DeepeshAhlawat

Applied ML Intern · Aftershoot

Aug 2026 — Present

New Delhi, India · On-site

Turning recent computer-vision research into working systems for a product that edits photos like a human would.

• Implement state-of-the-art computer-vision papers from scratch, reproducing results and adapting them to production constraints.
• Bridge the gap between research and product, translating academic methods into models that hold up on real-world photography at scale.

Agentic AI Intern · ConnectHEOR

Jan 2026 — Jul 2026

London, UK · Remote

Choosing the frameworks, automating the tuning, and shipping agents to production on AWS.

• Benchmarked DSPy against LangChain for agent development, weighing reliability, control, and room to optimise — then built an automated prompt-optimisation pipeline on the GEPA and MIPRO optimizers that improves agent performance without hand-tuning a single prompt.
• Deployed and auto-provisioned agents and agentic harnesses on AWS Bedrock, turning a manual setup into a repeatable production workflow.
• Pushed emerging tooling into the stack (Claude for Excel, Claude Code + Manim) to extend agents into spreadsheet automation and programmatic visualisation.

Physics-informed neural networks for space-weather detection, validated against international catalogues and headed for peer review.

• Built and trained an end-to-end deep learning system (PyTorch) over large-scale time-series solar-plasma data, owning the pipeline from preprocessing and feature engineering through validation and optimisation.
• Applied symbolic regression to uncover hidden variables in the Dst equation, sharpening how solar-wind activity is modelled as it drives geomagnetic disturbance in Earth's magnetosphere.
• Folded those physical constraints into a PINN that outperformed the existing ARCANE baseline for coronal mass ejection detection — 94.5% true-negative rate and 63.1% detection against the Richardson/Cane ICME catalogues.
• Authoring the manuscript on methodology and results, targeted at an international peer-reviewed journal.

AI Intern · DigitalPaani

Apr 2025 — Sep 2025

Gurugram, India · On-site

Shipped production AI end-to-end, data pipeline to cloud: conversational AI, real-time anomaly detection, document intelligence.

• Shipped a full-stack conversational AI assistant to production — FastAPI backend, OpenAI APIs, Oracle Cloud, complete REST integration.
• Built and deployed a real-time anomaly-detection service for live industrial signals, owning preprocessing, training, evaluation, and cloud deployment.
• Engineered LLM document-parsing pipelines that convert unstructured text into structured, rule-compliant formats, eliminating manual processing across 10+ document categories.
• Built production RAG pipelines (LangChain + Pinecone) to query and synthesise insights from dense technical documentation.