Shashank Durgad
agentic AI systems and the infra that make them measurably better
An animated sequence: a pair of eyes rendered as particles closes and bursts apart, then the fragments settle around a pair of doors. The left door leads to experience and education, the right to projects.
Work & Education
Anglo-Chinese School (Independent) — International Baccalaureate — 42/45
Jan 2017 — Dec 2022 · Singapore
- Vice-Chairman, Robotics Technology Society
- Head of Operations, Young Entrepreneurs' Society
University College London — BSc Computer Science — Predicted First Class Honours
Sep 2025 — Jun 2028 · London, UK
- UCL Data Science Society — Vice-President
- UCL AI Society — First Year Representative
- Foundry Labs (Cohort 4) · Bloomsbury Startup Academy (Demo Day Winner)
Onflow — Founding Engineer (AI and Growth)
Oct 2025 — Jul 2026 · London, UK
- Won 1st Place at the Bloomsbury Startup Academy Demo Day; 2nd Place at the Gemini 3 London Hackathon, judged by Google and industry leaders.
- Architected an agentic platform using LangChain to simulate user behaviour and generate actionable feedback, with Playwright for browser automation.
- Authored a white paper on simulated agentic web-use interaction.
Stack: Next.js, GitHub API, Playwright, LangChain, Gemini CLI
Overmind — Software Engineering Intern
Jul 2026 — Present · London, UK
- Ran end-to-end agent optimisation experiments on the Overmind platform, driving iterative prompt/tool/logic improvements through a state-machine-based optimiser with automated eval scoring.
- Working full-stack on the agent improvement pipeline — from trace/dataset ingestion and data engineering through agentic optimisation experiments, model fine-tuning and production rollout.
Stack: Python, TypeScript, Django, Docker, OpenTelemetry, PostgreSQL, Cursor SDK
scroll
Projects
Sentinel
Autonomous agents negotiating satellite collision avoidance
Environment Track — NVIDIA GTC Hack for Impact
- A decentralised system where autonomous agents negotiate satellite collision avoidance manoeuvres in seconds, reducing debris cascades.
- Dual-path collision risk classifier pairing rule-based thresholds with an XGBoost model (8 features, 4 risk classes) to capture compound interactions beyond hard rules.
- Real-time feature extraction from live CelesTrak conjunction data and NOAA space weather feeds.
Stack: Python, XGBoost, NumPy, FastAPI, Pydantic
Sentinel on GitHubSentrix
Multi-agent compliance monitoring with a patrol swarm
1st Place — UCL BUILD AI Festival
- A multi-agent AI compliance monitoring system, built around a patrol swarm that escalates findings through a staged investigation pipeline.
- LangGraph and NVIDIA Nemotron drive the swarm using pheromone-based sampling and quorum voting.
- A 4-stage investigation pipeline fed by an ETL layer ingesting A2A messages, emails, PRs and documents into SQLite, backed by NetworkX for inter-agent graph analysis.
Stack: Python, LangGraph, SQLite, NetworkX, FastAPI
Sentrix on GitHubOnflow
Simulated agentic web use, turned into UX feedback
1st Place — Bloomsbury Startup Academy Demo Day
- An agentic platform that simulates real user behaviour against a live product and turns the traces into actionable UX feedback and codegen fixes.
- LangChain drives the agents, Playwright drives the browser, and the GitHub API and Gemini CLI close the loop with automated fixes.
- Synthetic automated user testing drove a 30% conversion rate in preliminary beta.
- 2nd Place at the Gemini 3 London Hackathon, judged by Google; shortlisted for the Google AI Futures Fund interview.
Stack: Next.js, GitHub API, Playwright, LangChain, Gemini CLI
Onflow on GitHub