I build systems
that remember.

A memory engine, an agent that refuses to guess, moderation you can audit. Four of them are running right now — every screenshot below is the live thing.

Ø1Recall Ø2Aegis Ø3PDFChat Ø4Code-Guardian

Kunal Kumar.

CS @ BITS Pilani '27 · frontend at Emergent (YC S24) · Bengaluru

READ THE WORK → RÉSUMÉ.PDF ↗

Ø1

Profile

識別

I build systems that reach production — not demos.

A memory engine for unfinished work. An agent that refuses to hallucinate. An arena where machines learn to negotiate. A moderation pipeline whose every decision you can open up and audit.

I work where memory, retrieval and reinforcement learning meet things that actually ship — and stay running long after the demo. Lately I'm chasing what makes them safe: I'm through to round 2 at MATS, and I read alignment work the way other people read the news.

Before the AI work there were two years of community: Hack Club Bengaluru, 500+ participants fed and online, and a teaching assistantship that put 50+ students through algorithms. The through-line is the same — build the thing that remembers, then hand it to people.

「作ったものが、なぜ作られたかを覚えている」Systems that remember why they were built.

◆ SPECIFICATION / 仕様

hostBengaluru, India

processFrontend Engineering Intern — Emergent (YC S24)

studyCS @ BITS Pilani — class of 2027

focusAI safety · MATS round 2

honoursActivate AI Fellowship — India's first cohort, 1 of 15 · HPAIR delegate, Harvard

stateopen to opportunities

local time--:--:-- IST

Ø2

Works

作品 — 4 SHIPPED, ALL LIVE

PLATE Ø1

LIVE

web-seven-puce-99.vercel.app

Recall

// because I kept losing my own unfinished work.

Work scatters — the tabs, the files, the half-finished chat — and the context evaporates between sessions. Recall is a local-first continuity engine: a browser extension and watchers stream activity into a daemon that stores events, scores recovery and ranks what mattered, with semantic search across all of it. A Qt launcher hands the thread back — one keystroke, 100% on your machine. No cloud, no telemetry, plain files you can read and delete.

↳ solo build — extension → daemon → recovery interface, end to end.

pythonqtbrowser extensiondaemon apissemantic search

LIVE ↗CODE ↗

0

signal streams — browser · file · dev

LOCAL

first — nothing leaves the machine

macOS

site & download shipped

PLATE Ø2

LIVE

aegis-one-livid.vercel.app

Aegis

// scaling trust shouldn’t mean scaling the moderation team.

A kids’ platform and an adult forum need different lines, and black-box calls can’t be appealed. Aegis scores content across 7 harm categories, applies a per-platform policy of thresholds and custom rules, then routes by calibrated confidence — allow, review, or block. Every decision keeps the offending span, the reasoning and the full routing trace, so it can be audited.

↳ solo — engine interface, policy system, review queue, dashboard.

next.js 16react 19typescriptprisma · sqlitevitest

LIVE ↗CODE ↗

0

harm categories, scored 0–1

0

routes — allow / review / block

FULL

audit trail — span · reasoning · trace

PLATE Ø3

LIVE

pdf-reviewer-ten.vercel.app

PDFChat

// agents lie when retrieval gets messy.

A confident wrong answer costs more than no answer. PDFChat extracts, chunks and indexes locally, rewrites queries against the relevant chunks, and every response either cites its pages or refuses clearly. Multi-turn chat keeps session context for retrieval only, and multilingual grounding answers in your language while still treating the document as the sole evidence.

↳ built the full retrieval → citation → refusal path.

pythonretrievalcitationsrefusal policymultilingual

LIVE ↗CODE ↗

CITED

every answer points at its pages

MULTI

lingual — one source of truth

50MB

per document upload

PLATE Ø4

LIVE

zeus1205-codeguardian-ai.hf.space

Code-Guardian

// a scanner report is not a fix.

Security tooling buries teams in findings; the hard part is judging what matters and doing something about it. Code-Guardian is an agentic system for security and maintenance work on real codebases, built on OpenEnv — agents investigate the repository, reason about risk in context, and propose concrete changes instead of dumping a list. Public demo on Hugging Face Spaces.

