Yash Kumar profile photo 4 on Tal
Yash Kumar profile photo 1 on Tal
Yash Kumar profile photo 2 on Tal
Yash Kumar profile photo 3 on Tal
Yash Kumar profile photo 4 on Tal
Yash Kumar profile photo 1 on Tal

Yash Kumar

Building AgentClash | AI agent evaluation on real tasks | AI Engineer @ Akatsuki AI | GSoC ’25

Bangalore Urban

1+ year of work experience

Jun 20264 moPresent

Remote

Building an platform for brands to collab easily with artists “no more emails”
Jun 20264 mosPresent

I’m helping build AgentClash, an open-source platform for evaluating AI agents on real tasks. My focus is the bridge between product, users, and distribution: • turn agent-evaluation research into useful demos, benchmarks, and technical content • talk with teams shipping agents and turn their failure modes into better product workflows • make replays, scorecards, and regression gates understandable enough to act on • grow the open-source community around practical agent evaluation AgentClash runs candidates against the same workload, tools, sandbox, and budget, then preserves the evidence behind every pass or failure. https://www.agentclash.dev
Jan 20265 mosMay 2026

Shipped Casting Doctor AI, a bilingual (JP/EN) internal chatbot for Hakuhodo Casting & Entertainment covering talent search, booking history, regulatory NG checks, and reliability scoring; authored a structured test-prompt suite covering five product requirements across both languages. Built and deployed the Onet LINE Nurturing Bot on the OpenClaw platform — AI-driven LINE channel funnel + advisor-facing recommendation tool (FUNC-01 & FUNC-02), with prompt-injection defense and APPI compliance for Japanese enterprise data. Analyzed 514,000+ rows of Veeva Vault CRM Japanese sales-rep visit notes for Shionogi Pharma; extracted competitor drug mentions and surfaced flu-season activity spikes (e.g. Tamiflu mentions 100 → 3,100/month), and architected the end-to-end stakeholder demo pipeline. Shipped client-facing bilingual web apps to Vercel (Hakuhodo Talent & Project Data Mart, AKTSK Voice Concierge) and authored internal Claude Code skills (pptx-template-proposals, pptx-akatsuki-proposals) now used by the proposals team; built a Gmail → Slack alert bot (Apps Script + Groq) with urgency-tiered private DMs.
OpenWISP
Jun 20254 mosSep 2025

Shipped 5 features + 2 production fixes to netjsongraph.js (Leaflet + ECharts), the visualization library powering OpenWISP deployments globally. All proposed objectives merged to main. Cluster Overlap Prevention — 4-stage pipeline (GeoJSON/NetJSON converter, KDBush spatial clustering, circular layout, repulsion forces) that visually separates co-located clusters of differing status (PR #396, #349, monitoring #668). Client-Nodes Overlay — modular attachClientsOverlay API with orbital placement (concentric rings, density-aware dot count) + recursive sidebar rendering (PR #411). Zoom-dependent labels for map (zoomend) and graph (graphRoam) modes (#407, #419). Resiliency: deduplicateNodesById (#355), tile-provider-aware max-zoom (#363), scoped CSS (#374), README/GIF demos (#441). Mentored by Federico Capoano, Nitesh Sinha, Gagan Deep.
Google Summer of Code
May 20255 mosSep 2025

bangalore

Improve netjsongraph.js resiliency and visualization
GitHub
@cestercian

Yash’s AI Stack

claude code
claude code
claude
claude
codex
codex
cursor
cursor
hermes agent
hermes agent
langchain
langchain

Education

Birla Institute of Technology and Science, Pilani

Birla Institute of Technology and Science, Pilani

Indian Institute of Technology, Ropar

Specialisation In AI/ML, Computer Engineering

Indian Institute of Technology, Ropar

20241 yr2025

Join Yash and 10k+ other professionals on

Tal Boss

Tal is the app where bosses meet engineers & builders directly, without middlemen

Download Tal Boss

Yash Kumar on Tal

Headline

Building AgentClash | AI agent evaluation on real tasks | AI Engineer @ Akatsuki AI | GSoC ’25

