▶ Play it yourself — steer with arrows or WASD, pick a theme, or hand it to the autopilot that plans with BFS and refuses moves that would box it in. Same commit calendar, five themes, no build step.
class KiranNarayanareddy:
def __init__(self):
self.location = "Delft, Netherlands"
self.role = "Builder of autonomous, auditable agent systems"
self.site = "https://kiranreddy.nl"
self.focus = [
"Agent harnesses with hard safety boundaries",
"Aerospace & physical-AI research pipelines",
"Compliance-as-code for EU regulated products",
]
self.method = "Six-week wonder → working prototype → evidence, or none of it ships"
@property
def skills(self):
return {
"languages": ["TypeScript", "Python", "JavaScript", "Rust", "Go", "Kotlin"],
"agents": ["Claude Code", "Codex", "MCP", "multi-agent orchestration", "sandboxed execution"],
"infra": ["GitHub Actions", "GitHub Pages", "Docker", "AWS CDK", "Serverless"],
"domains": ["aerospace", "EU AI Act / GPSR / DSA", "fintech", "accessibility", "robotics"],
}
@property
def ship_rate(self):
return "A new public repository nearly every day, most with a live Pages build"I build in public and I publish the caveats. Most of what follows is a prototype with a working slice — each repo's README states exactly what is real, what is mocked, and what is not implemented yet. That honesty is the point of the whole portfolio.
Agent harnesses with hard boundaries · research pipelines where every number traces to a recorded computation · compliance gates that block instead of warn
| If you'd hire me to… | …this is the evidence | |
|---|---|---|
| 🛰️ | Make an LLM pipeline reliable enough to trust | A nine-stage research pipeline where a stage cannot advance without an evidence hash, and a refund agent whose hard rules live outside the model so a persuasive customer can't talk it out of policy. |
| 🛡️ | Turn a regulation into something a build can actually fail | EU AI Act, GPSR and DSA expressed as executable gates — exhibit, redflag-eu-compliance, listguard, clausewindow — so the artefact is a verdict, not a PDF nobody reads. |
| 🦾 | Put an agent next to hardware without it hurting anyone | Safety envelopes, verification ledgers and domain-scoped operators (reflex-physical-ai, so101-domain-operator, jointcheck) where an unsafe action is refused before it is attempted. |
| 🧪 | Set up an agent that produces research I can defend | A manuscript pipeline with a hash-chained evidence registry and a CFD stage that refuses to submit on a labelled synthetic benchmark. |
| 🌍 | Make 50,000 repos, 12,000 tools or 38 site builds legible | Self-hosted, $0-hosting sharded indexes with sub-5 ms search and a 60 FPS 3D explorer — the scout family below. |
38 live GitHub Pages sites · 40 live deployments · 138 original repositories · 112 stars · 7 forks · 3 followers
TypeScript 49 · Python 36 · JavaScript 12 · HTML 6 · Rust 4 · Go 2
Everything below is deployed and running on GitHub Pages (or its own host), straight from CI.
| Live site | Stack | Stars | What it does | |
|---|---|---|---|---|
| 🌐 | sallie-returns-ai-agent · source | Python |
— | Omnichannel shoe-retail returns pipeline: a 6-node deterministic state machine with 48h stock holds, defect assessment and human supervisor triage |
| 🌐 | Aeroscribe · source | TypeScript |
★ 1 | Nine gated stages from literature archaeology to a submission-ready manuscript, every claim traceable to a recorded computation |
| 🌐 | toolscour · source | JavaScript |
★ 1 | 3D knowledge galaxy and stack architect over 11,000+ open-source AI tools, models, runtimes and skill packs |
| 🌐 | AI-Arsenal · source | JavaScript |
★ 1 | Schema-first, Markdown-native encyclopedia of AI engineering — readable by humans, LLMs and agents |
| 🌐 | AI-Daily · source | JavaScript |
★ 1 | A calm, curated daily briefing for the AI ecosystem, built on a Scout → Challenge → Agree editorial model |
| 🌐 | gitscour · source | Python |
★ 1 | Queryable knowledge base of 50,000+ high-star GitHub repositories, sharded and served free from Pages |
| 🌐 | Aerorepos · source | Python |
★ 1 | Curated directory of open-source aerospace engineering projects — aero, propulsion, structures, GNC, space |
| 🌐 | WebsitedesignandPrompts · source | HTML |
★ 4 | Democratising the paywalled web: reusable landing-page templates with the full build prompts |
| 🌐 | Bunqsy · source | TypeScript |
★ 2 | Always-on financial guardian: 7-agent risk oracle, proactive interventions, nightly Dream Mode |
sallie-returns-ai-agent — returns pipeline · recorded 2026-11-07 · plays inline
toolscour — 3D knowledge galaxy · recorded 2026-11-07 · plays inline
▶ 18 more recordings (click a cover to play it in the reel player)
Why the grid opens a player instead of playing here: GitHub strips JavaScript from a README and strips loading=lazy from README images, so every image downloads and nothing can be started by a click. Playing all 20 inline would cost every visitor ~44 MB; this section costs ~439 kB of covers plus the two inline recordings. Open the player to browse all of them with hover previews. Each is captioned with its own date — none is a live feed.
