Works I've Done
A clinical backend for AI-assisted dermatology. It routes analysis requests across three LLM providers with priority-ordered failover, runs a three-round conversational diagnostic pipeline over patient-uploaded skin images, and keeps Celery task state recoverable so a worker crash never loses a patient session.
Generates client-ready proposal and case-study decks from a short brief. The pipeline pulls brand assets, drafts the slide content, and enforces layout and colour rules, so output stays on-brand without a designer reviewing every deck.
Plans and publishes social content across platforms from a single brief. It drafts the copy, generates matching visuals, and posts on a schedule, so a week of content gets set up in one sitting instead of daily.
A scraper that decides per page whether it needs a browser at all. Static pages go through plain HTTP; JavaScript-rendered ones fall back to headless automation. I built it after too many runs silently returned empty markup on client-rendered sites.
A question-answering system for Indian law, built on a domain-fine-tuned LLM over statutes, precedents and judgments. The goal was letting someone ask a question in plain language instead of needing the right search terms to find the right case.
A browser chess game whose opponent is trained with Deep-Q Learning rather than a hand-written evaluation function. React and WebSockets on the front, Chess.js for move validation, and a difficulty that tracks the player instead of sitting at a fixed depth.
A smart glove that reads hand signs. Flex and motion sensors on an ESP8266 classify gestures at the edge, with heavier analytics offloaded to a Flask service on EC2. Built to find out how much of a gesture pipeline can actually run on a microcontroller. 95.7% on the test set.
Voice activity detection for real-time audio: deciding, frame by frame, whether anyone is actually speaking. The hard part was making the threshold adapt to room noise instead of being tuned per environment.
Detects doors and windows in architectural blueprints. A YOLOv8 model fine-tuned on annotated drawings and served behind FastAPI, returning element coordinates, types and confidence scores as JSON for downstream take-off tools.
Turns a photo of a card into structured fields. Built on Google Gemini, with camera capture or file upload, single cards or front/back pairs, and JSON or clipboard export.
Document question answering over a private corpus. LangChain for orchestration, Pinecone for vector storage, and a hybrid of dense and keyword retrieval with re-ranking, because dense-only kept missing queries that hinged on an exact term.
Streaming fraud detection over transaction data. Kafka for ingest, Spark for scoring in flight, Cassandra for state, and a dashboard on Spring Boot and Flask. The design constraint was scoring a transaction before it settles rather than flagging it afterwards.
Predicts whether an ad will be clicked, trained on 50,000+ labelled impressions during my Innovate Intern placement. Feature engineering mattered more than model choice here: the ensembles only pulled ahead once the categorical encoding was right.
Segments retail customers by spending behaviour. K-means over 4,000+ records, with the elbow method for choosing k, settling on five groups. Built during my Prodigy Info Tech placement.
Recognises a dish from a photo and estimates its calories against a nutrition database. A classification model plus a lookup, wrapped so the whole interaction is one photo and one number.
Recognises handwritten characters from scanned input. A CNN trained during my Code Alpha internship, reaching 94% on the held-out set.
Predicts house prices from property features with linear regression, reaching R² > 0.85. Built during my Prodigy Info Tech placement; most of the work went into feature engineering and finding which attributes actually move the price.
An SVM image classifier, built during my Prodigy Info Tech internship to see how far classical ML gets on a task CNNs own outright. 88% on the test set, with hand-designed features doing the work.
