A construction materials trading startup was running leads, orders, payments, and vendor relationships off spreadsheets and WhatsApp — no role separation, no credit visibility, no site-stage tracking tied to demand.

A full-stack CRM (FastAPI backend, MongoDB via Motor, React Native/Expo as a web-first PWA) with role-based access for Admin, Sales, Accounts, and Operations. Tracks leads through a sales funnel, generates estimates/sales orders/purchase orders as PDFs, records customer-wise payments with overdue alerts, and maps each customer's site through a 12-stage construction timeline (Excavation → Completion) for predictive material demand.

Differential diagnosis tools that optimize only top-1 accuracy hide how confident a model actually is. A system that's 95% sure it's right looks identical to one that's barely ahead of its second guess — until a patient gets the wrong one.

Fine-tuned DistilBERT and PubMedBERT on DDXPlus (1M+ simulated patient cases, 49 pathologies) using KL-divergence soft-label training to optimize the full probability distribution, not just the top guess. Wrapped into a multimodal app: BERT symptom diagnosis, chest X-ray screening, OCR-based lab report analysis, and natural-language explanations across 5 LLM providers.

Pre-development documentation — architecture decisions, scope breakdowns, tech stack rationale — was eating 3–5 days per project at the agency. Engineers were writing the same boilerplate from scratch every time.

A multimodal AI pipeline that ingests a project brief, reasons over it using GPT-4 and Claude in parallel, and outputs structured docs with D3.js dependency graphs and a recommended stack. One prompt → full pre-dev package.

Node.js project setup — scaffolding, TypeScript migration, Docker config, linting — is repetitive manual work that eats the first hours of every new project.

A multi-tool MCP server built on the early FastMCP SDK, five months after Anthropic released the MCP specification. Exposes scaffolding for Express/Next/React/TS, automated JS→TS conversion (including JSX→TSX), Docker and ESLint/Prettier setup, test generation, and live npm registry lookups — plus proper MCP resources and lifespan management.

Students couldn't find specific moments in long lecture videos. Scrubbing through 2-hour recordings to find a 30-second explanation was killing retention and study efficiency.

Whisper for transcription + Tesseract & EasyOCR for slide text extraction, combined into a timestamp-indexed search engine. Students type a concept, get the exact minute and a summary. Containerised with Docker & K8s for scale.

Skin cancer affects 3M+ people yearly per WHO. Dermatologists rely on visual assessment which is prone to bias — especially across 7 lesion types that share similar visual characteristics. Rural access to specialists is limited.

CNN trained on the combined HAM10000 + ISIC 2019 dataset (35K+ images, oversampled to ~52K) classifying 7 lesion types: Melanoma, BCC, SCC, Actinic Keratoses, Vascular Lesions, Melanocytic Nevi, Dermatofibroma. Oversampling + data augmentation resolved class imbalance. Deployed as a Flask/Streamlit web app for real-time clinical screening. Published in IJCNIS Vol 16(5), 2024.

Breast cancer prediction models often overfit or underperform on skewed clinical datasets. Achieving high accuracy without sacrificing sensitivity required careful feature engineering and normalisation.

Trained on 10,000+ patient records. Applied advanced normalisation techniques that reduced data skewness by 18%, improving generalisation. Compared SciKit-Learn ensemble methods against TensorFlow deep models, selecting the best-performing architecture per metric.

Existing job boards surface irrelevant listings — wrong city, wrong experience level, wrong stack. Fresh graduates waste hours filtering noise instead of applying to real matches.

A Selenium + BeautifulSoup scraper with layered filters: location radius, required skills intersection, experience bracket, and salary range. Results ranked by match score. Data deduplication and normalisation improved pipeline efficiency by 15%.