Vahid Faraji works in applied AI.


Maximise ROI. Minimise cost.

I build workflows, data pipelines and agents that teams use in their daily work. I write the spec, build the tools the agent can use, and measure whether it actually helps: evidence more than demos.

VF · scoreboard2025 against 2026 →
Applied AI, Simplified · companion to my book

Applied AI, simplified.

We map, read, explain and source applied AI: the jobs, the tools and the docs, said plainly. Free, in five languages.

ConceptConfidentialLLM PipelineOct 2026
Yapay Zekâ İzi: AI in 1.6M Job Ads
Problem, approach, result
ProblemHow fast is AI really entering Turkish job ads since the GPT moment of 2022, and which tools do employers name? Keyword counts cannot tell a requirement from boilerplate across 1.6M ads written in free Turkish text.
ApproachThe design of an LLM pipeline over 1.6M kariyer.net ads, from 2022 to the first half of 2026: sentence-level reading of every ad, generative AI kept apart from classic AI and machine learning, tool names normalised so mentions can be counted, and a concept for an internal analysis site, Yapay Zekâ İzi (“the trace of AI”), with periods, tools, sectors, a validation page and the method. Every quote on a card is taken from the ad verbatim.
ResultDesigned to give a clear, period-by-period picture of how fast generative AI moved from novelty to requirement in Turkish job ads, and which tools employers actually name, with every finding traceable to the sentence it came from.
23.3× GenAI ads 1.6M ads in scope 2022→2026 H1
LLM Pipeline·Turkish NLP·Classification·Validation·Data Product
EnterpriseShippedAug 2026
ProjectLens: 450 hours saved across departments
Result450 hours saved across departments in total; pricing a project went from 2 hours to 10 minutes. 91% of the manual effort eliminated. 64h of hand work reduced to 5.5h, freeing 59 hours (7.3 working days). 31 Excel files no longer created.
Problem, approach, result
ProblemEvery project quote started life as a fresh Excel file. Effort estimation, pricing, proposal drafting, approval and follow-up lived in spreadsheets and inboxes. No shared history, no auditable numbers, no way to ask the data a question.
ApproachBuilt with the Pricing & Revenue team: one tool that runs request intake → effort calculation → pricing → proposal → approval → follow-up end to end. Wired into Microsoft Teams and Azure DevOps so requests arrive where the team already works. Ships a read-only SQL console (SELECT/WITH only, every query in a rolled-back transaction on its own connection) plus a presentation mode that masks figures during screen shares.
WorkBeforeAfter
Pricing a project2 hours10 minutes
Collecting a request30 minutes5 minutes
Arithmetic and copy errorsYesDesigned out
Waiting in an approval queueYesNo
Cycle time, conversion, marginNot measuredMeasured

Pricing alone: 31 runs × (2h − 10min) + 4 intake requests × (30min − 5min) ≈ 59 hours. The 450-hour total covers every department that uses the tool. The 2h manual baseline is a team estimate from 2026-07-30, not a measurement.

Applied AI·Microsoft Teams·Azure DevOps·SQL·Workflow Automation
EnterpriseProduction
Enterprise AI Transformation: 60% less manual work for 200+ colleagues
Result60% less manual work for 200+ colleagues. Adopted as group-wide AI standard.
Problem, approach, result
ProblemManual reporting workflows across Sales, Finance, R&D consuming analyst time with no self-service capability.
ApproachMulti-agent orchestration (LangGraph, MCP, Agno) for text-to-SQL, report automation, and AI feedback loops with human checkpoints.
LangGraph·MCP·Agno·Azure AI Foundry
Data ProductText-to-SQL
Metadata for text-to-SQL: 300M+ data points unified
Result300M+ data points unified. Became the foundation for later AI and LLM work: embedding generation, RAG retrieval, and context grounding at scale. It also set the structure for text-to-SQL as an internal product.
Problem, approach, result
ProblemText-to-SQL is only as accurate as the metadata under it. No unified metadata layer across a live job platform. Fragmented schemas, no lineage, analytics bottlenecked by ad-hoc queries.
ApproachDesigned metadata architecture with DataHub governance, real-time pipelines, and product-analytics instrumentation for strategic decisions. On top of it, the text-to-SQL concept: measured on BIRD-bench, structured as an internal product, tuned for accuracy and checked in user tests.
Text-to-SQL·BIRD-bench·DataHub·Azure Synapse·BigQuery·Real-time Pipelines·Retool
More at Kariyer.net
  • TaxonomyCV extraction and matching, built on the job taxonomy.
  • Salary insightsExpected-salary calculation at scale, with outlier detection and a database stream. Built with the BI and Data Platform teams.
  • Sales Coach reportingPower Automate workflows that send each person a personal sales report, 200+ people every month, so Sales Coach can scale. On Medium →

Research profile, papers and data →

ACM CAIS '26peer-reviewed
Vahid Faraji et al. · ACM Conference on AI and Agentic Systems, San Jose 2026
arXiv
Vahid Faraji et al. · arXiv 2604.08290 · April 2026
arXiv
Vahid Faraji et al. · arXiv 2601.22885 · January 2026
Ships to productionTools in daily use, not demos: ProjectLens, Text-to-Report, finance checks.
Proves the impact91% less manual work, a 40% smaller backlog, up to 95% accuracy, all measured.
Brings people along400+ people taught, C-level AI training, career mentoring at TEV.
Owns it end to endProblem, spec, build, evals, hand-off. Research to roadmap.
2025–now

Senior Applied AI Specialist

Kariyer.net (ilab group)

I lead applied AI across Sales, Finance and R&D: from the business problem to agents running in production.

