Nafi Ullah Shafin
ML Engineer · AI Systems & MLOps
3+ years building AI systems end to end — labelling and training my own vision models, exporting them light enough to serve cheaply, hosting them behind low-latency inference APIs, and tracing every prediction in production. Lately: LLM agents and real-time voice systems built on the same discipline.

About Me
I am an ML engineer who builds the whole path, not one slice of it. Data gets collected and labelled through annotation tooling I help build, models get trained and evaluated on metrics rather than impressions, checkpoints get exported small enough to serve on modest hardware, and inference runs in production behind APIs that log every call. When a prediction looks wrong, I can trace it back to the model version and the dataset snapshot that produced it.
My computer vision work runs from YOLO detection and segmentation to measurement extracted from masks — automated vehicle image enhancement at App-Concept, cattle weight estimation from a single phone photo, and the perception stack for an autonomous vehicle team. The recurring theme is making a model survive contact with real, messy, field-condition input rather than a clean validation set.
On the LLM side I build agent systems that actually do things: stateful LangGraph orchestration, ReAct tool calling, and RAG grounded in a client's own knowledge base so answers stay sourced. That includes a multi-agent sales representative that qualifies and nurtures leads on its own, and Market Flick AI, an agentic market research system that plans its own research steps.
Most recently I have been working in real time — a Bengali speech-to-speech voice agent where streaming STT, LLM reasoning, and streaming TTS have to cooperate inside a sub-400ms turn budget. Coming from a frontend and product background, I also care about what the AI feels like to use, not just what it scores.
AI & ML Engineering Experience
Model ownership, production deployment, and the MLOps that keeps it all traceable.
3+ Years Building AI Systems
From first OpenCV experiments to production inference pipelines and real-time voice agents.
- 2022 - 2023
Foundations in ML & Computer Vision
Moved from software engineering into machine learning at SUST — Python data pipelines, classical ML, and first hands-on OpenCV work. Led the team that placed 1st Runner-Up at NASA Space Apps Bangladesh with a geospatial fire-detection app selected for global judging.
PythonOpenCVClassical MLData Pipelines - 2023 - 2024
Deep Learning & Custom Model Training
Started training my own models instead of consuming pretrained ones — YOLO detection and segmentation on self-labelled datasets, augmentation strategies, and evaluation on mAP rather than eyeballing outputs. Placed 1st Runner-Up in the TensorFlow ML Olympiad CO2 emissions challenge.
YOLOUltralyticsImage SegmentationModel Evaluation - 2024
First Production CV System — Cattle AI
Shipped a vision model into real hands: cattle weight estimation from a single smartphone photo, trained from field imagery, containerized, and deployed on AWS EC2 behind a credit-metered inference API with S3-backed image storage.
YOLOOpenCVFastAPIAWS EC2Docker - 2024 - Present
ML Engineer — App-Concept GmbH
Own the whole loop for an automotive imaging platform: a custom in-house annotation tool feeds versioned datasets into training, models are exported lightweight enough to run cheap, and every production inference is logged and traced back to the model version that produced it.
MLOpsModel ServingInference TracingCI/CD - 2024 - Present
Perception Systems Lead — SUST Autodrive
Direct the perception stack for an autonomous vehicle team — multi-sensor integration, object and lane detection, and scene understanding — while mentoring junior leads through research and experimental validation.
PerceptionSensor FusionLane DetectionResearch - 2025
Agentic AI & LLM Orchestration
Shifted from single-shot prompting to real agent systems — stateful LangGraph orchestration, ReAct tool calling, and RAG over Pinecone. Placed 1st Runner-Up at the Vivasoft AI Hackathon with Market Flick AI, took Champion at Code Samurai 2024, and completed DeepLearning.AI's Agentic AI programme.
LangGraphRAGReActMulti-agent - 2026 - Present
Real-time Voice AI — Kormi AI
Building phone-native Bengali voice agents on a speech-to-speech pipeline — streaming STT into LLM reasoning into streaming TTS — with turn-taking logic tuned hard enough to hold sub-400ms response latency in optimized flows.
LiveKitStreaming STTBengali TTSSub-400ms Latency
Roles & Impact
ML, computer vision, and agent-engineering roles across remote teams and client work.
ML Engineer · App-Concept.com GmbH
2024 - PresentGermany · Remote
Own the end-to-end machine learning lifecycle for an automotive imaging platform. A custom in-house annotation tool feeds versioned datasets into YOLO segmentation training; the resulting lightweight models are exported, containerized, and served as low-latency inference APIs on AWS EC2 — with every prediction logged and traceable back to the model version and dataset snapshot that produced it.
