I'm an AI Systems Engineer focused on building production-oriented AI systems, particularly around agentic workflows, multimodal AI, and real-time applications.
I enjoy taking an ambiguous idea, understanding the underlying problem, breaking it into systems and components, evaluating the trade-offs, and turning it into something that actually works.
My focus is not just on using AI models, but on how AI fits into a reliable, controllable, and well-engineered system.
I work across areas such as:
- 🤖 Agentic AI & LLM-based applications
- 🎙️ Real-time AI and voice interactions
- 👁️ Multimodal AI / Vision-Language Models
- ⚙️ AI workflows and orchestration
- 🧠 Structured information extraction and reasoning
- 🔄 Stateful applications and lifecycle-driven systems
- 🐳 Production-oriented deployment and infrastructure
I'm particularly interested in the engineering problems that appear after the initial AI demo works — reliability, failure handling, observability, controllability, and system design.
A platform for building customizable AI interviewers from natural-language requirements and conducting interviews through real-time conversation.
SkillViewer allows a user to describe what they want to assess — including the skill, evaluation criteria, and interview requirements — and turns that specification into an executable AI interviewer.
The interviewer then conducts the assessment through a real-time conversational interface.
Instead of giving every candidate the same predefined interviewer, the system allows users to define and customize the interviewer itself.
The workflow looks roughly like:
User Requirements
↓
Structured Interview Specification
↓
Interview Skill
↓
AI Interviewer
↓
Real-time Interview
↓
Evaluation
The interesting engineering challenge wasn't simply building a voice agent.
It was figuring out how to represent an interviewer's behavior as a structured system that an agent can reliably execute.
- Agentic AI
- Structured LLM workflows
- Natural language → structured specifications
- Real-time voice interaction
- Stateful workflows
- Agent orchestration
- System design and architecture
- Reliability and controllability
A multimodal AI pipeline that analyzes videos and generates captions in multiple styles without requiring manual video review.
The system processes a video through multiple stages, first understanding its contents and then generating captions from that understanding.
Rather than allowing a vision-language model to directly produce the final caption, the system introduces a structured intermediate representation:
Video
↓
Video Understanding
↓
Canonical Video Report
↓
Caption Generation
↓
Multiple Caption Styles
The Canonical Video Report (CVR) acts as a boundary between video understanding and language generation.
This separation makes the system easier to reason about, debug, evaluate, and extend.
The pipeline was designed with practical constraints around:
- Runtime
- Model resource requirements
- Docker image size
- Batch processing
- Concurrent execution
- Failure isolation
- Data integrity
Individual failures are handled without bringing down the entire pipeline, while different stages remain independently testable and observable.
- Vision-Language Models
- Multimodal AI
- Pipeline architecture
- Structured intermediate representations
- Failure isolation
- Concurrent processing
- Testing and reliability
- Dockerized deployment
- AI workflow design
🔗 Explore Video Captioning Agent →
When I encounter an interesting system, I don't just think about how to use it.
I tend to ask:
How does it work?
Why was it designed this way?
What are the constraints?
What would I change if I built it myself?
That curiosity shapes how I approach projects.
Before implementing, I prefer to understand the relevant concepts, decompose the problem, explore possible architectures, and evaluate them against the actual constraints.
Then I build, test, observe the behavior, and iterate.
I'm particularly interested in systems where:
- inputs can be ambiguous
- failures are expected
- AI behavior needs to be controlled
- system state matters
- decisions need to be observable
- individual components need to evolve independently
I've documented my projects and the way I approach engineering on my portfolio.
You'll find:
- 🎬 Introduction & project walkthrough
- 🚀 Project demos
- 📝 Detailed devlogs
- 🧠 Engineering decisions
- ⚙️ System architecture
- 🔍 Project breakdowns
- 📫 Direct contact & scheduling
- 🌐 Portfolio: https://ai-engineer-portfolio-sigma-brown.vercel.app/
- 💼 LinkedIn: https://www.linkedin.com/in/adityamore2k/
- 🐙 GitHub: https://github.com/adityadmore2000
I enjoy turning ambiguous ideas into well-engineered systems.