AI / ML Engineer · Founder · Delhi, India

Yashvardhan Goel

I build production AI agents and research the problems they expose, from memory architectures serving thousands to benchmarks and interpretability work on how agents anticipate and fail.

View projects Let's talk
Scroll
01

About

Most of my work lives where AI meets the real world. I build agents that remember what matters, find the right information, and take action on their own, and I make sure they hold up when people depend on them. Putting these systems in front of thousands of users reveals problems no demo ever shows, and those problems have become my research.

Today I work on both sides: building agents that are fast and dependable, and studying how they think ahead, where they go wrong, and why. What I enjoy most is owning the full arc, from a rough question to a system running quietly in production.

M.Sc. Biological Sciences
BITS Pilani · 2020–2025
B.E. Manufacturing Engineering
BITS Pilani · 2021–2025
02

Experience

  1. Jun 2026 – Present Delhi, India

    Founder

    Forefetch

    • An anticipatory memory system for multi-agent AI pipelines. Instead of retrieving context only when an agent explicitly asks for it, Forefetch learns the temporal patterns of a pipeline, such as which agents typically run after one another and how sessions usually begin.
    • It uses these patterns to pre-stage the context an agent is likely to need next, reducing retrieval latency and making multi-agent workflows faster.
  2. Apr 2026 – Jun 2026 Bengaluru, India

    Co-Founder & CTO

    Virel

    • Built an AI-powered content intelligence platform for creators that analyzes competitor videos to surface the hooks, formats, pacing and patterns that actually work.
    • Turned those insights into creator-ready scripts in 90+ languages.
    • The product is still publicly available, but I've stepped back from active development to focus on anticipatory memory systems.
  3. Nov 2025 – Apr 2026 Singapore

    Software Developer, AI / ML

    Mira · Aroha Labs

    • Built production-grade AI agents on the Agno framework with custom memory architectures and Milvus-powered hybrid retrieval for context-aware reasoning.
    • Designed a learning pipeline that lets agents improve through an experiential feedback loop, using Natural Language Inference and Human-in-the-Loop systems.
    • Engineered MCP-compatible integrations with Google Workspace, M365, HubSpot and more, enabling cross-platform agent actions.
    • Architected event-driven microservices on AWS using SQS/SNS, ECS, ALB and VPC-isolated services, with CI/CD via GitHub Actions.
  4. May 2024 – Oct 2025 Singapore

    AI Engineer Intern

    Mira · Aroha Labs

    • Led the AI backend for Delphi Intelligence's chatbot using LangChain, Milvus, hybrid retrieval, advanced memory and intelligent routing for fast, context-aware answers.
    • Pioneered Mira Flows, a plug-and-play AI platform now serving 5,000+ users, on a microservice architecture.
    • Built a Web3 security agent over the GoPlus APIs using the Model Context Protocol for real-time, automated security intelligence.
  5. Jan 2024 – Apr 2024 Bengaluru, India

    Machine Learning Intern

    Rocketium · Mycelium Technologies

    • Fine-tuned a U²-Net model to analyze digital ads, driving a 50% increase in engagement.
    • Benchmarked Azure, AWS Rekognition and Google Vision for image tagging, and automated data extraction with Google Apps Script.
  6. May 2023 – Jul 2023 Assam, India

    Machine Learning Intern

    Hindustan Unilever (HUL)

    • Partnered with a Big Four firm on ML-driven labor allocation, saving ~1.5 hours daily.
    • Used a GBM regressor for quality checks, cutting cycle time by 20 minutes, and prototyped an energy-sustainability model to reduce carbon footprint.
03

Projects

A few things I've shipped and explored: agents, apps, and research tooling.

Benchmark · Agents 2026 →

Antbench

JANUS asked whether prefetch is measurable in synthetic workflows; Antbench asks whether it survives contact with real software. It instruments 57 supervisor–worker delegations across 9 repositories from SWE-bench Verified, with ground truth derived mechanically from diffs and test runs rather than from a judge. Workers turn out to waste 34.6% of their retrievals, and free IDF matching recovers half of what a worker will actually need (recall@8 of 0.50 against 0.06 for random, p < 0.0001).