- Shipped the full annotate → train → host → production-inference loop, cutting model iteration time from weeks to days
- Trained mobile-friendly lightweight segmentation models that hold accuracy at a fraction of the parameter count
- Brought median inference latency under a second on CPU-only instances through pre-processing and export optimization
- Built inference logging and tracing so any production prediction can be reproduced from its exact model and dataset version
Voice AI Engineer · Kormi AI — E-commerce Voice Agent
2026 - PresentRemote
Built a production-grade Bengali voice agent for e-commerce automation — outbound order confirmation, abandoned-cart recovery, upsell conversations, and inbound customer call handling — on a real-time speech-to-speech pipeline of streaming STT, LLM reasoning, and natural Bengali TTS running over live phone calls.
- Achieved sub-400ms turn latency in optimized flows via persistent WebSocket TTS connections and streaming generation
- Designed turn-taking and barge-in logic that keeps phone conversations natural under real-world network jitter
- Implemented agent workflows for intent detection, order verification, address confirmation, and objection handling
- Emitted structured post-call outcomes that flow straight into downstream CRM automation
AI Agent Engineer (Freelance) · Fuel AI — CRM Automation Platform
2025Remote
Architected a multi-agent sales-representative system with LangChain and LangGraph that automates the sales workflow end to end — lead qualification, nurturing, and meeting scheduling — using stateful agent orchestration and advanced RAG retrieval grounded in the client's own product knowledge base.
- Built stateful LangGraph orchestration so agents resume long-running lead conversations without losing context
- Implemented advanced RAG over Pinecone for grounded, source-backed answers to prospect questions
- Enabled autonomous tool and API calling with the ReAct pattern, removing manual CRM data entry from the loop
- Exposed the agent graph as FastAPI services deployed with Docker on AWS infrastructure
AI Engineer (Freelance) · Cattle AI
2024Remote
Built an end-to-end vision system that estimates cattle weight from a single smartphone photo — YOLO-based detection and segmentation with OpenCV geometry converting masks into physical measurements — wrapped in a credit-metered inference backend with admin analytics.
- Delivered a working weight-estimation model from raw field imagery through annotation, training, and evaluation
- Designed real-time inference services with per-user credit logic and admin-facing usage analytics
- Deployed the full stack on AWS EC2 and S3 with Docker and CI/CD workflows
Director of Perception Systems · SUST Autodrive
2024 - PresentSylhet, Bangladesh
Lead the design and implementation of autonomous vehicle perception systems — multi-sensor integration and computer vision pipelines for object detection, lane detection, and environmental perception — while mentoring junior leads through perception research and experimental validation.
- Architected a modular perception framework letting detection, lane, and fusion modules evolve independently
- Optimized object and lane detection for reliable real-time scene understanding during autonomous navigation
- Supervised and mentored junior leads in advanced perception research, system design, and experimental validation
AI Engineer · JobsNavi
2023 - PresentGermany · Remote
Built the AI layer of a job matching platform — automated job scraping, LLM-based structured data extraction that turns messy postings into clean typed records, and recruiter-side automation services exposed through a FastAPI backend.
- Automated job ingestion and structured extraction, replacing manual posting cleanup
- Used GPT-4 with schema-constrained prompts to keep extracted records reliably typed and machine-consumable
- Shipped recruiter automation services that match candidates against structured job requirements
Selected AI Work
Production ML systems and agent platforms — computer vision pipelines, MLOps loops, multi-agent orchestration, and real-time voice. Where a recording exists, the demo is the real thing.
App-Concept.com GmbH
2024 - PresentAnnotation Studio → Train → Host → Inference
The loop that keeps the model improving. Reviewers draw and correct polygon masks in a custom in-house annotation tool; each labelled batch is snapshotted as a versioned dataset; YOLO segmentation training runs are tracked on mAP rather than vibes; and a single path takes the winning checkpoint to containerized hosting. Production inference logs feed failure cases straight back into the next labelling round.
App-Concept.com GmbH
2024 - PresentCar AI — Production Vehicle Image Intelligence
The customer-facing end of that pipeline. A photo goes in, a lightweight segmentation model behind a containerized inference API returns masks in under a second, and the enhanced result comes back to the app. Every call is logged with its model version, latency, and confidence, so a bad output is a traceable event rather than a mystery.

