$ whoami

Hi, I'm Saeed Mahdian

I work as a Senior Data Engineer

I turn messy, scattered data into reliable pipelines and platforms that teams can trust, from raw sources all the way to a single source of truth.

See my work Say hello

Currently: Data Engineer @ Mofid Securities

SELECT * FROM saeed LIMIT 1;

scroll ↓

01 / impact

Impact in numbers

Hover or tap a number for the story behind it.

02 / career

Career pipeline

My career as an Airflow DAG run: an industry lane and a teaching lane. Hover the tasks, poke them, and click one to read its logs.

dag_id saeed_career schedule @lifelong tasks 0/10 running

201020122014201620182020202120222023202420252026

industry

Mofid

Mofid Securities

Divar

Balad

Mahsan

teaching

Allameh Helli

Teaching

YouTube

Data Engineer Data Scientist AI Engineer Software Engineer teaching

[2025-12] task_id=mofid running

Data Engineer @ Mofid Securities

Dec 2025 – Present

Built the company's first unified text-data platform on Elasticsearch and Kibana, consolidating four sources (call transcripts and chat logs from MongoDB, customer SMS messages and stock-market search events) into one denormalised schema enriched with customer and trading data.
Designed Airflow ETL pipelines that join session metadata with transcripts, rebuild each conversation as a chronological agent–customer dialogue and load flat, search-ready documents; backfilled a full year of history.
Replaced N+1 per-record ClickHouse lookups with a single bulk fetch and in-memory mapping, cutting a batch enrichment job from 70 minutes to under 5 (>92% faster).
Secured Elasticsearch and Kibana with Active Directory (LDAP) sign-in, role-based access control and field-level security masking PII such as national IDs, plus Filebeat-based audit logging.
Deployed a 3-node Elasticsearch cluster with Ansible and Docker, then right-sized shard and index settings after a production capacity review to cut memory overhead and inter-node traffic.
Refactored the shared ETL library used across Airflow DAGs, fixing database connection leaks and redundant engine creation, and extended the customer cash-flow pipeline with new transaction types.

ElasticsearchKibanaApache AirflowMongoDBClickHouseAnsibleDockerPythonLDAP

Education

2010 – 2013 M.Sc. Artificial Intelligence Iran University of Science and Technology

2005 – 2010 B.Sc. Computer Engineering Isfahan University of Technology

03 / work

Selected work

A few problems I'm proud of solving.

Mofid Securitiesfeatured

Mofid's first text-data platform

4 text sourcesEnrichElasticsearchRBAC + maskingKibana

Impact: One secure, searchable place where every team can read what customers say, write and look for, next to who they are.

Problem

Customer conversations, messages and searches were scattered across systems in different shapes. Nobody could search across them or see them next to the customer they belonged to, and they contained sensitive personal data.

Approach

Built the company's first unified text-data store on Elasticsearch, with Kibana on top. It currently holds four sources: call transcripts and chat conversations, both enriched with customer data; SMS messages sent to customers; and customers' stock-market searches, enriched with their trading data. Access goes through Active Directory, with RBAC and field-level security that mask PII such as national IDs.

ElasticsearchKibanaMongoDBClickHouseLDAPRBACPython

ingest(sources=4)

extract(calls, chats, sms, searches)→ 4 sources
enrich(customer, trades)→ joined
index(elasticsearch)→ searchable
secure(rbac, mask_pii)→ national_id → ••••
explore(kibana)→ dashboards

# illustrative run, not real data

Mofid Securities

70 min → 5 min: killing N+1

BatchBulk fetchIn-memory mapEnrich

Impact: Over 92% faster, with batch time dropping from 70 minutes to under 5.

ClickHousePythonPerformance

Mofid Securities

Owning the production Elasticsearch stack

ProvisionSecureAuditCapacityRe-shard

Impact: A secure, observable cluster with much lower shard overhead on heap and less traffic between nodes.

ElasticsearchKibanaAnsibleDockerFilebeatLDAPCapacity planning

Divar

One source of truth for CRM data

CRM sourcesPySparkQuality checksMaster table

Impact: A reliable single source of truth for cross-functional analytics, with far fewer downstream errors.

PySparkApache AirflowData QualitySQL

Divarfeatured

Ad-locator: an AI agent that finds your ad

User questionFind adPick keywordsSearchScreenshot

Impact: Significantly fewer ad-visibility tickets, and users get an answer they can see for themselves.

Problem

A steady stream of support tickets came from users asking "why can't I see my ad?". Each one meant a support agent searching by hand.

Approach

An autonomous LLM agent identifies the user's ad, chooses the keywords most likely to surface it, builds an optimised search query, then replies with a direct link and a screenshot showing exactly where the ad appears.

LLMAI AgentsPydantic AIPythonTool calling

agent.run("where is my ad?")

find_user_ads(user)→ 1 active ad
pick_keywords(ad)→ 3 keywords
search(query, city)→ found · page 1
screenshot(result)→ proof.png
reply(link, image)→ done

# illustrative run, not real data

Balad

Cleaning the map with NLP

POIsNLP matchConfidenceMerge

Impact: More than 100,000 duplicate points of interest removed.

NLPPythonSpark

Balad

Notifications people actually open

POI activityETLSegmentsPersonalised push

Impact: A 28% increase in user engagement.

ETLPythonSparkPersonalisation

Divar

Observability for data-intensive services

DecoratorsPrometheusGrafanaAlerts

Impact: Troubleshooting became standard, fast and measurable.

PrometheusGrafanaObservability

Baladfeatured

Restaurant recommendations on the map

LocationSimilar usersSimilar placesHybrid rankTop picks

Impact: Personalised restaurant suggestions on a map with over 10 million weekly active users.

Problem

Balad's map knows thousands of restaurants, but a long list sorted by distance doesn't tell anyone where they'd actually enjoy eating.

Approach

A hybrid recommender blends collaborative filtering, which uses what similar users liked through a user-relationship graph, with content-based similarity between places (cuisine, price, style). Both are built from similarity matrices produced by ETL pipelines.

Recommender SystemsCollaborative filteringGraphsPythonSpark

04 / stack

Stack

The tools I reach for, shuffling around like puzzle pieces. Hover one to see where I used it.

Data & Orchestration 11

Languages & Frameworks 10

Infrastructure & Observability 8

AI & MLOps 7

Used at → Mofid Securities Divar Balad Mahsan Allameh Helli High Schools Teaching YouTube · Python as Candy

05 / teaching

I also teach.

I started teaching programming at Allameh Helli, a school for gifted students, in 2010 while I was doing my master's. I taught there full-time for a few years and kept a part-time class going until 2019, even after moving into industry. Since 2021 I've taught classes from time to time, and I put free Python lessons on YouTube.

2010first class

4-yearcurriculum I designed

FreePython lessons on YouTube