Ayesh

Ayesh

Agentic AI Engineer | Building Production-Grade, Reliable LLM Systems

Based in: Karachi, PakistanMost recently: Agentic AI & Automation Engineer Intern, BizBuddy

Hire me

CS grad from FAST-NUCES specializing in agentic AI and automation. Proven experience building multi-tenant SaaS agents handling 50+ daily conversations per client at BizBuddy. Expert in designing reliable LLM pipelines with evidence-linked extraction, durable background jobs, and containerized environments.

Experience

Agentic AI & Automation Engineer Intern

BizBuddyKarachi, PakistaninternshipJun 2026 – Sep 2026

Summary

Worked directly with the founding team on production systems and with 15-20 clients on onboarding and training.

What I did

  • Built and maintained AI WhatsApp ordering agents for 15+ restaurants on a multi-tenant SaaS platform (n8n, LangChain, OpenAI/Gemini), handling 40-50 customer orders and conversations per day across queries, recommendations and delivery.
  • Rebuilt a legacy tag-based agent system as modular, tool-based workflows shaped by each client's requirements.
  • Investigated live prompt, session-memory and order-routing failures in production after launch and shipped fixes.
  • Integrated WhatsApp Business API, PayFast, Syrve POS, Blink, Whisper and Gemini Vision via custom API payloads and webhooks.
  • Ran Playwright automated tests across multiple browsers on 10-15 client setups at a time to validate the ordering workflows.

Results

  • Handled 40-50 customer orders and conversations per day for 15+ restaurants.
  • Successfully embedded with 15-20 clients from onboarding through post-launch fixes.
n8nLangChainOpenAIGeminiWhatsApp Business APIPayFastSyrve POSBlinkWhisperGemini VisionPlaywrightWebhooksAPI Payloads

Skills

Technical

Agentic LLM Systems
PythonPython
Tool use orchestration
LangGraph
LangChain
Structured outputs
Claude Code
FastAPIFastAPI
RAG
Cursor
Schema validation
ChromaDB
pgvector
PostgreSQLPostgreSQL
n8n
Zod
Model Context Protocol (MCP)
RedisRedis
Next.jsNext.js
ReactReact
Langfuse
TypeScriptTypeScript
SQLSQL
Zapier
DockerDocker

Projects

Gatekeeper

Lead Developerpersonal
Built a regression gate for a self-improving RAG agent: each lesson drafted from a failure is re-run on a sealed held-out golden set (13 Q&A) with a deterministic exact-match harness and rejected if it breaks any passing answer. Proved the gate with an adversarial lesson rejected in 4 of 4 runs and an ablation where held-out accuracy stayed 13/13 with the gate on but fell to 12/13 with it off; exposed ask, improve and report as MCP tools.
PythonOpenAIChromaDBMCP
View the code

LoopAI

Lead Developerpersonal
Built a LangGraph agent pipeline behind a FastAPI /chat API that uses conversation history for multi-turn intent; ran up to 3 retrieval rounds with LLM-refined queries over embedded document chunks, returning a fixed safe reply when retrieval is weak.
PythonFastAPILangGraphChromaOpenAI
View the code

Nothing

personal
Eight-agent LangGraph app builder, structured-output debate nodes.
LangGraph

RedHanded

personal
LLM-as-judge, strict JSON.
LLM-as-judgeJSON

Resonance

personal
Local-first Spotify history documentary.
Spotify API

COUNSEL

personal
Hybrid BM25 + semantic retrieval, reranking, citation checks.
BM25Semantic RetrievalDockerPostgreSQLpgvectorFastAPI

InquiryOps

Lead Developerpersonal
Built an inquiry-to-lead automation system for service businesses using Next.js and Groq. Features include Gmail OAuth intake, source-linked fact extraction, and owner-approved drafts tied to specific versions. Implemented durable background jobs with idempotent keys and Zod validation to ensure no lead follow-up is missed, even if provider calls fail.
Next.jsTypeScriptSupabasePostgreSQLGroqZodGmail API

Aria Voice Agent

Lead Developerpersonal
Built a 7-step streaming voice pipeline: injection scan, pgvector semantic cache, token-budgeted context builder, gpt-4o-mini intent routing, GPT-4o streaming with an agentic tool loop, post-processing and per-step latency tracing. Kept multi-turn state in three memory tiers (Redis session turns, Postgres user facts, pgvector cache) within a 1,200-token prompt budget; wrapped every external service in a circuit breaker with fallbacks.
PythonFastAPIWebSocketsOpenAIDeepgramElevenLabsRedispgvectorMCPDocker

Education

B.S. Computer Science

FAST-NUCESComputer ScienceSep 2023 – Jun 2027

Recognition

Certifications

Supervised Machine Learning: Regression and Classification (Andrew Ng)

Coursera2024

Introduction to Generative AI

Google Cloud2024

Introduction to Large Language Models

Google Cloud2024

Generative AI Learning Path

Google Cloud2024
Hire AyeshReach out about a role, a contract or a conversation.For recruiters

Ayesh's twin is AI, it can make mistakes.

More people on Kanvis

Browse more peopleCreate your own Kanvis