  • ProjectLens: pricing moved from Excel to one tool, 91% less manual work, 31 spreadsheets retired
  • Text-to-Report agents answer data questions in plain language: analytics backlog down 40%
  • My context-engineering framework became the company standard, cutting token spend while improving accuracy
  • Finance exception detection across statements from 4 major Turkish banks
  • Secured a USD 40K Perplexity enterprise grant
  • Google Generative AI Leader certification
Claude CodeAgent SDKLangGraphMCPLangfuseAzure DevOps
2022–2025

Product Manager, Data

Kariyer.net

I built the data foundation the AI work now stands on.

  • Metadata design for 300M+ data points, so every number has an owner and a trail
  • Real-time labour-statistics pipelines with data quality, governance and executive dashboards
  • Led AI adoption training for C-level executives
DataHubBigQuerySynapseGovernance
2021–2022

Business Analyst, Data Analytics

WorqCompany

I gave leadership numbers they could plan with.

  • Financial forecasts and revenue simulations used for mid-range planning
  • Dashboards and process automation for finance and operations
ForecastingSQL
2019–2021

Co-Founder

Defaro.io · backed by Özyeğin University

I built a data product from zero: labour-market and compensation analytics.

  • Data pipelines, benchmarking models and product strategy from scratch
  • ~1BTokens with Claude Code58.0B
  • 83Public commits1,136
  • 1Pull requests198
  • 0Papers3
  • 2Talks and courses3
The full scoreboard, with the detail behind each row →

Enterprise solutions you can count on.

What I bring
Tools I’ve built with
Applied AI
  • Claude Code
  • Claude Agent SDK
  • Copilot Studio
  • Azure AI Foundry
  • Vertex AI
  • LangChain
  • LangGraph
  • Deep Agents
  • Agno
  • ElevenLabs
  • Vapi
Data & retrieval
  • RAG
  • Vector DB
  • BigQuery
  • Google Cloud
  • SQL
  • Text-to-SQL
  • DataHub
  • Hugging Face
  • Perplexity AI Search
Product engineering
  • Python
  • TypeScript
  • Next.js
  • Vercel AI SDK
  • Microsoft Agents SDK
  • Azure DevOps
  • MCP
  • Langfuse
  • Power Automate
  • Zapier
Pencil drawing of a wooden abacus with beads counted along its rods
Fig. 1 Counting before computing. Graphite on paper by Homa Mofidian.
The problem I solve

Messy text in,
decisions out.

Eight years in industry, alongside four years of research, on one problem shape. The domain keeps changing. The problem doesn’t.

Shipped in Talent Pricing Finance Sales R&D Developer tools Data infrastructure
How I got here
01 · Languages

Reading people

Turkish native, English C1, Spanish at teaching level, Portuguese and Japanese on the way. Living in several languages taught me that meaning sits in context, not in the words.

02 · Data products

Structuring it

Defaro.io, then Kariyer.net: 300M+ data points under one metadata layer, with governance and lineage. A number is only as good as the trail behind it, so I built the trail.

03 · Applied AI

Letting models read

Agents that map job titles in five languages to O*NET and ESCO, a Turkish skill-extraction benchmark, Tokalator for the context budget. The model reads the language; the system keeps the lineage.

Same problem, a new reader. Languages taught me to read people, data products taught me to keep the evidence, and applied AI lets both run at scale.

Education

03
  • M.A. Economics, Ibn Haldun University
  • Thesis: Real-Time Labor Market Analysis
  • Teaching assistant, Econometrics I

Credentials

09
  • Claude Certified Architect: registered, on the way
  • ACM Professional Member
  • Google Generative AI Leader
  • Anthropic: MCP, Agent Skills, Claude API, Claude Code
  • LangChain: LangGraph, LangSmith

Languages

05
  • Turkish: Native
  • English: C1
  • Spanish: teaching level
  • Portuguese: Basic
  • Japanese: Learning

Teaching

03
  • 18h Applied AI Course, Techcareer
  • AI Career Mentor, TEV Foundation
  • Enterprise prompt engineering, university lectures

Build

Agent systems that survive production.

What that looks like
  • Multi-agent orchestration on MCP, LangGraph or Agno
  • Text-to-SQL and report automation with human checkpoints
  • Evaluation harnesses and observability before launch, not after

Advise

Architecture review before the spend.

What that looks like
  • Context and token strategy: where the budget actually goes
  • Build-versus-buy calls across the agent tooling landscape
  • Data and metadata readiness for RAG and grounding

Train

400+ people trained, 50+ courses.

What that looks like
  • Applied AI for engineering and data teams, 8–18h formats
  • Agents, MCP and tool design, hands-on
  • AI-native development workflow adoption
Vahid Faraji CV 2 pages · PDF · updated with every deploy
CV page 1 of 2: summary, experience and selected projects CV page 2 of 2: projects, technical skills, publications, talks and education
Vahid Faraji Animated CV 01 Who I am

Suggest a time

Pick a slot in your time zone. I'll confirm within 24h.

My time --:-- · You --:--

Your name, email and topic go to my Google account only so I can reply. See Privacy.

or book directly on my calendar