FastAPI · React · TypeScript · ChromaDB · Groq · NVIDIA NIM · Qwen (fine-tuned)

• Fine-tuned Qwen-2.5-Coder-7B on 50+ custom Manim animation examples built from scratch
• Two-stage pipeline: NVIDIA Nemotron → JSON storyboard → Qwen → Manim Python → MP4
• Production RAG: arXiv → chunking → ChromaDB → citation-grounded LLM chat
• Solved 75s+ async timeout via FastAPI BackgroundTasks + asyncio executor isolation
• 95%+ JSON parsing success via automatic repair logic

Python · LLMs · Multi-Agent · STT/TTS · Runtime Tool Synthesis

• Agent synthesises and executes tools at runtime — 40+ dynamic capabilities, zero pre-written tool list
• Full hands-free: wake-word detection → STT → multi-agent LLM orchestration → TTS
• Multi-step task planning and intent understanding at system level
• Voice-controlled automation across a unified tool-calling interface

FastAPI · Next.js · Go · MCP Server · OAuth · LLMs

• Multimodal: text, voice, documents → LLM-powered Sharpe Ratio, VaR, AI rebalancing
• Go-based MCP server for OAuth cross-account financial data aggregation
• Built backend infra and LLM context layer together end-to-end

Python · PyTorch SDPA · LLMs · Muon Optimizer

• Ported Karpathy's H100-only agent to free T4 GPUs via PyTorch SDPA
• 22 autonomous overnight experiments — multi-key API rotation + resume-safe checkpointing
• Agent loop: LLM proposes diffs → trains → parses val bpb → commits or reverts
• Open-sourced for reproducible zero-cost ML research

PyTorch · HuggingFace PEFT · TRL · Kaggle T4

Benchmarked 6 PEFT methods (LoRA, QLoRA, DoRA, VeRA, AdaLoRA, IA3) on SmolLM2-1.7B — same model, same data, same hardware. Built to answer one question: which method should you actually use?

Key findings: DoRA → best final loss · IA3 → LoRA quality at 10× fewer params · QLoRA → honest default for free GPU users · Phase 2: safety alignment degradation testing in progress

AdaptSRNet: Enhancing Image Steganalysis via Adaptive Filter-Attention Fusion

Compact deep architecture: learnable SRM filter banks + SE/CBAM multi-scale attention fusion for high-precision binary steganalysis. Evaluated on WOW stego images at 0.2 bpp with full ablation studies.

Deep Learning Attention Mechanisms Digital Forensics Steganalysis

Cold-Start Learning Under Domain Uncertainty: A Strategy Selection Perspective

Reframed cold-start forecasting as a strategy selection problem. Used normalised Wasserstein distance as domain divergence criterion — no target labels needed. Showed scratch training outperforms transfer learning under measurable domain mismatch.

Cold-Start Learning Transfer Learning Domain Adaptation