60% fewer manual tasks โ€” shipped to real users
a. Open Source
b. Research
research ACM CAIS '26
Multi-Agent Position Classification

MCP-based multi-agent system for cross-lingual normalization of free-form job titles to O*NET & ESCO taxonomies. Root Content Agent delegates to specialized sub-agents (web search, content extraction, DB queries, trend analysis) with multi-layered memory. Evaluated on 14K+ positions โ€” 72.5% reduction in classification time, 86.2% rule-based accuracy, 95% LLM accuracy (0.94 avg confidence). Accepted as demo paper at ACM CAIS '26.

arch: planner_agent โ†’ [search_agent | db_agent | extract_agent | trend_agent] โ†’ evaluator โ†’ taxonomy_output
MCPยทLangGraphยท O*NETยทESCOยทMulti-AgentยทEvaluation
c. Enterprise
production enterprise
Enterprise AI Transformation

Agentic workflows across Sales (report automation), Finance (text-to-SQL), and R&D โ€” designed, evaluated, and handed off to enterprise operations. AI-powered feedback automation serving 100+ users with 60% reduction in manual tasks. Token-efficient prompt and context engineering adopted as org standard.

arch: task_decomposer โ†’ specialist_agents โ†’ human_checkpoint โ†’ eval_loop โ†’ enterprise_handoff
FastAPIยทLangGraphยท AgnoยทPydanticยทArize
production
Text-to-SQL โ€” Natural Language Reporting

Natural-language to SQL pipeline turning business questions into structured reports. RAG-augmented query generation agent with schema grounding, guardrails for query safety, and human escalation for ambiguous inputs.

arch: nl_parser โ†’ schema_rag โ†’ sql_generator โ†’ validator โ†’ report_renderer
PythonยทSQLยท RAGยทBigQueryยทAzure Synapse
production
Self-Learning Agents

Autonomous agents that improve through feedback loops โ€” evaluation-first design with Arize observability, LangSmith tracing, structured failure recovery, and controlled autonomy boundaries.

arch: executor โ†’ outcome_evaluator โ†’ feedback_store โ†’ self_refinement โ†’ human_gate
LangGraphยทMCPยท AgnoยทArizeยทLangSmith
ACM CAIS '26 demo paper
Tool-Augmented Multi-Agent Systems for Job Position Normalisation
Vahid Faraji et al. ยท Kariyer.net R&D ยท ACM CAIS 2026 ยท Accepted
arXiv 2601.22885
Vahid Faraji et al. ยท arXiv preprint ยท 2026
arXiv 2604.08290
Vahid Faraji ยท arXiv preprint ยท 2026
2025 โ€” Present

Senior Applied AI Specialist

Kariyer.net
Designed agentic reporting pipelines (Sales automation, Finance text-to-SQL) and implemented evaluation loops for LLM-powered workflows. Introduced token-efficient prompt and context engineering standards adopted org-wide. AI feedback automation serving 100+ users, 60% reduction in manual tasks.
โ†‘ Google GenAI Leader ยท Perplexity ilab $40K ยท ACM CAIS '26 first author
2022 โ€” 2025

Senior Data Product Manager

Kariyer.net
Built AI-powered search, agentic workflows, and cost optimization systems for Turkey's largest job platform. Led cross-functional teams delivering data products to 20M+ users. Normalized 30K+ job titles into O*NET & ESCO taxonomies โ€” replaced 6 months of manual work.
โ†‘ 20M+ users ยท 30K+ positions normalized ยท agentic search shipped
2021 โ€” 2022

Business Analyst

WorqCompany
Financial modelling, risk logic, and KPI design for HR-tech products.
2019 โ€” 2021

Co-Founder

Defaro.io
Labor market analytics startup โ€” built the data pipeline and product strategy from scratch.
Google Dec 2025 โ€” Dec 2028

Generative AI, Google Gemini, AI strategy & leadership.

LangChain Jan 2025

LLM observability, tracing, evaluation and debugging with LangSmith.

Techcareer.net 2026
Uygulamalฤฑ AI Araรงlar, Agent'lar ve Otomasyon

Practical AI tools, agents and automation โ€” agentic workflows and enterprise AI delivery.

Agent Systems

  • LangGraph / MCP / Agno
  • OpenAI Agents SDK
  • LangChain
  • Agent Eval & Benchmarking
  • Observability (Arize, LangSmith)
  • Guardrails & Safety
  • Agent Garden (Google)

Languages

  • Python (FastAPI, Async)
  • TypeScript
  • SQL / Text-to-SQL
  • Pydantic

Data & Retrieval

  • PostgreSQL / Supabase
  • Azure Synapse / BigQuery
  • Vector DBs / RAG
  • ETL / Data Ingestion
  • DataHub
  • Metadata Management

Automation

  • n8n
  • Power Automate
  • GitHub Copilot Workflows

AI Dev Kit

  • Docker
  • Azure / Google Cloud
  • Next.js / React (Vercel, V0)
  • VS Code, Claude Code

~ Currently Building

  • โ”œโ”€โ”€ Long-horizon agent reliability patterns
  • โ”œโ”€โ”€ Structured evaluation frameworks for agentic systems
  • โ”œโ”€โ”€ Cost-aware multi-agent orchestration at enterprise scale
  • โ””โ”€โ”€ Tokalator v2 โ€” expanded model coverage + agent-native context APIs

~ Currently Reading / Shipping

  • โ”œโ”€โ”€ Agent eval benchmarks beyond single-turn accuracy
  • โ”œโ”€โ”€ Prompt caching break-even economics across frontier models
  • โ”œโ”€โ”€ Production traces: what trustworthy autonomy looks like
  • โ””โ”€โ”€ Open-weights vs. hosted cost curves for agentic workloads