Rushikesh
Patade.
Building game engines and agent systems. Right now that means ludexel — prompt-to-playable across engines — and fabri, the memory and orchestration layer that runs it.
"Developer by nature who wants to have an undeniable god complex someday soon."
Founding Engineer at Soca, building the backend, state layer, and infra for a Go matching pipeline.
Alongside that I build ludexel — a platform that turns a prompt into a playable game, one plugin per engine. GBA works today; UEFN, Roblox, and Minecraft are in the queue. Underneath it runs fabri, the local memory and agent-orchestration framework that keeps the content agents remembering what they did.
Everything else — CV, Rust, systems, older mobile work — is scattered below. Some of it is old. All of it is real.
Where I've worked.
Founding Engineer · Soca
Jun 2026 – Present · 2 mosFull-time · Hybrid- Built and own the core matching engine: candidate generators, an LLM judge with reciprocity and facet-veto policy, and a constrained allocator, with an offline eval harness.
- Own the Postgres task queue: at-most-once claims, stale-job reclaim, and pool budgeting that closed exhaustion and silent match loss.
- Built the state layer (realtime sync bus over React Query + Zustand) and the admin surfaces on it: a curator console and a server-paginated Cost & Usage page.
- Built an LLM cost ledger enforcing spend (daily kill-switch, per-user caps), shipped Backed (a WebGL collectible-card product), and closed four concurrency races while owning releases.
GoReact.jsTanStack QueryZustandPostgresNxLiveKitGitHub ActionsSenior Software Engineer · VectorShift
Jan 2025 – May 2026 · 1 yr 5 mosFull-time · YC S23 · Remote · India- Senior Software EngineerJan 2026 – May 2026 · 5 mos
- Software EngineerJan 2025 – Dec 2025 · 1 yr
- Founded the frontend citation system for an AI search product: cross-format highlighting across PDF/DOCX/Excel, message-level indices, and coordinate-accurate multi-page overlays.
- Primary owner of the document-viewer surface: PDF/DOCX/XLSX/PPTX rendering plus a server-side PPTX-to-PDF pipeline.
- Led the evaluations product: revamped the framework, built a scenario × test-case playground, founded the Evaluators module, and led Workbooks. Promoted to Senior within the year.
- Founded the Rust evaluator subsystem in a ~540k-line engine (axum / tokio), shipped a petgraph hot-path optimization, and parallelized config preprocessing to cut deploy time 25%.
React.jsRedux ThunkRustAI EvaluationsRAGMember of Technical Staff · Sage (formerly Fyle)
Jun 2023 – Dec 2024 · 1 yr 1 moFull-time · Remote- Member of Technical StaffJul 2024 – Dec 2024 · 6 mos
- Frontend Engineer InternJun 2023 – Dec 2023 · 7 mos
- Shipped 2 customer-facing features and 2 UX improvements in 4 months; owned frontend deployment releases.
- Developed E2E tests for a legacy page and migrated it to Angular with zero production issues.
- As an intern, shipped production fixes within days and a feature productizing recurring support requests, with strong usage in 2 days.
AngularTypeScriptTestingTechnical Lead · Black Cherie Media
Feb 2024 – May 2024 · 4 mosFull-time · Hybrid · Pune District, Maharashtra, IndiaTechnical Project Leadership
Featured projects.
- 01
ludexel
GBA engine live · agent-driven pipelinePrompt to playable Game Boy Advance ROM.
Type a prompt, get a real .gba ROM. ~24k lines of Python generators driving a hand-written C engine, compiled through devkitARM. Backed by ~800 automated tests, including headless regression runs of the compiled ROM against a real mGBA build, plus a five-scene playable demo. React/FastAPI product surface, multi-arch images on tagged releases, self-hosted. UEFN and Roblox exist as engine-seam scaffolds; GBA is the one that ships.
PythonCdevkitARMAgentsvisit ↗ - 02
fabri
Apache-2.0 · live on PyPI · v0.13A self-improving agent engine.
An agent orchestration engine with memory. Every run writes a JSONL trace, the trace is mined for failures and successes, each finding is compressed into a short guideline stored in a vector store, and the next run retrieves the relevant ones into its system prompt, so agents stop repeating mistakes without anyone hand-editing a prompt. Guidelines dedupe by cosine similarity and promote from tactical to strategic after recurring across sessions. Hybrid BM25 + dense RRF retrieval lifted recall@5 from 0.79 to 0.94, locked behind a CI eval gate that fails any merge regressing below baseline. Runs recursive sub-agents with per-model cost accounting that rolls grandchild spend up to the parent, five LLM providers (Gemini, Anthropic, OpenAI, OpenRouter, Bedrock) behind one interface, two memory backends (Qdrant and SQLite FTS5), an eight-state outcome taxonomy built to defeat false-success (a run where every tool call failed can't score as a success), an ask_user human-in-the-loop primitive, OpenTelemetry spans across sub-agents, and a fail-closed sandbox. 1,008 tests, green in 93s.