Kormi AI
2026 - PresentKormi AI — Bengali Voice Agent
A phone-native voice agent that confirms orders, recovers abandoned carts, upsells, and handles inbound customer calls in natural Bengali. Streaming STT feeds LLM reasoning which feeds streaming TTS over a persistent WebSocket connection, with turn-taking logic tuned to hold sub-400ms response latency in optimized flows — close enough to human timing that callers stop noticing.
Vivasoft AI Hackathon 2025 · 1st Runner-Up
2025Market Flick AI — Agentic Market Research
Give it a product or a market question and it plans its own research: decomposing the question into sub-queries, gathering and cross-checking sources, and returning a structured, citation-backed market brief instead of a wall of unsourced text. Built under hackathon time pressure and placed 1st Runner-Up.

CRM Automation Platform
2025Fuel AI — Multi-agent Sales Representative
A sales-rep agent system that qualifies leads, nurtures them across conversations that span days, and books meetings on its own. Stateful LangGraph orchestration keeps each lead's context alive between touches, and Pinecone-backed RAG keeps every reply grounded in the client's real product knowledge rather than model imagination.
Freelance
2024Cattle AI — Vision-based Weight Estimation
Farmers photograph an animal from their phone; a YOLO segmentation model isolates the body, OpenCV geometry converts the mask into physical measurements, and the system returns an estimated weight — no scale, no handling stress. Served through a credit-metered inference API with admin analytics behind it.

SUST CSE Carnival · 1st Runner-Up
2024Reelify — AI Text-to-Video
Turns a text prompt into a finished short-form reel. An LLM drafts the script and breaks it into scenes, then a programmatic rendering pipeline composes frames, voiceover, and captions into a video file — no human editor in the loop. Built in a 24-hour hackathon and still live.
More AI & ML Work
Competition builds, research projects, and the observability groundwork behind production inference.

JobsNavi — AI Job Matching
The AI layer of a job matching platform: automated job scraping plus LLM-based structured extraction that turns messy free-text postings into clean, typed records, feeding recruiter-side automation and candidate matching through a FastAPI backend.

Inference Observability Stack
A three-tier service instrumented end to end with OpenTelemetry — distributed traces and metrics exported to Jaeger and Grafana, the whole microservice stack containerized. The same tracing discipline is what makes production model inference debuggable rather than mysterious.

CO2 Emissions Prediction
A regression pipeline for the ML Olympiad CO2 emissions challenge — feature engineering over vehicle and fuel attributes, model selection, and hyperparameter tuning validated against a public leaderboard rather than a private notebook.

AI Ecommerce Business Analyst
An AI system that reads raw ecommerce store data and returns commercial insight — sales pattern analysis, product-level diagnosis, and plain-language recommendations for the operator. Built at the international Collaborative AI Hackathon.

Wildfire Geospatial Intelligence
Team-led geospatial app that maps fire-affected regions from satellite data and renders a 3D projection of fire spread, with a community forum layer on top. Selected for global judging among 200 teams.

EcoSync — Smart Waste Management
A city-scale waste management platform for DNCC combining route-aware collection planning, geospatial dashboards, and a field-worker mobile app. Took Champion at Bangladesh's largest inter-university hackathon against 406 teams.
Achievements
AI and ML competition results first, then the wider engineering and research record.

1st Runner Up, Vivasoft AI Hackathon 2025 — Market Flick AI
Built Market Flick AI, an agentic market research system that plans its own research steps, gathers and cross-checks sources, and returns a structured, citation-backed market brief instead of an unsourced wall of text. Built under hackathon time pressure on LangChain and LangGraph, it placed 1st Runner Up at the Vivasoft AI Hackathon 2025.

1st Runner Up, Collaborative AI (International Hackathon)
Built an AI system for ecommerce business analysis — turning raw store data into actionable commercial insight — at the international Collaborative AI Hackathon organized by SJ Innovation LLC and CloudCamp.

1st Runner Up, ML Olympiad 2024 (TensorFlow User Group)
Competing as team HabiJabi in the CO2 Emissions Prediction Challenge hosted by United International University and TensorFlow User Group North Bengal — feature engineering, model selection, and regression tuning under a public leaderboard.

