AI-focused software engineer and published generative-AI researcher. I build applied-AI systems end to end: edge computer-vision pipelines with YOLO and ONNX, RAG-based LLM applications with LangChain, and deep-reinforcement-learning agents. I ship them inside production-grade, real-time React and TypeScript products for clients in the US and Europe.
My edge is the full path from model to product: training and evaluating the model, wrapping it in deterministic rules and human review where it matters, and building the interface people actually use. I work AI-native every day with Claude Code, MCP and agent workflows, and I'm certified in it by Anthropic Academy, Google and Oracle.
Computer vision at the edge, retrieval-augmented LLMs and reinforcement learning. Each one taken from idea to a working system.
Inside Guardian Lens
One frame, five layers of judgement.
01 · RTSP stream
A camera frame arrives
It starts with the CCTV and IP cameras already on site. Their live RTSP streams are read frame by frame on a machine at the edge, next to the cameras.
02 · OpenCV
OpenCV cleans it up
Each frame is decoded, resized and normalised, so the model always sees the same clean input whichever camera it came from.
03 · YOLO · ONNX Runtime
YOLO finds people and PPE
Locally trained YOLO models pick out each person and their protective gear. They run at the edge on ONNX Runtime, and every detection carries a confidence score.
04 · Safety rules
Rules decide what counts
A detection is not a violation yet. Deterministic rules, such as no helmet inside restricted Zone B, make that call, so every alert can be explained.
05 · Review → Dashboard
A person has the final say
A reviewer approves or dismisses each flagged event before anything is recorded. Verified violations are logged to PostgreSQL through FastAPI and appear on the dashboard.
The model proposes, the rules decide, a person confirms.
Every application starts as a document. The PDF is parsed into clean text with its sections kept in order, so layout never affects the result.
02 · Chunks · embeddings
It is split and embedded
The text is split into small chunks, and each chunk becomes an embedding: a vector that captures what the passage means, not just the words it uses.
03 · Vector store · top-k
Each requirement finds its evidence
For each requirement in the job description, the retriever pulls the top-k closest chunks from the vector store. Those passages are the evidence. A missing skill finds none.
04 · LLM judge
An LLM scores it against a rubric
A prompted rubric has the model score each requirement using only the retrieved evidence, and give its reason. It runs on LangChain and the OpenAI API, served by FastAPI.
05 · Score + reasons
A score you can question
The recruiter sees a match score with a bar and a reason for every skill. The pipeline reaches 90%+ screening accuracy and ships through Dockerized CI/CD.
Every score points back to its evidence. Reasons, not just keywords.
RAG pipeline with FastAPI, LangChain and OpenAI APIs for contextual candidate evaluation and role matching.
Designed prompt workflows and LLM evaluation pipelines reaching 90%+ screening accuracy.
Shipped with Dockerized CI/CD pipelines.
PythonFastAPILangChainOpenAI APIDocker
Inside IntelliTraffic
One junction that learns to switch.
01 · SUMO simulation
A junction in simulation
Training happens in SUMO, a traffic simulator: one four-way junction, cars arriving on every approach, and queues building behind a red light.
02 · State · TraCI
The agent reads the queues
At every step the agent reads the queue length on each lane through TraCI. Those numbers are its state, everything it knows about the road.
03 · DQN · Keras
A network scores each move
A Deep Q-Network built in Keras takes that state and estimates a Q-value for each action: how much waiting it expects to save by holding the phase or switching.
04 · Action → reward
It switches, then gets graded
The agent takes the highest-value action, here switching green from north-south to east-west. The reward then reports whether waiting went down. That is what it learns from.
05 · Fixed-time vs DQN
It beats the timetable
Against fixed-time signals, the trained agent cuts simulated vehicle wait time by 35%. Custom reward tuning made training converge 28% faster.
No fixed timetable. The signal learns from the queue in front of it.
Mohali, India · Offshore software development firm
Contributed to Guardian Lens, the company's computer-vision workplace-safety platform: training YOLO models and building the edge inference pipeline (see Selected work).
Working in a Git pull-request workflow with regular code reviews, debugging in Chrome DevTools and using AI-assisted development (Claude Code) daily.
Client product · United States
CityWide CWS
Security workforce management platform
01
Built real-time dispatch status, guard tour tracking and call management views consuming live Socket.IO event streams.
Integrated REST APIs against backend contracts, handling loading, error and empty states across data-heavy tables and live views.
Cut unnecessary re-renders on high-frequency live updates with React.memo, useMemo and useCallback.
ReactTypeScriptReduxSocket.IOREST
Client product · Spain
Happy Whole Human
Multi-role wellness platform
02
Developed role-based dashboards for Admin, Coach and Client with reusable, typed React components.
Integrated Stripe payments for coaching plans and subscriptions.
Separated data fetching from UI with RTK Query: caching, cache invalidation and async state over MongoDB-backed APIs.
Whether it's an AI engineering role, a product that needs applied AI shipped end to end, or just a good technical conversation, my inbox is open. I usually reply within a day.