EP.00Agent online

SALMANZAFAR

Next-gen agentic systems · built for the AGI era

Senior AI engineer and team lead in Dubai. Ten years across IoT, e-commerce and ERP, and now AI. I take LLM-powered systems from prototype to production: agentic workflows, RAG pipelines and multi-agent orchestration. Then I lead the teams that ship them.

Open the character file

Salaam! I'm Munshi, Salman's AI secretary. Ask me anything about his work.

500+
users on enterprise platform
90×
faster critical responses
35%
lower LLM inference cost
10+
years in software engineering
THE PROTAGONIST

Character file

Applied AI engineer who ships agents to real users, then builds the team and the standards around them.

SALMAN ZAFAR
File no. SZ-0001Status: active · Dubai
Name
Salman Zafar
Class
Senior AI Engineer · Engineering Team Lead
Roles played
Engineer · Team lead · Project manager
Story arcs
IoTE-commerceERPAI
Base
Dubai, UAE
Guilds
Publicis Groupe (teos.ai) · promptev.ai (co-founder)
Specialty
Context engineering, agentic workflows, RAG and GraphRAG, multi-agent orchestration, evals
Languages
English · Urdu
Honors
Employee of the Year2018

Abilities

SKILL TREE

Applied AI & Agentic Systems

Multi-agent orchestrationTool / function callingReActPlan-and-ExecuteLLM integrationContext engineeringPrompt engineeringSelf-healing agents

Models & Providers

OpenAIAnthropic / ClaudeGoogleMistralOpen-source modelsDynamic routing

Agentic Frameworks

LangChainLangGraphAutoGenCrewAILlamaIndexSemantic KernelOpenAI Agents SDKAzure AI Foundry

RAG & Retrieval

GraphRAGMultimodal RAGAgentic RAGEmbeddingsPineconeWeaviateMilvuspgvectorHybrid searchRe-ranking

MCP & Multimodal

MCP servers & toolsMultilingual Context Engine / PipelineOCR & extractionAzure Doc IntelligenceClaude visionDocument · image · audio

AI Safety & Evaluation

GuardrailsContent filteringPrompt-injection defenseHuman-in-the-loopAccuracy · safety · latency evals

Languages & Frameworks

Python / FastAPINode.js / ExpressTypeScriptReact.jsNext.js

Data & Infrastructure

PostgreSQLRedisAWS · Azure · GCPDockerCI/CDTerraformPulumiObservabilityRow-level securityRoute-level permissions

APIs & Integrations

REST API design & developmentPayment integration (Stripe, PayPal)ERP integrationsE-commerce platformsReal-time device APIs

IoT & Devices

Mobile device management (MDM)GeofencingRemote lock & trackingOTA update portalSensor data & debuggingReal-time device communication

Performance & Reliability

p50 / p95 / p99 latencyResponse-time optimization (up to 90x)99.9% uptimeObservability

Leadership

Project managementTeam leadership & mentoringTechnical architectureStandards & deliveryStakeholder management
THE STORY SO FAR

Episodes

The career so far, in airing order. Two episodes are still running.

EP.01Aired

Innowi Inc.

Software Engineer

2017–2022 · Lahore, Pakistan

  • Built mobile device management (MDM) for Innowi Checkout, a payment device: remote locking, tracking, geofencing and sensor-data debugging.
  • Built a separate OTA portal to push software updates to devices in the field.
  • Built full-stack web portals and IoT-enabled systems with RESTful APIs for real-time device communication at 99.9% uptime.
  • Employee of the Year 2018.
EP.02Aired

TeqXS

Senior Software Engineer

2020–2022 · Lahore, Pakistan

  • Built Khansaama and Amdan.pk, two startup products that have since wound down.
  • Optimized AWS infrastructure to 99.9% uptime and cut deployment time by 50%.
  • Designed relational and NoSQL data models; partnered with mobile, marketing, and front-end teams.
EP.03Aired

Vfairs

Team Lead / Senior Full-Stack Developer

2021–2022 · Lahore, Pakistan

  • Built a large-scale virtual events platform: virtual event services, exhibitor booths and webinars.
  • Led and mentored a cross-functional team of 6.
  • Architected a maintainable platform at 99.9% uptime with modular REST APIs.
EP.04Now airing

