AI Software Engineer · HumAI, Dubai · 2025 —
I design and ship the services behind a conversational CRM assistant, voice agents, an enterprise knowledge base and a customer context graph — LangGraph and DSPy orchestration through to continuous release on Kubernetes — and the benchmarking, regression tests and monitoring that keep language-model systems dependable once they are live.
Principal contributor to Huscribe, HumAI’s AI platform: five production services in Python and FastAPI, orchestrated with LangGraph and DSPy, released continuously to Google Kubernetes Engine. The work on this page spans all of them.
Speak once and Huscribe turns the words into clean CRM data — notes, tasks, follow-ups — inside Salesforce, HubSpot, Dynamics 365, Zoho and Pipedrive. The app is on iOS and Android. The platform behind it is what I build: the reasoning that turns a sentence into the right record, the knowledge base, the customer graph, and the evaluation that keeps all of it honest.
Screens from the public App Store and Google Play listings. Huscribe and Mairit are products of Humai FZCO.
“Log a call with Sara at Al Noor Trading — she wants pricing for forty seats, follow up next Tuesday.”
Automatic field resolution and record normalisation across Salesforce, HubSpot and Zoho, conversational intent routing, and deal-health and lead-scoring pipelines — with per-task model routing so each step runs on the model best suited to it.
A three-stage agent compiles each ingested document into cited, cross-linked pages, and an index-guided retriever answers questions from only the pages it needs — so nothing is asserted without a citation.
Ingests Slack, Salesforce, Zendesk and Shopify, extracts entities through a combined rule-based and LLM pipeline, resolves identities across systems, and opens the result to conversational query.
Led the Arabic and English speech-recognition evaluation: three candidates benchmarked on word and character error rate across standard corpora and real customer dictations. The incumbent turned out to vary run to run and intermittently drop clauses.
Replaced hand-written prompts with automated prompt optimisation across nine coaching agents, scored against the originals.
The voice interviewer for HumAI’s recruitment platform, and the legal-translation agent for Mairit — contracts in PDF, Word and scanned form taken through recognition into structured, reviewable output.
OBAM-AI treats a deployed language-model application as an opaque HTTP endpoint: it builds a benchmark grounded in the project’s own documents, replays it against the live service on a schedule, and scores every reply for groundedness and helpfulness — with no SDK, instrumentation or code change inside the system under test. The working implementation was validated over a 24-day scheduled run against a live deployment.
Began as my final-year project at UET Mardan, where it was built and run end to end.
Huscribe’s CRM reasoning layer, the multi-tenant knowledge base, ContextGraph, the Arabic and English ASR evaluation, the DSPy-optimised coaching agents, the voice interviewer and the Mairit legal-translation agent — detailed above.
Fine-tuned CodeStral-22B to generate error-free PineScript trading indicators, paired with a sandboxed execution loop that fed compiler errors back to the model so it corrected its own output.
Built Bots For AI, a configurable chatbot platform — a knowledge base per customer with automatic content syncing, deployed across web, WhatsApp, Messenger, Instagram, Slack and Salesforce — and a WhatsApp inventory assistant answering product, pricing and stock enquiries from live business data.
Fine-tuned TinyBERT for sentiment classification and shipped it as a web application; built an e-commerce support assistant on FastAPI and LangChain, and an interview-preparation tool that generated questions, scored answers and returned per-answer feedback.
Final-year project: OBAM-AI, the LLM observability platform behind the ICEAI 2026 submission. Degree attested by the Higher Education Commission of Pakistan.
DeepLearning.AI certifications: Generative AI with Large Language Models · Natural Language Processing Specialization · Deep Learning Specialization · Machine Learning Specialization.
Natural language to SQL and MongoDB queries over your own databases — FastAPI, a ReAct agent, streamed responses and similarity-based caching.
meeting-botUpload a meeting transcript and chat with it — FastAPI, LangChain and a small web client, containerised.
rag-essentialsA retrieval-augmented generation reference build with LangChain and FAISS, plus a step-by-step AWS EC2 deployment guide.
finetuning-llmsWorking notebooks for fine-tuning large language models.
realtime-voice-agentA voice agent on the OpenAI Realtime API with a React client; a Deepgram variant lives in a sibling repository.
healthconnect-aiA RAG doctor-recommendation assistant over a doctor database scraped from public sources — LangChain and ChromaDB behind an interactive question-and-answer flow.
Let’s talk about agentic systems, evals — or a role.
Dubai, UAE · GMT+4 · local time --:--