1st Runner Up, SUST CSE Carnival AI Hackathon — Reelify
In a day-long hackathon we shipped Reelify, an MVP that turns plain text into a finished reel video: an LLM drafts the script and scene breakdown, and a programmatic rendering pipeline composes the frames, voiceover, and captions.

Champion in Code Samurai 2024
We became champions in the biggest inter-University Hackathon 'Code Samurai 2024', competing against 406 teams. Our project involved designing a comprehensive waste management software for DNCC. We built both mobile app and web app

Finalist, DevOps Hackathon
Built a three-tier application instrumented end to end with OpenTelemetry — distributed traces and metrics exported to Jaeger and Grafana for remote monitoring, with the whole microservice stack containerized. The same tracing discipline underpins how I log and debug production model inference today.

1st Runner Up, NASA Space Apps Challenge Bangladesh 2023
As team leader, built a geospatial wildfire app that maps fire-affected areas and renders a 3D projection of fire spread, with a community forum layer. Selected for global judging among 200 teams.

Champion, Mobile App Development Competition 2023
As the sole team member, I created a game app that allows users to explore Bangladesh’s history while playing. This project won the Champion title in the Mobile App Development Competition 2023 in the Sylhet region. The app combines educational content with engaging gameplay.

2nd Runner Up, LU CSE Carnival 2023
Secured the 2nd Runner Up position in the LU CSE Carnival 2023 for developing an app that facilitates paid appointments between experts and aspiring individuals abroad. The project used Flutter, React, Node.js, MySQL, and Google Maps. This achievement was a result of a 24-hour hackathon.

2nd Runner Up, Cozmo Clench Tech Fest
My team take place 'Cozmo Clench' segment at Tech Fest, IIT Bombay. The competition involved robotics skills. We qualified for the final round demonstrating strong teamwork and technical expertise.

2nd Runner Up,BUET Science Festival 2022
Led Team Supernova to secure the second runner-up position for the Robo-Defense project in the Project Display (Mechanical) category. Also won the 2nd position in the Science quiz at the '2nd Science Festival 2022' powered by Walton Laptop and BUET.

Prize Winner, National Science Festival
Won a prestigious prize for my mechanical project at the National Science Festival powered by Prothom Alo. This achievement marked a significant milestone in my formative years, laying the foundation for my future in engineering.
AI Stack & Tools
The tools I use to take a model from raw data through training to traced production inference — and the agent frameworks layered on top.
ML & Computer Vision
PyTorch
TensorFlow
YOLO / Ultralytics
OpenCV
scikit-learn
ONNX Runtime
LLM & Agents
LangChain
LangGraph
RAG Pipelines
Pinecone
OpenAI
Hugging Face
Data & Serving
Python
FastAPI
Pandas
NumPy
Redis
PostgreSQL
MongoDB
MLOps & Cloud

Docker
AWS
GitHub Actions
OpenTelemetry
Grafana
Linux
Realtime & Product
LiveKit
WebRTC
NextJS
TypeScript
Git
What Clients Say
Verified reviews from Upwork and LinkedIn across client engagements and collaborations.

Verified client feedback from an Upwork engagement — delivered on scope, on schedule, and maintained after handoff.

A recommendation from a collaborator on LinkedIn, on how I work through ambiguous problems with a team.

Verified client feedback from an Upwork engagement — clear communication and a system the client could run themselves.

A recommendation from a collaborator on LinkedIn, on ownership and follow-through on delivery.

Verified client feedback from an Upwork engagement — technical depth applied to a real business outcome.

A recommendation from a collaborator on LinkedIn, on turning research ideas into shipped software.

Verified client feedback from an Upwork engagement — delivered on scope, on schedule, and maintained after handoff.

A recommendation from a collaborator on LinkedIn, on how I work through ambiguous problems with a team.

Verified client feedback from an Upwork engagement — clear communication and a system the client could run themselves.

A recommendation from a collaborator on LinkedIn, on ownership and follow-through on delivery.

Verified client feedback from an Upwork engagement — technical depth applied to a real business outcome.

A recommendation from a collaborator on LinkedIn, on turning research ideas into shipped software.
Get In Touch
If you are building something with models, agents, or real-time AI — or you need an existing pipeline taken to production — reach out on LinkedIn or by email and I'll get back to you whenever I can.