000
Open to remote roles · Worldwide

RL Environments Engineer

Production AI Systems

Backend & Cloud Infrastructure

Kartik Aslia

Environments that can't be gamed.Systems that stay up.

I build RL environments, graders and reward-hacking defenses for agent training — and the production AI platforms, pipelines and cloud infrastructure behind them.

Agentic task authoringAutomated gradingReward-hacking defenseTrajectory reviewDocker-in-Dockercgroups & netnsSandboxesGrafana LGTMLangGraphLangChainLangSmithGeminiNode.jsTypeScriptPythonPostgreSQLAgentic task authoringAutomated gradingReward-hacking defenseTrajectory reviewDocker-in-Dockercgroups & netnsSandboxesGrafana LGTMLangGraphLangChainLangSmithGeminiNode.jsTypeScriptPythonPostgreSQL

About

I build agentic coding environments and the QC gates that stop agents from gaming them. Before that, I was the founding engineer on three live AI products: I scaled a compliance platform to 1000+ concurrent users and moved it from Oracle Cloud to GCP on my own. Either way, I care whether it holds up under load, on deadline, with real users.

◦ Gurugram, India◦ B.E. Computer Science · Chandigarh University◦ Remote-first, worldwide
0×

Faster capture

Solo-built Docker-in-Docker farm: 50-task batch ~150h → ~4h, built in 3 days

0%

Of a 50-task delivery

36 of 50 tasks shipped personally, top of five contributors (next-highest: 10)

0+

Concurrent users

Compliance-heavy due-diligence platform, owned end-to-end

0ms

API latency

Down from 200ms via caching, async task queues and worker processing

0%

Fewer tokens

1M+ → 75K per report via Gemini File API & a shared OCR registry

0

Tables migrated

Oracle Cloud → GCP single-handedly, with KMS envelope encryption & CMEK

Journey

Roots run deep

Like mycelium under a forest floor: quiet, connected, load-bearing. Here's the network I've grown so far.

Apr 2026 — Present

Contractor · Remote

Bespoke Labs

RL Environments Engineer — Expert

Turning environment ideas into validated agentic coding tasks at scale, and keeping them honest as volume grows.

Aug 2026 — Present

Expert Reviewer

AlphaRecon

  • Review complex agentic coding environments for prompt quality, environment complexity and up-to-date research/planning context, so they teach agents to research, plan and review.
  • Gate human- and LLM-produced trajectory annotations for correctness before they enter training and evaluation sets.

Jul 2026 — Aug 2026

RL Environments Engineer

Byzanite

  • Built a parallel Docker-in-Docker capture farm solo in 3 days. The whole team adopted it: 50-task batch ~150h → ~4h (≈37×).
  • Wrote delivery-QC gates that execute every cited query, mirror-check repos and guard against internal-material leaks: 0 errors / 0 warnings on the client validator.
  • Built root-cause-analysis incident worlds on real codebases (Ethereum nonce, Solana watermark, Redis/Postgres cascades, Kafka/JWT, KYC/AML) with graders calibrated to a target pass@1.
  • Joined ~2 working days before the deadline and landed the 10-task pilot. Then shipped 36 of the 50-task final delivery (72%) across 214 commits, top of five contributors.

Apr 2026 — Jul 2026

Pod Lead → Senior Reviewer

Nebula Aurora

  • Led Pod 6: authored and QC'd agentic RL tasks against a strict client data contract; onboarded contributors and cleared blockers.
  • Promoted to Senior Reviewer. Posted the pod's highest client-accepted delivery count in a ~15–20 day window, against peers with 6+ months.

Feb 2025 — Present

Founding engineer

Clarigen AI

Director of Technology & Product Lead

First engineering hire, owning three live AI products end-to-end: architecture, AI, infrastructure and delivery under audit timelines.

  • Scaled a compliance-heavy due-diligence platform to 1000+ concurrent users and cut latency from 200ms to 50ms with caching, async task queues and workers.
  • Cut report cost 80% and tokens 93% (1M+ → 75K) with a shared OCR registry and Gemini File API inside a 6-node LangGraph pipeline.
  • Migrated 107 SQL tables, document stores, containers and 597 CDN assets from Oracle Cloud to GCP single-handedly.
  • Hardened to zero-trust: KMS envelope encryption, CMEK storage, SBOM + vulnerability gating, 200+ real-SQL integration tests.