📸 More of the wall (pages 2–3 — all 33 reachable deployments)
Real captures taken headlessly by screenshots.yml, recaptured only when a site changes. Unreachable deployments are detected and left off the wall rather than shown broken — coverage is in data/shots-manifest.json. Clickable index.
📚 The rest of the live catalogue — every deployed site
|
|
- 🚀
hack-for-good— live · pushed 2026-11-08 - 🛰️
benchmark— live · pushed 2026-11-08 - 🧠
toolscour— live · pushed 2026-11-07 - 🧪
AI-Daily— live · pushed 2026-11-07 - 🔭
AI-Arsenal— live · pushed 2026-11-07 - ⚙️
Aeroscribe— live · pushed 2026-11-06 - 🛡️
Courserefresh— live · pushed 2026-11-05 - 🎛️
proofpilot— live · pushed 2026-11-04
Three builds, and what each one actually produced.
|
A manuscript, produced by a pipeline that can't be talked into a claim Nine gated stages run literature archaeology → hypothesis → CFD → writing. Each stage writes to an append-only, hash-chained evidence registry, and the writing stage is checked against it — so a sentence with no recorded computation behind it is a blocked build, not a footnote. The CFD stage calls SU2 with a grid-convergence check. Without a real SU2 binary it runs a labelled synthetic benchmark and refuses to release the manuscript. That refusal is the feature: the pipeline produced a submission-ready manuscript for the cases it could verify, and would rather block than guess.
|
A refunds agent that talks itself out of policy — and the breaker that catches it Built at Agents Gone Rogue (SAAF × Just Eat Takeaway × TAG, Amsterdam). The agent is scored on CSAT, writes itself lessons, and rewrites its own instructions every four tickets — so by ticket 36 the genesis policy is no longer in the prompt at all. Priya asks for €21.35 against a €20 cap with no photo and a 1-star threat. The breaker is four stages — forge, rehearse, sentinel, verify — running against a real support workflow: a 36-ticket rehearsal over 16 personas, then the live Sentinel calling the real TypeSafe Jev API. Hard rules live outside the model, so a Jev With the watchdog off the agent leaks money. With it on, it pays €0.00 against policy. The workpaper is then re-verified offline with zero model calls: alter a verdict without rehashing and verification fails.
|
Placed at bunq Hackathon 7.0 — an always-on financial guardian BUNQSY doesn't wait for a prompt. A 30-second heartbeat syncs every account, computes a 0–100 health score from balance, velocity, goals and upcoming obligations, and puts seven concurrent risk sub-agents to a vote. Cross the threshold and it acts: voice narration, plan-gated card freezes, drafts cancelled — every action routed through PLAN → CONFIRM → EXECUTE. At night, Dream Mode analyses patterns in an isolated worker and returns a DNA card plus three suggestions. Built on the real bunq REST API with RSA-2048 request signing.
|
| Track | The idea | Flagships | |
|---|---|---|---|
| 🛰️ | Agent harnesses | Frontier models are only as good as the cage around them: confinement rings, isolation, durable state, review loops | apex · Harness-Zero · Autoorg |
| 🚀 | Aerospace research | An autonomous research lab where every quantitative claim traces back to a recorded computation | Aeroscribe · Aerorepos |
| 🛡️ | Compliance-as-code | Regulation as an executable gate, not a PDF: EU AI Act, GPSR, DSA evidence packs | exhibit · redflag-eu-compliance · listguard · clausewindow |
| 🦾 | Physical AI | Safety envelopes, verification ledgers and domain-scoped operators for robots that touch the real world | reflex-physical-ai · so101-domain-operator · jointcheck |
| 💸 | Agentic finance | Agents that act before damage is done, with a gate between proposal and execution | Bunqsy · Guardian · LEGACYVAULTV1.0 |
| ♿ | Accessibility | Interfaces that adapt to the person instead of asking the person to adapt | clarion · accesshold · LitPlay |
Built under a clock, shipped to a live demo, documented afterwards.
