I like problems where the model's answer isn't enough by itself. So I build what goes around it: agents that run and check the tools they write before keeping them, evaluations that don't grade themselves, and systems that record why each decision was made.
Lately that's meant agent tooling over the Model Context Protocol, causal ML where correlation would mislead, and small models that run on hardware you already own.
llm-agents mcp evals causal-inference small-models ai-observability on-device-ml
Also: PaddleOCR-VL in MLX for local document parsing on Apple Silicon, AEGIS, per-patient causal estimates for Type 1 diabetes (47.5% lower error than OLS), and CampaignIQ, doubly robust causal estimates explained by a local Mistral-7B.
now B.Tech in AI & ML · Saveetha Engineering College, graduating 2027
2026 AI engineering · Matrix Maven
open-source context + policy layer for coding agents, a 15-tool MCP
interface, and models quantized to MLX for on-device inference
2025–2026 Tech Society · Saveetha Engineering College
built the eval pipeline that graded 1,500+ hackathon submissions;
ran five student tech communities, 400 → 600+ active members
2024–2026 Research · GNS-AI
tested an LLM as a prior-authorization reviewer on 166 cases:
63% agreement, and its stated confidence said nothing about accuracy
I've talked about context management for small on-prem models at the NASSCOM Research Conference, and taught SLM fine-tuning to a room of 70+ students.
Models PyTorch · Transformers · PEFT / LoRA · MLX · ONNX · Ollama · DoWhy
Agents Model Context Protocol · LangGraph · tool calling · Firecracker sandboxes
Systems Python · TypeScript · Go · FastAPI · gRPC · Postgres · Docker · Kubernetes · GCP
If you're working on agent reliability, evals, or getting models to run somewhere they weren't meant to, I'd like to hear about it.