Publicis Groupe

Senior AI Engineer / Engineering Team Lead

2022–now · Dubai, UAE

  • Lead design, build and deployment of teos.ai, the enterprise platform serving 500+ users: job descriptions, headcount, financial forecasts, client vitals and org structure.
  • Built row-level data access and route-level permissions, with workflows and approvals by role and module.
  • Shipped AI agents on the platform: Taskmorph, and TEOS, a self-healing SDLC agent where production errors open their own pull requests for human review.
  • Built RAG/GraphRAG pipelines, evaluation frameworks, and a multimodal Context Engine.
  • Lead a team of engineers, own technical standards and delivery, and drive AI adoption org-wide through Internal Digital Transformation (IDT).
EP.05Now airing

promptev.ai

Co-founder & Lead Engineer

2025–now · Dubai, UAE

US startup registered in Florida, co-founded with a US-based partner

  • Built context-aware AI infrastructure end to end: agents answer from company data and act through 198 tools, limited to what each person asking may see.
  • Open-sourced the context engine (Apache-2.0), with Python and TypeScript clients on one Postgres schema.
  • BYOK by design: customers bring their own models, database and storage, and run it managed, in their own cloud or on-prem.
SPECIAL MOVE

Production errors open their own pull requests.

TEOS is the self-healing SDLC agent I built for teos.ai at Publicis Groupe.

Why deduplication comes first

One bad deploy can throw the same exception 400 times. Without deduplication that is 400 work items and 400 pull requests: a denial-of-service attack on your own backlog. With it, it is one.

  1. 01

    An error hits production

    The error tracker fires a webhook into the agent.

  2. 02

    Check for a known issue first

    Before writing any code, the agent searches Azure DevOps. A known issue gets the error attached to its work item; a new one gets a work item created.

    Known → attachNew → create
  3. 03

    Investigate

    Reads the stack trace, pulls the relevant files, and works out what actually broke.

  4. 04

    Implement a fix

    Writes the change against the repo.

  5. 05

    Open a pull request

    A human reviews it. Nothing merges itself.

MAIN QUEST

promptev.ai

The company I co-founded with a US-based partner, and where I lead the engineering from Dubai. promptev is privacy-first, context-aware AI for teams: every answer is limited to what the person asking is allowed to see. promptev.ai ↗

What I built

  • The context engine: ACL-aware retrieval where the permission check lives inside the search query, so a document someone cannot open is never fetched, ranked or shown to the model.
  • The agent platform around it: connectors, tool calls with human approval gates, and agents that answer in Slack, WhatsApp, a website widget, an API and MCP.
  • A BYOK architecture: customers bring their own models, database and storage, and run it managed, in their own cloud or on-prem.
Five search methods fused into one ranking, all filtered by the same ACL predicate. Open source under Apache-2.0, with Python and TypeScript clients on one Postgres schema.
Semanticpgvector
Trigrampg_trgm
Full-textPostgres FTS
Accent-freeunaccent
Graphoptional
Fused into one ranking
ACL in the SQL predicate
PII masked, sources cited

Open-source work

FREE TO USE
Library · Apache-2.0

promptev context engine

ACL-aware retrieval you can read, audit and run on its own, with no account and no call home.

pip install promptev-context-engine
pypi.org/project/promptev-context-engine ↗
Debugger · v0.3

ctxdiff

A git diff for your agent's context window. Turn-by-turn diffs, token budget by role, schema-bloat detection and cache-prefix breaks. Local-first.

pip install ctxdiff && ctxdiff demo
github.com/salmanzafar949/ctxdiff ↗
WINS

Battle log

Results from production, measured in users, milliseconds and money.

Scale

Enterprise AI platform

Architected and delivered a platform serving 500+ users across 7 modules in the MENA region.

Cost

35% lower inference cost

Cut LLM cost through model selection, dynamic routing, and caching.

Speed

Up to 90× faster

Optimized critical API and process response times by up to 90×.

Efficiency

2.5× less overhead

Reduced process overhead by 2.5× through automation and re-architecture.

Multimodal

Multimodal Context Engine

Ingests and extracts structured data from documents, images, and audio (Azure Document Intelligence + Claude OCR).

Defense

Enterprise-grade security

RBAC, PII redaction, encryption, and audit logging for regulated, sensitive-data environments.

TRANSMISSIONS

Dispatches

Notes from building agents in production. Newest first.

Diff your context, not your promptscontext-engineeringagentsHello, world — why this site existsmetaContext engineering is the job nowcontext-engineeringagentsWhy I built promptev on BYOK (and never stored a customer's data)promptevarchitecturePrompt injection is an agent problem nowsecurityagentsRAG in 75+ languages: what breaks firstragmultilingual
All 14 posts →
JOIN THE PARTY

Let's build what comes after chat.

Open to senior applied-AI roles, technical leadership and agentic systems collaborations.

salmanzafar949@gmail.com
GitHub ↗LinkedIn ↗promptev.ai ↗
© 2026 Salman Zafar · Dubai, UAETO BE CONTINUED