PythonAgentsQdrantLLMgithub ↗ - 03
Nakshekadam
SIH 2022 · Winner · 10★SIH 2022 winning project.
Smart India Hackathon 2022 winner. Cross-platform Flutter app addressing a national problem statement, my most-starred repo and the first real ship.
FlutterDartgithub ↗ - 04
IoT-Architecture
3★A curated hub of IoT architectural patterns.
Reference architectures and patterns for IoT systems, written in Mermaid so the diagrams live next to the reasoning.
MermaidSystemsgithub ↗ - 05
appwrite_ui_auth
7★Flutter auth template library.
Drop-in Appwrite authentication UI + flow package for Flutter: email, OAuth, session handling wired up so a new app can start from a signed-in state.
FlutterDartAppwritegithub ↗ - 06
ssd-implementation
VGG16D + SSD, from scratch in PyTorch.
Single-Shot Detector built ground-up on a VGG16D backbone. An exercise in reading the paper and typing every layer, not calling torchvision.
PyTorchComputer Visiongithub ↗
The wider archive.
Six years of side projects and experiments. Not everything is polished — most of it exists because I wanted to know how it worked.
AI, Agents & LLM tooling
1Games & engines
6ML & Computer Vision
2Systems & low-level
5Mobile
6Six years, one line.
- 2020
First commits.
Started shipping code publicly on GitHub. Small tools, game clones, learning by breaking things.
- 2022
Nakshekadam wins Smart India Hackathon.
Cross-platform Flutter app takes SIH 2022, the moment building for real users clicked.
- 2024
The LLM / agent shift.
Full turn toward agent systems, orchestration, and the tooling around large models.
- 2026
ludexel and fabri.
Prompt in, real Game Boy Advance ROM out, built on a self-improving agent engine I wrote. Other engines are scaffolded; GBA is the one that ships.
Polyglot, by habit.
I pick the language the problem wants. These are the ones I've shipped something in.
- Python
- TypeScript
- JavaScript
- Go
- Rust
- C
- C++
- Dart
- Assembly
- Lua
- Shell
- React
- Next.js
- Angular
- Node.js
- Express.js
- Django
- Astro
- Flutter
- PyTorch
- Vite
- PostgreSQL
- MySQL
- Oracle SQL
- MongoDB
- Firestore
- Qdrant
- Pinecone
- MinIO
- GBA / devkitARM
- UEFN / Verse
- Docker
- Kubernetes
- GitHub Actions
- Cloudflare
- Multi-agent AI delivery (Claude Code + Codex)
- Agent orchestration
- RAG
- LLM tooling (OpenAI, Bedrock, Azure)
- Computer vision
- Playwright / Jest testing
Education & awards.
MIT World Peace University
Bachelor of Technology, Computer Science and Engineering
July 2024 · Pune, Maharashtra, India
CGPA: 9.58 / 10.00
Coursework: OS, CN, OOP, RDBMS, Project Management, DSA, AI, ML
Technical & IT Head · Innovation Hub
Jul 2022 – Apr 2023 · Pune, Maharashtra, India
- Led and managed 15 diverse projects while overseeing a team of 100+, implementing technical guidance and process enhancements that boosted efficiency by 100%.
- Mentored and empowered 20+ entry-level team members, fostering a culture of continuous learning and growth.
- Organized and led 10+ activities focused on DSA and competitive-programming skills.
- 2022Smart India Hackathon 2022Winner, All India Rank 1 in a problem statement set by the Ministry of Education.
- 2024Datathon @ Cummins CollegeWinner, ranked 1st out of 200 teams.
- 2021IBM ReImagineFinalist, top 150 among 38,000+ international teams.
- 2024AWS x Riot Data HackathonManaged a 500GB dataset and processed it to vector embeddings.
- 2024HackMITWPU '24 · Video Analytics by Maha-MetroRanked in the top 5 at a state-level hackathon.
- 2021Resideo Smart HomesFinalist, top 26 out of 4,000+ international teams.
Tell me what
you're building.
If you're working on agent systems, game tooling, or something quietly ambitious — I'd like to hear about it.