↳ agent design, environment and the demo deployment.

agentsopenenvpythonhf spaces

LIVE ↗CODE ↗

AGENTIC

investigation over static scanning

OPENENV

the environment agents act in

DEMO

public, on HF Spaces

◆ ALSO BUILT / その他 — no public artefact

Watershed

A multi-agent RL arena for scarce resources — GRPO over 100k+ self-play episodes, with trust dynamics and coalition formation modelled into the reward, served behind FastAPI. Run it long enough and the agents stopped simply competing: alliances formed.

rl · grpotrlfastapimulti-agent

private repository

Support Triage

A deterministic support agent that answers from the docs or refuses, every time — hybrid BM25 + TF-IDF retrieval with reranking, multilingual injection and retrieval-poisoning defense, PII redaction and calibrated confidence. 132/132 evaluation tests passing.

llamaindexbm25 · tf-idfrerankingprompt defense

CODE ↗

Ø3

Runtime

経歴 — WHERE I'VE BEEN

◆ SWIPE THE TABLE →

pid	process	role	since	state
001	Emergentyc s24 · ai startup	frontend engineering intern	jun 2026 —	RUNNING
002	Open Source Clubscaler school of technology	operations lead	2026 —	RUNNING
003	Aspirants Relationsscaler school of technology	head of immersions	2026 —	RUNNING
004	Ostrius AIshipped auth & verification end to end	swe intern	apr–jun 2026	COMPLETE
005	Hack Clubcampfire · daydream · codeday bengaluru — 500+ participants	growth & ops lead	nov 2024 – mar 2026	COMPLETE
006	DSA Teaching50+ students mentored — algorithms & structured thinking	teaching assistant	may–jun 2026	COMPLETE
007	Apple Dev Academyapplication in progress	applicant	2027 cohort	QUEUED

◆ COMMIT MAP — LAST 12 MONTHS / 記録GITHUB / KUNALKUMAR-13 ↗

Ø4

Second Brain

第二の脳 — KUNAL_HQ v10

July 2026: an internship, college, three applications and a life — all carried in working memory. The result was fog. The fix wasn't discipline; it was a place. So I built one: an external brain that sorts, remembers, and shows only what matters now.

vaultplain-text files on my machine — the single source of truth. journal, habits, goals, projects, reviews.

agentan AI that reads and writes the vault — plus a silent pulse at 8:23 every morning that pulls the world in: weather, calendar, GitHub ×2, LeetCode.

pageone dashboard that mirrors it all — and always opens to exactly one thing.

capture from any terminal: $ brain <thought> — two seconds, zero judgment, sorted later.

the game layer earns XP only from real, verified data — never invented. it runs my mornings; this section is just its shadow.

◆ NEURAL LINK — kunal_hqLOCAL DEMO

Ø5

Modules

道具 — WHAT I REACH FOR

interface

界面

react
typescript
next.js

api

通信

fastapi · node
pydantic · sql
go — learning
java

intelligence

知能

llm orchestration
rag · retrieval
rl · agents · faiss
pytorch · llamaindex — daily

infra

基盤

docker · k8s — when it matters
aws · linux
cloudflare · git

Ø6

Colophon

奥付

SAY
HELLO.

click to copy · click again to open mail

GITHUB ↗ LINKEDIN ↗ RÉSUMÉ.PDF ↗

● available for 2026 — full-time & collabs

Typeface / 書体: Oswald — condensed display
Archivo — running text
JetBrains Mono — labels
Built with / 制作: Hand-written HTML, CSS & vanilla JS.
No framework. No build step.
One page, one script.
Plates / 図版: Every screenshot is the actual running deployment, captured at 1280×800 — no mockups, no renders. They sit in halftone until you point at one, then develop into colour.
Live data / 実データ: Bengaluru clock, GitHub commit map and streaks are fetched, not typed. If a feed is down it says so.
Field guide / 秘密: ⌘K — command palette
1–4 — jump to a plate
G — leave a mark

Kunal Kumar.

Profile

Works

Recall

Aegis

PDFChat

Code-Guardian

Runtime

Second Brain

Modules

interface

api

intelligence

infra

Colophon

SAYHELLO.

SAY
HELLO.