Saram Hai Hire me

Senior AI Engineer · Karachi · open to remote

I build AI agents that talk, reason and run in production.

Four years of shipping multi-agent systems, real-time voice agents and RAG pipelines. As Head of AI Engineering at The Botss I've delivered for 30+ clients in legal, healthcare, e-commerce and enterprise across the US, Europe, Australia and the Gulf.

See my work
4+
years in production AI
30+
client engagements
9+
production AI systems
~58%
LLM cost cut on Arthur
Saram Hai in a grey suit, standing in an office
Hi, I'm Bo, Saram's AI sidekick. Move your cursor, I'm watching. 👀

How I work

I take messy business problems and turn them into agents that talk, reason, take payments and book meetings. Then I make them cheap, fast and reliable enough to run every single day, for real customers, in production.

01 / Selected work

Systems people actually use

Agents, voice and document AI, built end to end and run in production.

Multilingual AI agent with voice~3,000 users

Arthur

An agent that speaks English, German and Russian, takes voice calls, books events and handles Stripe payments across separate production instances. I split a 6,000+ line monolithic prompt into layered sub-agents and made every system prompt editable by the client from an analytics dashboard.

98.5%prompt cache hit rate
~58%lower inference cost
2providers: OpenAI and Gemini
  • OpenAI Agents SDK
  • ElevenLabs
  • Twilio
  • RAG
  • Gemini
  • PostgreSQL
AI legal intakeLive on the App Store

Accident Assist AI

The agent interviews US accident victims, builds structured case files through tool calling and routes qualified leads to law firms. I also built the admin backend and moved it to a hardened, client-owned AWS environment.

  • OpenAI Agents SDK
  • FastAPI
  • PostgreSQL
  • AWS
Document AI · PII redactionAustralia

RedactIQ

Classifies, renames and redacts PII in council planning documents, sized for ~25,000 documents a year. Inference is pinned to Vertex AI's Sydney endpoint to keep data in Australia.

  • Gemini on Vertex AI
  • AWS Sydney
Multi-agent legal AIPakistan

EzQanoon

Nine jurisdiction-specific agents (Federal, Punjab, Sindh, KPK and more) with RAG over statutes and judgments, agent handoffs, tool calling and streaming responses.

  • OpenAI Agents SDK
  • Vector Store RAG
Voice ordering agentRestaurants

Speedoy

A phone and web voice agent that takes takeaway orders live and sends them to the restaurant dashboard through secured webhook tools.

  • ElevenLabs Conversational AI
  • FastAPI
  • Webhooks
Omnichannel commerce agentShopify App Store

Commify AI

A shopping assistant on WhatsApp, Instagram, Facebook and the web. It recommends products, adds to cart, runs cash-on-delivery checkout, raises support tickets and hands off to a human.

  • OpenAI
  • WhatsApp Business API
  • Meta APIs
  • Shopify
PPT to narrated videoSaaS

Audifyer

Multi-tenant SaaS with JWT/RBAC, multi-provider TTS with fallback, voice cloning, SRT/VTT captions and in-browser MP4 rendering. I was also PM and QA lead.

  • FastAPI
  • PostgreSQL (RDS)
  • ElevenLabs
  • Gemini TTS
WhatsApp automationKSA · Canada · Pakistan

Inquiry & ops bots

Lead bots for real estate (Estraha), immigration (CanVisas) and automotive (Suzuki Pakistan), plus VendorSync AI, a building issue tracker where tenants register over WhatsApp.

  • WhatsApp Business API
  • OpenAI
  • Express
  • Prisma

Live demo · voice AI

Don't just read about it. Talk to it.

Bo is a real-time voice agent running on Gemini Live. Ask about Saram's work out loud, and interrupt whenever you like. This is the kind of agent he builds for clients.

Tap to talk · uses your microphone · up to 5 minutes

  • “What has Saram built with voice AI?”
  • “How did he cut LLM costs by 58%?”
  • “Is he open to remote work?”

02 / Experience

Where I've built

  1. Jul 2023 to now

    Head of AI Engineering · The Botss

    Karachi · clients in the US, UK, Europe, Australia and the GCC

    • I own AI architecture and hands-on builds for 30+ client engagements: model selection, agent design, evals and production rollout.
    • I turn business problems into agent designs, tool specs and evaluation criteria, working directly with clients.
    • I cut LLM inference cost by ~58% on a ~3,000-user multilingual agent with prompt caching and a multi-provider layer.
    • I wired Stripe, Shopify, GoHighLevel, Airtable, Google Calendar, WhatsApp, Messenger and Telegram into agent tool calls.
  2. Dec 2024 to Feb 2025

    AI Engineer · SOC Solutions LLC

    Wyoming, USA · remote contract

    • Owned AI brand tagging for Social Tip: detects every logo in a photo, then generates hashtags and tags the brands.
    • Built a pipeline that crawls a website and has an LLM fill in government grant applications from what it finds.
  3. Dec 2023 to Dec 2024

    Machine Learning Engineer · Neural Internet

    USA · remote

    • Designed a decentralised RAG chatbot on Bittensor handling 1,000+ messages a day across Discord and Telegram.
    • Ran 30+ subnet validators and built a real-time monitor reading validator logs directly from the nodes, which helped lift APY by 2%.
  4. Nov 2022 to Dec 2023

    Machine Learning Engineer · Automaatia

    Karachi

    • Built an AI news anchor that turns one photo and a script into a lip-synced presenter video, using Wav2Lip, First Order Motion Model and MMOCR.
  5. 2021 to 2022

    Research Intern · MERL  /  AI Intern · Kwick Doctor (NIC)

03 / Toolkit

What I work with

Agentic AI & LLMs

OpenAI Agents SDK, Google ADK, LangGraph, LangChain, CrewAI, Pydantic AI, handoffs, tool calling, structured outputs, prompt caching, MCP

Models

GPT-5 family, Gemini 3 on Vertex AI, Claude, Grok, DeepSeek, Qwen, Kimi, Llama, Falcon

Voice AI

ElevenLabs Conversational AI & voice cloning, OpenAI Realtime API, Twilio Voice/SMS, Gemini TTS, STT/TTS pipelines

RAG & data

Hybrid retrieval, OpenAI Vector Stores, Qdrant, pgvector, Pinecone, Weaviate, Chroma, crawling and scraping pipelines

Evals & LLMOps

Golden-set evals, regression testing, pytest, observability, cost and latency optimisation, context window management

Backend & cloud

Python, FastAPI, async SQLAlchemy, Node.js, TypeScript, Prisma, PostgreSQL, AWS, GCP, Azure, Docker, GitHub Actions

Integrations

WhatsApp Business API, Meta, Telegram, Slack, Discord, Stripe, Shopify, GoHighLevel, Airtable, Zapier, Make, n8n

ML & vision

PyTorch, TensorFlow, Hugging Face, scikit-learn, OpenCV, Stable Diffusion, Wav2Lip, OpenMMLab, logo & object detection

Education

BS Artificial Intelligence, Sindh Madressatul Islam University, 2020 to 2024

Certification

Machine Learning & Deep Learning (6 months), NAVTTC, 2022

Community

Microsoft Learn Student Ambassador; GDSC and Notion community contributor

04 / Writing

Notes from production

22 Sep 2026 · agents · architecture

Stop building one big agent: the supervisor pattern for production LLM systems

One agent with twenty tools looks clean in a demo and falls apart in production. Here's the supervisor pattern I use instead, with Semantic Kernel code.

Read the post →

05 / Contact

Have an agent that needs to actually work?

I'm open to remote roles and project work in any timezone.