I build deep learning systems that make it all the way to production — training and evaluating models, squeezing them onto real hardware, wrapping them in APIs, and keeping the data and GPU infrastructure behind them healthy. Currently an AI Engineer at Paraxial Technologies, working on large-scale 3D medical image analysis.
I'm an AI Engineer and Data Scientist with two years of professional experience across deep learning, computer vision, LLM applications and data engineering. My work usually starts at the model and ends somewhere much less glamorous — a container, a GPU driver, an evaluation sheet — because that's where most AI projects quietly fail.
At Paraxial Technologies I own core parts of a production 3D imaging platform: I cut peak inference memory by 76% so models could run on consumer GPUs instead of $7,000 cards, rebuilt the object-tracking logic that links detections across time, analysed 11,365 annotated samples to re-tune detector anchors, and shipped a microservice that returns results in about 3.6 seconds. I also maintain the Docker releases, CUDA environments and ten GPU servers the team works on.
Before that I worked at Blackcoffer on LLM tooling, vector-search based computer vision and large-scale data extraction, and interned at upGrad building CNN classifiers. I like problems where the answer isn't a bigger model — it's better data, sharper evaluation, or an engineering decision nobody wanted to make.
Everything here is something I've shipped with, not just read about.
Day-to-day in Python; comfortable dropping into shell, SQL or JS when the task needs it.
Training, fine-tuning and — more often — figuring out why a model looks good on paper and bad in production.
Building applications on top of language models — retrieval, structured extraction, local inference.
2D and 3D — detection, segmentation, tracking, re-identification and registration.
Getting messy real-world data into a shape a model or a stakeholder can actually use.
Containers, GPUs, APIs and releases — the part between "the notebook works" and "the client can use it".
Core engineer on a production 3D imaging AI platform. I own the detection and segmentation models, the longitudinal tracking pipeline, automated reporting, container releases and the GPU infrastructure the whole team runs on.
Delivered client-facing data science and computer vision projects end to end, from scraping and modelling through to deployed APIs and dashboards.
First hands-on exposure to applied deep learning: training image classifiers and shipping a working prototype.
A mix of production systems and things I built because I wanted to know if they'd work.
An end-to-end deep learning platform that ingests volumetric scans and runs detection, segmentation, multi-timepoint registration and automated report generation. I own the model layer, tracking logic, module architecture and release builds.
A deep-dive into why a 3D detection stack demanded 44 GB of VRAM, and a rework that brought it down to 10.59 GB — moving an entire product line off $7,000 workstation cards onto consumer GPUs without changing a single prediction.
A standalone service built from scratch that pulls studies from an imaging server, runs segmentation on demand and returns per-slice boundaries in roughly 3.6 seconds — fast enough to sit inside an interactive review tool.
Scrapes a user's public Reddit activity and builds a structured behavioural persona with a locally hosted LLM — no API keys, no data leaving the machine. A study in getting reliable structured output out of small models.
Keeps a consistent identity for every player across multiple camera angles of the same match, combining detection, multi-object tracking and appearance embeddings to solve the ID-switching problem.
Aggregates coverage of a given company, runs comparative sentiment analysis across articles, and speaks the summary back in Hindi — a full pipeline from scraping to audio in one app.
A fault-tolerant extraction system covering 32+ structurally different legal portals, with proxy rotation and retry logic. Turned a six-hour daily manual task into a thirty-minute automated run.
A high-throughput visual matching system built on a Qdrant vector database, reaching 98% accuracy on the client's benchmark and deployed on GPU cloud infrastructure for low-latency lookups.
An integrated pipeline connecting a local BLIP image-captioning model with an Ollama-served LLM (Mistral). Extracts structured fashion attributes (neckline, silhouette, etc.) from unconstrained product images with graceful regex fallbacks.
An intelligent web platform that predicts meal demand and peak rush hours using historical reservation data. Built to help students avoid queues and assist administrators with data-driven inventory management.
Longer write-ups of problems I've actually solved — including the parts that didn't work.
The model only ran on a $7,000 card. Profiling showed the memory wasn't going where I assumed, and the fix was a configuration decision rather than a smaller model.
Read the case studyI inherited a pipeline whose tests were marked as passing. They were producing plausible-looking wrong answers, which is a harder problem than an outright crash.
Read the case studyWriting the scrapers was the easy half. Making the output trustworthy enough that someone would stop checking it by hand was the actual project.
Read the case studyHappy to talk about ML systems, inference optimisation, or anything else in applied AI.