| Event | What shipped |
|---|---|
| bunq Hackathon 7.0 | BUNQSY — always-on financial guardian with a 7-agent risk oracle |
| Vultr Cloud Hackathon | SENTINEL — fail-closed financial oversight agent behind a 3-signal gate |
| Slack Agent Builder Challenge | Clarion — accessibility agent, live-tested in a real Slack sandbox |
| Prosus Build Weekend | Magpie — autonomous classifieds scout that haggles and closes |
| TechEx Amsterdam | Agentcrash — SAAF policy circuit breaker for support agents |
| TwitchCon Codex Hackathon | twitch-codex — four live-audience, control and accessibility builds |
| OneGov / Alkemio | onegov2-spatial-assistant — Dutch spatial-assistant challenge entry |
| F1 × Elastic × Hopsworks | F1Elastichopsworkshack — explorable 2023–24 race workspace |
Take something too large to browse, shard it until it fits in a static file, and give it an interface you can actually think in.
|
🛰️ toolscour — 12,771 AI tools, models, runtimes and skill packs across 10 domain sectors, sharded into a < 1.8 MB gzip tier-1 index with sub-5 ms tokenized search. The 3D knowledge galaxy is a hand-rolled 60 FPS canvas projection — no Three.js — and the stack architect scores a proposed pipeline for hardware, primitive and licence compatibility.
|
🔭 gitscour — 50,000+ repositories above 500 stars, classified on four orthogonal axes and queried in-browser through a WebAssembly SQL console. Read-only is enforced by validating every
|
|
🧰 AI-Arsenal — a 1,598-file encyclopedia where every entry is Markdown with schema-validated YAML frontmatter, so the same repo is a browsable knowledge base for a human and a queryable data API for an agent. Structure first, UI later.
|
🎨 WebsitedesignandPrompts — the paywalled web, rebuilt in public: 38 live templates, each shipped with the full prompt that builds it, a build log and an adaptation guide. Cinematic scroll stories, 3D WebGL camera flights, glassmorphic product pages — engineering as reproducible craft rather than inspiration screenshots.
|
These are live windows: the newest capture inside a frame that animates — a pulsing LIVE pill, a sweeping scanline — because GitHub strips <iframe>. Captured by screenshots.yml — the same pipeline that builds the wall above, so these can't go stale.
| Project | Why it exists | |
|---|---|---|
| 🛰️ | toolscour | 11,000+ AI tools mapped as a 3D galaxy, with a stack architect that scores compatibility |
| 🔭 | gitscour | The same sharding pipeline applied to 50,000+ high-star repos |
| 🧰 | AI-Arsenal | A machine-readable encyclopedia an agent can consume as easily as a human |
| 📰 | AI-Daily | Scout → Challenge → Agree, so the news you read was argued with first |
| 🎨 | WebsitedesignandPrompts | The paywalled web, rebuilt as reusable templates plus the prompts that make them |
| 🧪 | 3dexperimentation | Pure-code animated film studio — Canvas, no assets, 60 FPS |
Drawn by build_badges.py — no third-party badge service, so nothing here can break.
▶ Signal dashboard + contribution terrain — hover a bar, lift a block, switch theme · play the contribution snake · browse the 20 recorded demos
One self-rendered card, no third-party widgets, drawn by build_dashboard.py from the GitHub contributions API — so the numbers are exact. This is an image: GitHub serves README images through its Camo proxy as <img>, which delivers no pointer events, so nothing inside it can react. The hoverable version is here.
371 blocks, one per day, height = that day's output. Drawn by build_3d_terrain.py from the same calendar the dashboard uses. This is an image too, for the same reason — Camo serves it as <img>. Hover any block to lift it and read its exact date in the interactive version.
kiranreddy.nl · Delft, Netherlands · Autonomous agents · Aerospace & physical AI · Compliance-as-code
Live sites, counters and cards refresh themselves daily at 04:17 UTC via GitHub Actions — see scripts/build_profile.py