Revanth Darshan D R
Loading the field…
AI / ML Engineer. Engineer in AI, Automation, ML.
I studied computer science at VIT Chennai, specialising in AI and machine learning, and finished in 2025 with a CGPA of 8.56.
Most of what I enjoy sits at the point where a model stops being a notebook and starts being something someone relies on. That usually means less time on the architecture than on the parts around it: what the input actually looks like, where it breaks, and whether anyone can tell when it’s wrong.
Away from the keyboard — sometimes I click photos as well :) I like nature, long walks, places I haven’t been yet. Exploring is the same curiosity that pulls me into new tech, just pointed at the world instead.
I also build random things — small ideas that show up and won’t leave until they’re real. Some turn into products. The rest teach me something.
I’m AWS certified as a Solutions Architect and Cloud Practitioner. I work in Python most days, and reach for C++, Java, or SQL when the job asks for it.
Python · TensorFlow · Keras · Computer vision · Deep learning · Pandas · NumPy
Prompt engineering · Agentic AI · NLP · RLHF · API integration
AWS · Azure DevOps · CI/CD · Git · SQL · Power BI
Prompt engineering, then generative tooling, then orchestration — each role a step further from the notebook.
I work as an AI process orchestrator — coordinating automation that spans document writing, database operations, and the screens people actually look at. RPA pulls structured requirements out of systems too old to have an API, and that feeds the pipeline. I also took an agentic requirements generator to the regional round of the internal AI hackathon.
Led GreenCode AI, a tool that reviews code for the energy it burns. Computing accounts for something like 7% of global electricity, and almost none of that shows up in a code review. I put the feedback where developers already are — a VS Code extension and a pre-commit hook — rather than in a dashboard nobody opens.
Refined more than 150 prompts against human feedback and moved from attempter to reviewer inside six months. On the Nightingale multimodal project I worked on pulling structure out of images, which lifted accuracy about 25%.
Built the collection layer for market intelligence work — scrapers that cut gathering time by about a third, feeding risk assessment and demand models.
Products that went out the door first, then the research behind them.
A matrimony and community app built end to end for a client and now live with real members. I owned the whole thing: React Native app, FastAPI backend, Postgres, Redis-backed workers and scheduled jobs, all containerised behind nginx — plus the release pipeline onto the Play Store and the infrastructure it runs on. Localised in English, Telugu, and Tamil, with verification, notifications, and analytics wired through.
A memory layer for personal AI that keeps facts consistent when one of them changes. Most systems store “I moved to Denver” and stop there; GEM tracks typed dependency edges and walks the chain, so the commute, the alarm, the departure time all get re-examined — while leaving unrelated facts alone. It scored 100% on 37 derived facts across deep-dependency scenarios, about 27 points ahead of flat vector memory.
A Windows app, published and live on the Microsoft Store. Small, but shipped — which is its own kind of exercise: packaging, store review, and updates for people you will never meet.
Deepfakes hold up well to a spatial look but leave traces in the frequency domain. LFAT reads both — a MobileNetV2 branch for what the image looks like, an FFT branch for what it's made of — and hands them to transformer encoders. It reached 99.34% on Celeb-DF at 5.6M parameters, roughly a quarter the size of the ResNet baseline it beat.
A CNN watching video for the moment attention goes. OpenCV handles the face, Keras the judgement, at about 90% accuracy — with augmentation doing most of the heavy lifting on the hard frames.
Feature selection and validation work that took the usable set from 500 samples to over a thousand and added five points of accuracy. Not the glamorous part of machine learning, but usually the part that decides whether it works.
LSTM models compared across univariate, multivariate, and multi-step horizons, scored on MAE, RMSE, and MAPE rather than whichever chart happened to look best.
I'd be glad to hear about it.
Nature, long walks, places I hadn’t been yet — the same curiosity that pulls me into new tech, pointed at the world instead. ( every frame here is mine )