Ekansh Sharma

Problem. Industrial requirement documents (Lastenhefte) are dense, bilingual, and full of tables and diagrams that text-only search cannot read, and in regulated industries every answer has to be traceable back to its source.

Outcome. Built an end-to-end multimodal RAG over 26 German and English industrial PDFs, using ColPali for visual page retrieval and a LoRA-fine-tuned BGE-reranker-v2-m3. Fine-tuning on 714 synthetic DE and EN technical queries lifted Hit@1 by 8.4 points and MRR by 5.1 over the off-the-shelf reranker. A LangGraph multi-agent setup (planner, retriever, validator, synthesizer) routes between local and API models, and the system is built to the EU AI Act and GDPR: Article 6 risk classification, Article 13 transparency, Article 17 right to erasure, and a full audit trail.

• Python
• FastAPI
• Next.js
• LangGraph
• ColPali
• pgvector
• Docker
• Langfuse

Problem. Most AI coding agents write changes straight to disk, so you only find out what they did after they have already done it.

Outcome. Built a terminal coding agent from scratch in TypeScript and Bun. You give it a goal in plain language; it reads and searches your codebase with sandboxed tools, then stages every edit in an in-memory overlay so you can review a unified diff and approve or reject it before anything touches disk. The same engine runs as an interactive CLI and a Telegram bot with inline approve and reject buttons, backed by a full unit test suite of 50 tests, strict TypeScript, GitHub Actions CI, a Dockerfile for one-command runs, automatic cost tracking from OpenRouter billing, and a benchmark harness that scores success rate and cost per task.

• Bun
• TypeScript
• Vercel AI SDK
• OpenRouter
• Telegraf
• Firecrawl

Problem. Driving footage carries segmentation, depth, and motion signals, but combining them into a single trainable pipeline is fiddly.

Outcome. Built a multi-task pipeline in CARLA: DeepLabv3-ResNet50 for segmentation, monocular depth estimation, and GRU/ConvGRU heads for ego-motion. Reported IoU and depth metrics with trajectory and point-cloud visualisations.

• PyTorch
• CARLA
• OpenCV