About

AI agents look great in demos. Real tasks are where the weird failures show up. I’m building AgentClash, an open-source platform for evaluating agents under the same workload, tools, sandbox, and budget. It keeps the replay, scorecard, artifacts, and failure evidence, so teams can choose the right agent, catch regressions, and ship with proof instead of vibes. I also build production AI systems at Akatsuki AI Technologies. Before that, Google Summer of Code 2025 with OpenWISP taught me what good open-source engineering looks like: real users, testable work, and maintainable systems. I’m interested in agent evaluation, tool use, multi-turn reliability, prompt injection, agent UX, open source, and the messy work of explaining technical products clearly. If you’re shipping agents and still judging them from demos, I’d love to compare notes: https://www.agentclash.dev

Location

Bangalore Urban

AI stack

claude code

claude

codex

cursor

hermes agent

langchain

Work experience

Founding Engineer at Artistant, Jun 2026 - Present · 4 mo, Remote, Building an platform for brands to collab easily with artists “no more emails”

Builder at AgentClash, Jun 2026 - Present · 4 mos, I’m helping build AgentClash, an open-source platform for evaluating AI agents on real tasks. My focus is the bridge between product, users, and distribution: • turn agent-evaluation research into useful demos, benchmarks, and technical content • talk with teams shipping agents and turn their failure modes into better product workflows • make replays, scorecards, and regression gates understandable enough to act on • grow the open-source community around practical agent evaluation AgentClash runs candidates against the same workload, tools, sandbox, and budget, then preserves the evidence behind every pass or failure. https://www.agentclash.dev

Artificial Intelligence Engineer at Akatsuki AI Technologies, Jan 2026 - May 2026 · 5 mos, Shipped Casting Doctor AI, a bilingual (JP/EN) internal chatbot for Hakuhodo Casting & Entertainment covering talent search, booking history, regulatory NG checks, and reliability scoring; authored a structured test-prompt suite covering five product requirements across both languages. Built and deployed the Onet LINE Nurturing Bot on the OpenClaw platform — AI-driven LINE channel funnel + advisor-facing recommendation tool (FUNC-01 & FUNC-02), with prompt-injection defense and APPI compliance for Japanese enterprise data. Analyzed 514,000+ rows of Veeva Vault CRM Japanese sales-rep visit notes for Shionogi Pharma; extracted competitor drug mentions and surfaced flu-season activity spikes (e.g. Tamiflu mentions 100 → 3,100/month), and architected the end-to-end stakeholder demo pipeline. Shipped client-facing bilingual web apps to Vercel (Hakuhodo Talent & Project Data Mart, AKTSK Voice Concierge) and authored internal Claude Code skills (pptx-template-proposals, pptx-akatsuki-proposals) now used by the proposals team; built a Gmail → Slack alert bot (Apps Script + Groq) with urgency-tiered private DMs.

Open Source Developer at OpenWISP, Jun 2025 - Sep 2025 · 4 mos, Shipped 5 features + 2 production fixes to netjsongraph.js (Leaflet + ECharts), the visualization library powering OpenWISP deployments globally. All proposed objectives merged to main. Cluster Overlap Prevention — 4-stage pipeline (GeoJSON/NetJSON converter, KDBush spatial clustering, circular layout, repulsion forces) that visually separates co-located clusters of differing status (PR #396, #349, monitoring #668). Client-Nodes Overlay — modular attachClientsOverlay API with orbital placement (concentric rings, density-aware dot count) + recursive sidebar rendering (PR #411). Zoom-dependent labels for map (zoomend) and graph (graphRoam) modes (#407, #419). Resiliency: deduplicateNodesById (#355), tile-provider-aware max-zoom (#363), scoped CSS (#374), README/GIF demos (#441). Mentored by Federico Capoano, Nitesh Sinha, Gagan Deep.

Open Source Developer at Google Summer of Code, May 2025 - Sep 2025 · 5 mos, bangalore, Improve netjsongraph.js resiliency and visualization

Education

Birla Institute of Technology and Science, Pilani

Specialisation In AI/ML, Computer Engineering, Indian Institute of Technology, Ropar

Public links

github: https://github.com/cestercian