I design and ship large-language-model systems end to end — from custom MoE architectures and RL alignment to the agentic backends and MLOps that put them in production.
| Focus | What that looks like |
|---|---|
| LLM Architecture | Sparse Mixture-of-Experts, Multi-head Latent Attention, SwiGLU, custom training loops in raw PyTorch |
| Alignment & RAG | SFT dataset curation (94k samples), GRPO reinforcement learning, hallucination-resistant retrieval grounding |
| Agentic Systems | Real-time multi-agent backends with FastAPI + Agno, WebSocket streaming, vector context retrieval |
| Media Automation | Script-to-video pipelines: LLM → Stable Diffusion → TTS → FFmpeg rendering |
| MLOps / Infra | Modal serverless deployment, Docker, Weights & Biases, reproducible UV environments |
🎓 AI/ML student researcher · 📦 43 public repos · 🗓️ on GitHub since Aug 2024 · 🧑🏫 led community ML workshops on the Agno framework
|
🔬 Building Pretraining + GRPO reinforcement learning on KT_GPT — my 992.68M-param Sparse MoE LLM — on Modal serverless, funded by a research grant. 🧪 Researching Hallucination-resistant RAG grounding — teaching models to refuse when they have no info, and to parse tool-result math correctly. |
📚 Learning Rust, C, and Java — getting closer to the metal under the Python. 🧑🏫 Teaching Hands-on ML workshops on the Agno agent framework, plus interactive tech events. |
A 992.68M-parameter Sparse Mixture-of-Experts language model, built from scratch in PyTorch. Architected for efficient long-context inference on edge GPUs — and aligned for trustworthy RAG.
| 🧮 Total params | ⚡ Active / token | 🕳️ Sparsity | 🗜️ KV compression |
|---|---|---|---|
| 992.68M | 141.12M | 85.8% | 8.8× |
Why it matters: MoE attacks the params-per-token cost while MLA attacks the KV-cache size — the two things that actually bite you at inference time. Multi-head Latent Attention shrinks the per-token KV cache nearly 9×, making long-context inference viable on edge GPUs, and the model is aligned on a custom SFT set engineered for no-info refusal and tool-result math parsing — so it grounds answers in retrieval instead of hallucinating. Grant credits fund ongoing pretraining and GRPO reinforcement learning.
| Component | Specification |
|---|---|
| Total parameters | 992.68M |
| Active parameters / token | 141.12M (85.8% sparsity) |
| Expert routing | 37 routed experts + 1 shared expert |
| Attention | DeepSeek-style Multi-head Latent Attention (MLA) |
| KV cache / token / layer | 1,408 → 160 values (8.8× reduction) |
| Feed-forward | SwiGLU experts |
| Load balancing | online bias-based |
| Alignment | custom 94k-sample SFT dataset for hallucination-resistant RAG grounding |
| Deployment | Modal serverless ($500 research grant + monthly credits) |
| Roadmap | continued pretraining + GRPO reinforcement learning |
🧠 Architecture data-flow — click to expand
┌──────────────────────────────────────────────┐
token ──▶ embed ──▶ │ Multi-head Latent Attention (MLA) │
│ KV cache: 1,408 ──compress──▶ 160 (8.8×) │
└──────────────────────┬───────────────────────┘
▼
┌──────────────────────────────────────────────┐
│ Sparse MoE router → top-k of 37 experts │
│ + 1 shared expert (SwiGLU FFN) │
│ online bias-based load balancing │
└──────────────────────┬───────────────────────┘
▼
992.68M total params · 141.12M active/token · 85.8% sparse · Modal serverless
Alignment focus — the 94k SFT set is tuned for hallucination-resistant RAG: explicit no-info refusal when context is missing, and correct parsing of tool-result math rather than guessing.
|
Real-time AI coding assistant — VS Code + FastAPI + Agno. WebSocket streaming suggestions, vector-based context retrieval, incremental parsing, and multi-agent routing.
|
🎨 ShowkaseAI-powered portfolio generator. A multi-agent backend (Orchestrator → Planner → Coder) turns your profile into a fully custom React site — with live Sandpack preview and one-click Vercel deploy.
|
|
Script-to-video automation pipeline. LLM script segmentation → Stable Diffusion image prompts → Kokoro TTS → FFmpeg MP4 rendering. Built for MLSA Project Wing 2025.
|
🖼️ KolamvisionComputer vision for traditional Kolam art — analyzing and generating Kolam patterns. Has community forks (related:
|
|
Persistent memory layer for LLM agents — durable state across agent sessions.
|
📄 PDF_RAGRetrieval-augmented generation over PDFs — chunk, embed, and query documents.
|
|
Clinical RAG decision-support system. Hybrid BM25 + Qdrant retrieval, cost-aware dual-model inference via vLLM, Redis semantic caching, and NLI faithfulness checks for hallucination-resistant EHR analysis. Live demo →
|
🎬 Professor Peter 🏆AI video educator — hackathon-winning project. FastAPI orchestrating FFmpeg lip-sync with ElevenLabs TTS narration to generate explainer videos.
|
More projects & teaching repos
| Project | What it does | Stack |
|---|---|---|
| bizman | Business / automation tooling | Python |
| lenai | Applied LLM application | Python |
| ai_interviewer | AI-driven mock interviewer | Python |
| Breast_cancer_detection | ML classification notebook | Jupyter |
| streetview_regionwise_scrapper | Region-wise Street View scraper | Python |
| LLM_Battle_Royale | LLM head-to-head arena | JavaScript |
| pytorch_workshop | Teaching repo — hands-on PyTorch workshop | Python |
| Pytorch_learn | Teaching repo — PyTorch learning material | Python |
🧑🏫 Community: Ran hands-on ML workshops on the Agno agent framework and hosted interactive tech events.
| Domain | Tools |
|---|---|
| Core |     |
| LLM Research |        |
| Agentic & Backend |        |
| Media / Automation |     |
| MLOps / Infra |      |
| Frontend |   |
| Learning |    |
Same stack, as a Python dict
stack = {
"core": ["Python", "PyTorch", "Transformers", "Tokenizers"],
"llm_research": ["MoE", "MLA", "SwiGLU", "LoRA", "GRPO RL", "RAG", "ChromaDB"],
"agentic_backend": ["FastAPI", "Agno", "WebSockets", "AsyncIO", "REST",
"PostgreSQL", "SQLite"],
"media_automation":["FFmpeg", "Stable Diffusion", "ElevenLabs TTS", "Kokoro TTS"],
"mlops_infra": ["Modal", "Docker", "Weights & Biases", "UV", "Git"],
"frontend": ["TypeScript", "React"],
"learning": ["Rust", "C", "Java"],
}| Win | |
|---|---|
| 🗜️ | Implemented & verified DeepSeek-style MLA (8.8× KV cache compression) inside the 992M-param KT_GPT model |
| 💰 | Secured a $500 Modal research grant + monthly credits for pretraining and GRPO RL |
| 🏆 | Hackathon champion with Professor Peter |
| 🧑🏫 | Led community ML workshops on the Agno framework |
I'm open to research collaborations, internships, and conversations about LLM systems. Fastest way to reach me is email.
Generative AI · Agentic Systems · LLM Architecture · MLOps — building from scratch, in the open.