- a.Built autonomous agents — smart agentic systems using open-source and closed-source LLMs that handle complex workflows and make decisions automatically.
- b.AI strategy & problem solving — identified the real business challenges clients faced and engineered the AI/ML solutions to fix them.
- c.Production-grade deployment — built and shipped AI models using FastAPI and microservices, running smoothly and scaling well on the cloud.
- d.Data pipelines — collected and cleaned messy structured and unstructured data so the models perform accurately.
Akhtar Shah — AI/ML Engineer
AI that reads your inbox, makes your calls, and hires.
Three years shipping production AI: an autonomous email-to-phone-call pipeline on AWS Bedrock, a recruitment platform that cut HR costs 70%, and voice agents that hold real conversations.
0%HR costs cut (AIREC)
0%Hiring efficiency gained
0AI systems built end-to-end
0 yrsShipping ML
01Profile
Currently: AI Engineer @ Vizz Web Solutions
“I build systems that act — agents that don’t wait for a prompt.”
Three years of building and deploying end-to-end AI applications across health, recruitment, productivity and generative AI. I specialize in large language models, autonomous agents, voice-calling agents and conversational chatbots — and in the backend engineering that makes them hold up in production: queues, caches, microservices, and data pipelines that don’t fall over.
At Vizz Web Solutions I build agentic systems that handle complex workflows and make decisions on their own. As a consultant at Airec I helped ship the recruitment platform that cut HR operating costs by 70%. Along the way I’ve run sprints, reviewed code and shaped technical roadmaps — I like owning a feature from whiteboard to deployment.
- Focus
- Autonomous agent orchestration
- Stack of the month
- LangGraph + Bedrock + Milvus
- Open to
- AI engineering roles & collaborations
02Experience
Three roles, reverse chronological
- a.AI solution development — designed and implemented advanced ML models, wrote custom APIs, and performed prompt engineering and fine-tuning aligned with business goals.
- b.System optimization — tuned AI systems for high performance with minimal latency and reduced API call overhead, scaling to hundreds of concurrent users.
- c.End-to-end ownership — managed the full lifecycle of AI features, from initial design and thorough testing to production deployment.
- d.Collaboration & documentation — wrote comprehensive technical documentation and worked with cross-functional teams to drive adoption.
- a.Data wrangling — collected, cleaned, preprocessed and integrated data from internal and external sources.
- b.Exploratory analysis — surfaced trends, patterns and anomalies, and summarized data insights for decision making.
- c.Modeling — built, trained and tuned supervised and unsupervised ML models with Python and scikit-learn for predictive and trend analysis.
I build systems that act — agents that don’t wait for a prompt.
03Selected Work
Six systems, built end-to-end
Cover Story — No 1
AIREC
An AI recruitment platform that cut HR operating costs by 70% and made hiring 98% more efficient.
The problem
For every open role, recruiters were manually screening piles of CVs, scheduling first-round calls, and judging soft skills — slow, expensive, and inconsistent. Strong candidates were getting missed.
What I built
- a.CV parsing and rank engine that matches candidates against the job description automatically.
- b.AI-led screening stage that asks role-relevant questions and evaluates answers in real time.
- c.Soft-skill evaluation layer so recruiters get a ranked shortlist with clear notes instead of raw resumes.
Python · FastAPI · LangGraph · Milvus · Gemini
Visit AIREC70%Cut in HR costs
98%Hiring efficiency
Hours →
minutesScreening time
minutesScreening time
“The screening process went from taking days to under an hour. We finally focused on real interviews instead of reading resumes.”
“Candidates were ranked with clear reasoning. The system felt like a senior recruiter who never gets tired.”
“It handled hundreds of concurrent users without slowing down. The API optimization made a real difference.”
The problem: People waste hours every week reading low-priority emails, scheduling appointments, and making follow-up calls. Important requests get buried.
What I built: Envia connects to Gmail, classifies incoming email, and automatically handles the whole workflow — booking calendar events, making outbound calls via AI voice, and remembering user preferences from conversation.
- a.Real-time email ingestion from Gmail, queued via AWS SQS for Lambda workers.
- b.Onboarding scans the last 90 days of email so the AI understands the user from day one.
- c.A lightweight model classifies relevance, escalating important emails to Claude on Bedrock for action.
- d.Conversational AI books appointments through outbound calls, creates events, and saves personal/family preferences.
Outcome: A fully autonomous scheduling and communication assistant that removes the back-and-forth from daily email.