Dec 2024 — Jan 2025

Contract

Stealth Web3 / Quant

Software Engineer

An autonomous market-intelligence engine for on-chain portfolios.

  • Fused on-chain signals, news feeds and 147+ technical indicators into risk-aware automated portfolio decisions.
  • Owned architecture and system design, trading off requirements directly with the founding team.

Jun 2024 — Nov 2024

Remote

Kishvi Studio

Software Developer Intern

Real-time backends and secure infrastructure.

  • Built a WebSocket transaction system (+22% efficiency) and a real-time broadcast server.
  • Built a microservices API gateway (+34% throughput) and an AES-256 encryption portal.

Aug 2023 — Jan 2024

Remote

Travel-tech startup (UK)

Machine Learning Engineer Intern

LLM itinerary generation with retrieval.

  • Fine-tuned an LLM with RAG and Self-RAG; an internet-connected pipeline added 47% more search context.
  • Built a token-cost logger that improved cost tracking by 30%.

Selected work

Systems I've shipped

Production · 04RL & Eval · 02Research · 01
01 / 07Production

System

  1. Data-room upload→
  2. Shared OCR registry→
  3. Async job queue→
  4. 6-node LangGraph→
  5. Human approval gate→
  6. Report cache

AI due diligence · Live · Architect & lead engineer

KarmicDD

Problem

Each analysis agent OCR'd the same data room on its own and pushed 1M+ tokens of raw text into the prompt. Reports were slow and expensive, and long requests blocked users.

Hardest part

Four agents racing to OCR the same documents. I fixed it with a per-entity, per-category registry where one run is shared and everyone else awaits it, so each document is read once however many analyses need it.

  • 1000+ concurrent users
  • 200ms → 50ms API latency
  • 93% fewer tokens, 80% lower API cost
Visit live ↗
02 / 07Production

System

  1. User task→
  2. LangGraph router→
  3. 96 tools · 11 categories→
  4. DeepSearch agent→
  5. SSE stream + recovery→
  6. Audit log

Agentic workspace · Live · Architect & lead engineer

CLAR AI Workspace

Problem

Research and document tasks run for minutes across dozens of tools. A chat-style request/response loop drops work when a connection or process dies.

Hardest part

Keeping multi-minute agent runs alive all the way to the browser. I layered crash recovery onto the SSE stream so a dropped connection doesn't throw away the run.

  • 96 tools in production
  • Up to 3M tokens / session
  • 3-layer stream crash recovery
Visit live ↗
03 / 07Production

System

  1. Word task pane→
  2. Entra SSO→
  3. Action router · 28 types→
  4. Worker-thread pipelines→
  5. ChangeTracker→
  6. Apply to document

Office.js add-in · Live · Product & engineering lead

CLAR for Word

Problem

Enterprise users won't leave Word, and their IT teams won't allow an AI tool that skips corporate sign-on or edits documents without a record.

Hardest part

Running translation and grammar over long contracts without stalling the API. Moving those jobs onto worker threads keeps the event loop free for every other user.

  • 28 document action types
  • Enterprise SSO via Entra
  • Full audit trail of AI edits
Visit live ↗
04 / 07RL & Eval

System

  1. Task batch→
  2. Slot scheduler→
  3. DinD workers→
  4. Time-shift→
  5. Parquet merge→
  6. LGTM stack

RL infrastructure · Sole engineer · built in 3 days

Capture Farm

Problem

Each RL environment's telemetry had to be captured one at a time. A 50-task batch took ~150h, longer than the delivery window.

Hardest part

Scaling past a few parallel slots hit Docker address-pool exhaustion, a cgroup-delegation race and slot collisions. I cleared all three to make the farm safe at full width.

  • ≈37× faster: 150h → 4h
  • 10-task batch: 21h → 4h
  • Adopted by the whole team
05 / 07RL & Eval

System

  1. Task package→
  2. Schema & enum checks→
  3. Execute cited queries→
  4. Git-mirror check→
  5. Leak guard→
  6. Client validator

Reward-hacking defense · Author · top contributor on delivery

Delivery-QC Gates

Problem

With five contributors racing to ship 50 tasks, manual review couldn't reliably catch grader loopholes, citations that didn't resolve, or internal material leaking into a package.

Hardest part

Citations that looked correct but returned nothing. The gate re-runs every cited query and fails the build if the evidence isn't there, which closed a whole class of grader loophole.