  • SWE-bench
  • Agent Delegation
  • Retrieval
View on GitHub
Benchmark · Quantization 2026

quant_cliff

A study of how agentic tool use degrades under quantization: 18 configs, two Qwen3 sizes, 193 tasks, with schema validity, tool selection and argument correctness each scored deterministically and no LLM judge anywhere. The finding is that the cliff is model-dependent, not precision-dependent. At Q3_K_M the 1.7B loses 23.8 points of task success while the 4B moves 0.5, yet perplexity reports the same 21% degradation for both. Domain-matched imatrix calibration wins 18.7 of those points back at zero size cost.

  • Quantization
  • Tool Use
  • Evaluation
View on GitHub
Benchmark · Python 2026 →

JANUS

An anticipatory-memory benchmark scoring both halves of memory: recall (finding the right past fact, auto-graded against known answers, no AI judge) and prefetch (having the next-needed context ready before it's asked, graded against a computable Bayes-optimal ceiling). Data is generated, not borrowed, so every answer and ceiling is known exactly.

  • Memory Systems
  • Benchmarking
  • NumPy
View on GitHub
MCP · Python 2026

Handoff

A lightweight MCP server that persists conversation-turn summaries to a shared CLAUDE.md, keeping Claude Code and Desktop in sync across sessions, later extended for multi-user collaboration over HTTPS with a platform-scoped CRUD API, fcntl locking and atomic writes for data integrity.

  • Model Context Protocol
  • Concurrency
  • Multi-agent
View on GitHub
Android · AI 2025 →

TL;DAI

Too Long; Did AI. An Android app aggregating AI & tech updates from arXiv, Reddit, GitHub Trending and Hacker News, with an AI pipeline that distills long papers into concise summaries. Privacy-first, with local-only preferences.

  • Summarization
  • Mobile
  • Privacy
View on Play Store
Research · GNN 2023 – 24

Computational Discovery of Battery Materials

Adapted deep-learning frameworks to predict chemical-compound properties at 90% accuracy, built a BERT-based extraction tool for mining research papers, and used hierarchical GNNs to surface novel battery materials. Supervised by Prof. Tribeni Roy.

  • Graph Neural Nets
  • BERT
  • Materials
NLP · BioInformatics 2023

Mining Biological Literature

Extracted research on rice-associated genes via NLP and the PubMed API, fine-tuned a BERT classifier on a self-curated dataset to label genes by isolation method, and shipped a database-driven catalog for researchers. Supervised by Prof. Rita Sharma.

  • PubMed API
  • Fine-tuning
  • Web
04

Blogs

Longer write-ups on Substack: what my experiments found, what broke, and what I think it means.

Read on Substack
05

On YouTube

I document my AI-engineering journey in public as frenzyScholar: weekly experiments in agent memory, retrieval and benchmarks, plus the wins and failures along the way.

Watch on YouTube
06

Research & honors

Sep – Nov 2025

TEEP Scholar, Taiwan

Ministry of Education, Govt. of Taiwan · Yilan County

Selected for the Taiwan Experience Education Program. Developed hybrid LSTM-GRU models to forecast time-series output voltage across microbial fuel-cell setups. Supervised by Prof. Wang Chin-Tsan.

Oct 2023 – May 2025

SPARKLE Research Fellow

“Let's Rewire”: EV battery charging with ML · Rajasthan

Received a ₹2 Lakh grant and led a 10-member team through experiments, data collection and model development.

Oct 2018

PRMO Scholar

Pre-Regional Mathematics Olympiad · Delhi

Among the top 200 from the Delhi region to qualify for the Regional Mathematics Olympiad.

07

Toolkit

AI / ML

  • Agentic AI
  • LLMs
  • Machine Learning
  • NLP
  • Deep Learning
  • Computer Vision

Backend & infra

  • FastAPI
  • Docker
  • Nginx
  • AWS
  • GitHub Actions
  • OOP

Data & tooling

  • PostgreSQL
  • Redis
  • Git
  • Milvus
  • LangChain
  • MCP

Open to opportunities & collaborations

Let's build something worth remembering.