I'm a machine learning engineer, and my work is split roughly evenly between computer vision and language models. A lot of my work starts with a question that doesn't have a clean answer yet. I like that stage: reading papers, building datasets, running experiments until something holds up. I like the stage after it just as much, when what worked has to become fast and reliable enough to ship, often on AWS (I'm an AWS Certified Machine Learning Associate).
What I work on
Computer vision
Detection, tracking, keypoints, 3D body reconstruction, action recognition, and diffusion models. Some of it runs on phones and edge devices, which is where quantization, TensorRT, ONNX, and TFLite come in.
Language models and agents
RAG and GraphRAG, knowledge graphs, agents with memory and tools, and MCP servers that let assistants like Claude use them. Mostly on AWS (Bedrock, Lambda, Neptune) behind FastAPI services.
What I've been building
Show
Swing3D
Valere Labs · computer vision
Measuring how baseball players move, from video: position, speed, acceleration, and full biomechanical analysis.
55%
lower end-to-end latency
29 FPS
on iPhone and Android
Detection. YOLO finds the ball, batter, pitcher, and bat.
Bat and ball in 3D. A top-down approach: RF-DETR finds the bat and ball, then a keypoint model I trained locates the bat, and a BlurBall model I trained locates the ball.
The body in 3D. SAM 3D Body reconstructs a mesh of the player, and OpenSim runs biomechanical analysis on it.
Breaking down the swing. On-device action recognition models (SlowFast and MS-TCN) find each swing and split it into phases: leg lift, stride, and so on.
None of the bat, ball, or swing models existed for this footage, so the hard part was all of it: generating the datasets, setting up validation, training, choosing loss functions, pushing accuracy, and then seeing whether it held up on real test videos.
Then I made it fast: TensorRT, post-training quantization and quantization-aware training, TFLite and ONNX exports, and multiprocessing and threading across the pipeline. End-to-end latency dropped by 55%. A detector for baseballs and golf balls runs at 29 FPS on both iPhone and Android.
PyTorch
YOLO
RF-DETR
BlurBall
SAM 3D Body
OpenSim
SlowFast
MS-TCN
TensorRT
TFLite
ONNX
FontSwap
Monotype · generative models
Change the font of the text in an image, and leave everything else exactly as it was.
65%
faster inference
Harder than it sounds: a small shift in stroke width or letter proportions turns one font family into another, and that counts as a failure.
It took text segmentation, edge detection, ControlNet, and a lot of experiments across Stable Diffusion, SDXL, SD3, and Flux with LoRA adapters, plus research into Rectified Flow. The end result was a custom diffusion model that shipped to production.
Inference got 65% faster with few-step LoRA adapters, TensorRT, quantization, and running CPU and GPU work in parallel.
Stable Diffusion
SDXL
SD3
Flux
LoRA
ControlNet
Rectified Flow
TensorRT
OrgGraph
Valere Labs · LLMs and agents · ongoing
An agent is only as useful as the context it has.
50%
lower AWS cost, same accuracy
OrgGraph syncs a company's Gmail, Google Chat, Calendar, and Slack, and builds a knowledge graph for each person that links to everyone else's.
It uses the graph to run short interviews with employees, filling in what the data can't show about how they work. With that context, it recommends agents to deploy on platforms like monday.com and Claude.
I also built its MCP server, so Claude and similar assistants can find those agents, run them with extra connectors, and query the graph directly.
It runs on AWS: S3 and Lambda for syncing, Bedrock for models, Neptune for the graph, and FastAPI services. I cut the AWS bill in half without losing accuracy by matching each task to the right model and thinking level, tuning Neptune, caching, parallelizing, and tightening prompts.
Knowledge graphs
GraphRAG
Agents
MCP
Bedrock
Lambda
Neptune
S3
FastAPI
CaseClips
Valere Labs · computer vision + LLMs
Hours of footage from criminal cases and proceedings, and someone who needs to find the few minutes that matter.
A vision-language model watches the video and summarizes what happens.
Those summaries go into a vector database, vector search finds the main events, and the video is clipped into short videos around them.
Vision-language models
Embeddings
Vector search
RAG
Video processing
Smaller builds
Proposal writer
Agentic RAG over a library of past projects to find what matches a job post, then drafts a proposal from it. Long-term and short-term memory keep track of each user's preferences.
Valere Labs
Logo tampering detection
A ResNet-based model that compares a reference logo with a test image, trained on a dataset I built. AUC 0.959.
Personal project
Pneumonia detection from chest X-rays
My first time working with medical images. A custom CNN at 95% accuracy that also held up in lab testing.
Humanli.ai
Data analysis agent
An agentic system for sales and marketing analysis and projections that writes the reports, built around a fine-tuned LLaMA model.
Humanli.ai
Study assistant
A RAG system that answers students' questions from their textbooks, with the source and a worked solution.
Humanli.ai
Tools I work with
Languages
Python
C++
C
Java
Machine learning and vision
PyTorch
TensorFlow
Keras
scikit-learn
OpenCV
Hugging Face
Ultralytics YOLO
NumPy
pandas
LLMs and agents
LangChain
Claude
Gemini
OpenAI
Amazon Bedrock
MCP
Optimization and deployment
TensorRT
ONNX
TFLite
FastAPI
Docker
AWS
Google Cloud
Firebase
Linux
Git
Also used
Streamlit
Django
Flask
MySQL
Raspberry Pi
Arduino
Experience
Apr 2025 – now
Valere Labs
Software Developer, AI/ML
Computer vision and LLM systems: Swing3D, on-device models at 29 FPS, CaseClips, and OrgGraph with its MCP server.
May 2024 – Mar 2025
Monotype
AI Trainee, Research
Research on diffusion models for FontSwap, through to a custom model that shipped, with 65% faster inference.
Jul 2023 – May 2024
Humanli.ai
ML Engineer
A small team, so I worked across everything: pneumonia detection from chest X-rays, an analysis agent built on fine-tuned LLaMA, and a RAG study assistant.
Research
Image Forgery Detection: A Survey of Recent Approaches
Amneet Kaur, Basav Bamrah, Prof. Jaspreet Kaur. Proceedings of the 4th National Research Scholars MEET, March 2023.
Sign Language Recognition Using Convolutional Neural Networks
Basav Bamrah, Prof. Jaspreet Kaur. Journal of Advanced Research in Electronics Engineering and Technology, Vol 9 No 1&2, 2022.