  • 0 errors / 0 warnings on client validator
  • 36 of 50 tasks shipped personally (72%)
  • 19/19 environments salvaged, zero recapture
06 / 07Production

System

  1. Oracle Cloud→
  2. Schema migrations→
  3. Cloud SQL + CMEK GCS→
  4. KMS envelope encryption→
  5. SBOM / Grype gate→
  6. Cloud Armor + IAP

Cloud & security · Sole engineer

Oracle → GCP Migration

Problem

A compliance platform with regulated customer data had to leave Oracle Cloud without data loss, and come out with a security posture auditors would accept.

Hardest part

Proving the migrated system behaved like the old one. 200+ integration tests ran against real SQL instead of mocks, so schema drift showed up before cutover.

  • 107 tables + 597 CDN assets migrated
  • Zero-trust: KMS + CMEK + IAP
  • 200+ real-SQL integration tests
07 / 07Research

System

  1. KSP telemetry→
  2. Sequence windows→
  3. LSTM layers→
  4. Dense head→
  5. Position estimate

Research · ML · Researcher

Dense-LSTM Spacecraft Estimator

Problem

Without GPS, a spacecraft has to estimate its position from its own telemetry, in real time, on limited onboard compute.

Hardest part

Holding accuracy while cutting compute. A shallow hybrid replaced a deep recurrent stack, so it stays light enough for real time.

  • 99.6% accuracy
  • −92% compute time
Repository ↗

Next chapter

Your hardest problem
could be card eight.

RL environments, evaluation infrastructure or production AI systems — I'm open to remote roles worldwide.

How I build

Four rules I ship by

01

Execute, don't trust

My delivery-QC gate runs every cited query instead of reading the text, which took a client package to 0 errors and 0 warnings.

02

Fail closed

The telemetry slot-shift refuses to merge on any time-window collision, and payment access is re-verified with the provider instead of read from a cached row.

03

Pay for expensive work once

One shared OCR pass feeds every due-diligence agent, and documents go in through the File API instead of the prompt, taking reports from 1M+ to 75K tokens.

04

Build the tool the team is blocked on

When serial capture was going to take ~150h, I spent 3 days on a parallel farm that the whole team then used, bringing it to ~4h.

Ecosystem

Tools, used in anger

RL Environments & Eval

01

Writing agentic tasks hard enough to matter, and graders strict enough that they can't be gamed.

Agentic task authoringAutomated gradingReward-hacking defenseTrajectory reviewpass@1 calibrationSWE-benchTerminal-BenchHarbor

ProofQC gates that execute every cited query took a client package to 0 errors / 0 warnings.

Platform & Infra

02

Isolated, parallel, observable execution, for sandboxes and for production alike.

Docker-in-Dockercgroups & netnsSandboxesGrafana LGTMParquet / pyarrowGCPCloud KMS & CMEKAWS

ProofA solo-built capture farm cut a 50-task batch from ~150h to ~4h; a solo Oracle → GCP migration moved 107 tables.

Applied AI

03

Multi-agent systems with human gates, audit logs and a token budget.

LangGraphLangChainLangSmithGeminiMulti-agent systemsRAG / Self-RAGOCR / document AI

ProofA 6-node LangGraph DD pipeline went from 1M+ to 75K tokens per report; a 96-tool workspace runs in production.

Backend & Data

04

Async-first TypeScript and Python services on the right store for each job.

Node.jsTypeScriptPythonPostgreSQLPrismaMongoDBRedisSSE & WebSockets

ProofCaching, async queues and workers took the platform to 1000+ concurrent users at 50ms.

2nd

TON Gurugram Bootcamp

Led the team — 1,500 USDT prize

AIR 5

IDEATHON 2023

Out of 500+ teams nationwide

Top 6

Smart India Hackathon

Team lead

2022 — 2026

B.E. Computer Science

Chandigarh University · CGPA 7.98 / 10

Interlude

Packet run

You're a request crossing the grove: eight quick hops from the edge to the origin. Dodge or blast what's failing, collect healthy requests, keep your replicas alive.

010

8 hops · edge → origin

Route the request.

all levels ↓

enter start · ↑ ↓ / W S move · space / J / X fire · P pause

Level select · 0/8 cleared

Every hop, at a glance

Watch any of them. To play one, clear the hop before it; Overtime opens once the origin falls.

Open to remote roles

Let's build what stays up.

RL environments, applied AI, backend infrastructure. If it has to be correct under load, I'd like to hear about it.

Copied to clipboard