The problem: Practicing social skills or rehearsing difficult conversations with a real person can feel awkward or high-stakes. People need a safe, realistic space to practice.
What I built: HiBFF creates persona-based digital humans that hold natural voice and text conversations, adjusting to emotional tone and behavior.
- a.Persona engine defines each avatar’s background, voice, and conversation style.
- b.LangGraph orchestrates multi-turn dialogue with memory and context.
- c.D-ID powers realistic video avatars; ElevenLabs provides lifelike voice.
- d.Milvus stores conversation memory so the avatar remembers past sessions.
Outcome: Users get a private, judgment-free conversation partner that adapts to them.
The problem: Health apps are usually fragmented — one app for diet, another for workouts, another for sleep. Users bounce between tools and lose the big picture.
What I built: A unified AI health backend covering meal planning, workout generation, sleep tracking, hydration analysis, and a community layer.
- a.Barcode lookups and nutrition analysis using Open Food Facts and AI image interpretation.
- b.AI generates personalized meal plans and workouts based on goals and restrictions.
- c.Redis caching and Celery workers keep heavy requests fast and responsive.
- d.Milvus stores user health embeddings for better recommendations over time.
Outcome: A single backend that connects nutrition, fitness, sleep, and hydration in one place.
The problem: Most fitness plans are static. After a few weeks they stop matching the user’s real progress, recovery, or available equipment.
What I built: GymAI creates AI-crafted training programs that adapt continuously based on performance, recovery, goals, and the equipment at hand.
- a.Collects goals, fitness level, injuries, and available equipment during onboarding.
- b.LangGraph agents adjust reps, intensity, and rest based on logged workouts.
- c.ChromaDB stores workout history so plans evolve with the user.
- d.Groq powers fast inference so plan updates feel instant.
Outcome: Workouts that stay useful instead of going stale after week three.
The problem: Small agencies spend too much time on repetitive client work — writing posts, replying to comments, following up leads. Growth becomes limited by human hours.
What I built: BizAI is an autonomous agency assistant that learns each client’s brand voice and runs marketing and engagement workflows automatically.
- a.Brand memory stored in ChromaDB keeps messaging consistent per client.
- b.LangGraph agents handle content generation, scheduling, and lead follow-up.
- c.Claude powers high-quality, context-aware marketing copy.
- d.FastAPI + MongoDB backend serves the multi-tenant agency dashboard.
Outcome: Agencies can handle more clients without adding linear headcount.
04Index of Skills
Nine categories, set as running text
Languages— 02
Python, Node.js.
ML & Data Science— 07
Scikit-learn, XGBoost, LightGBM, Pandas, NumPy, feature engineering, model evaluation.
Deep Learning & AI— 07
TensorFlow, PyTorch, Hugging Face, spaCy, fine-tuning (LoRA), LLMs, generative AI.
NLP & RAG— 06
LangChain, LangGraph, LlamaIndex, Sentence-Transformers, Milvus, ChromaDB.
Computer Vision— 07
OpenCV, CNNs, YOLOv8, object detection, PaddleOCR, DeepFace, FaceNet.
MLOps & Deployment— 07
MLflow, Docker, Airflow, Jenkins CI/CD, Git, model monitoring, experiment tracking.
Data Eng. & Cloud— 09
PostgreSQL, MySQL, ETL pipelines, DataMart, T24, AWS SageMaker, FastAPI, Flask, RunPod.
TTS & STT — Voice— 06
ElevenLabs, Vapi, Retell AI, Whisper, Deepgram, AWS Transcribe.
Leadership— 05
Team management, sprint planning, code reviews, technical roadmap, hiring & onboarding.
* yes, engineers can run sprints.
05Appendix
Education · Research · Certifications
A.Education
Bachelor’s in Computer Science
Relevant courses: Artificial Intelligence, Machine Learning, Deep Learning, Computer Vision.
B.Research
Car Facelift Design Generation Using GenAI
Assessed the feasibility of generative networks for producing realistic, diverse car facelift designs to assist designers — on a custom-built car front-design dataset.
C.Certifications
Supervised Machine Learning: Regression and Classification
Generative AIs, LLMs, and ML
Become A Certified Python Programmer: Python Practice Tests
Intro to Coding in Python
06Correspondence
Letters to the editor
Let’s put AI
to work.
akhtargharsheen@gmail.com
Click to copy — opens your mail app too
Open to AI engineering roles — replies within 24h