4×3modalities benchmarked across three chip vendors
1intern mentored, plus knowledge-sharing sessions for teammates
How to read this site
Follow one task through an agent.
The agents I build handle a task in the same order: an orchestrator takes it, picks tools, checks itself against guardrails, acts, then reflects on the result. This page runs in that order. Scroll, and your visit is the task.
01Orchestrator
Your task lands with an engineer who builds agentic systems for production.
I'm Shriyans, an AI engineer in Bengaluru with two-plus years of building GenAI systems that ship. Today I'm at Torinit, building AI features that help estimators in the HVAC industry. Before that I led a platform refactor that split an agentic AI platform into independent services. My work sits where agents meet the real world: orchestration, tool calling, model routing, guardrails and evaluation, so an agent behaves in production the way it did in the demo.
2026.07 → now
Torinit · AI product
Software Engineer – AI · features for HVAC estimators, remote
2026.05 → 06
Torinit · internal POC
End-to-end LLM pipeline, model and framework evaluation
2024.06 → 2026.04
NStarX
Data Scientist / AI Engineer / ML Engineer · where my scope grew fastest
2024.01 → 05
Futurense
AI Data Engineer · analytics and ML
2024.08
Jain University
B.Tech Computer Science (AI & ML) · CGPA 8.0
NStarX, where the curve bent upward
I joined before I'd even graduated and left leading platform design. Same two years, a very different scope. Tap a point.
02Tools
The toolbox the orchestrator can call on.
Grouped by the job each tool does inside an agentic system.
Agent tooling and MCPtools agents can call safely, Claude Code workflows
LLM platform architecturegateways, model routing, streaming, rate limits
Forward deployed engineeringsitting with a team and shipping on their problem
Inference serving and benchmarkingNIM, vLLM, latency on real hardware
FLAGcan do, not my core
Frontend glue in ReactI ship features across React and FastAPI when the work needs it
Data analysis and BI dashboardsPandas, SQL, Power BI
Model fine-tuningdone it with NeMo and DAPT; held back by data and GPU access, not interest
Retrieval pipelinesbuilt plenty before; now as memory for agents
04Execute
Platform layers: one agentic platform, split into parts anyone can plug in.
NStarX's GenAI Lab ran several agentic AI use cases on a single monolith. I led the design and build of four shared layers that pulled it apart into independently deployable Docker services, provisioned per use case with Helm on Kubernetes. The highlighted boxes are mine.
Result: use cases stopped shipping their own copies of the same plumbing, deployments got simpler, and teams could switch models or backends without touching their code.
Agent systems
NStarX
Multi-agent event planner
Peer-to-peer agents on NVIDIA NAT for idea generation, theme refinement and participant retrieval, with read-only SQL tools, structured outputs and calendar file generation. Llama 3.1 8B on NIM.
NVIDIA NATMulti-agentTool callingNIM
Product engineering
Torinit
AI features for HVAC estimators
Started on an internal LLM proof of concept, evaluating models and frameworks and wiring LLM capabilities into an existing system. Now on the core team of an AI product for the HVAC industry, adding new features that help estimators do their job faster, across React and FastAPI.
OpenAIAnthropicFastAPIReact
Inference and models
NStarX
AI PC benchmarking
GenAI workloads across text, audio, image and video on Intel, AMD and Qualcomm machines. System metrics (CPU/GPU, power, memory, temperature) captured next to model metrics (TTFT, ITL, end-to-end latency).
BenchmarkingInferenceEdge hardware
NStarX
Fine-tuning as a service
A domain-specific fine-tuning service on the NeMo framework. I worked on data curation, domain-adaptive pre-training and custom tokenizers.
NVIDIA NeMoPyTorchDAPT
Earlier
NStarX
Retrieval work
A policy chatbot for an enterprise client and the move of an in-house RAG studio onto NVIDIA NIM. That groundwork is what agent memory layers are built on now.
MilvusNIMLangChain
05Reflect
Figure it out, then make it hold.
My default is a figure-it-out mindset. New stack, thin docs, vague brief: I break it down until it has a shape, then build. I don't start with a model. I start with what goes in, what must come out, and everything that can go wrong in between.
RequirementsWhat problem, for whom, and what counts as done.
Inputs and outputsPin the contract: shapes, limits, latency budget.
SystemDesign the pieces so each one can be swapped.
Failure pathsRetries, fallbacks, timeouts, every edge case I can think of.
MeasureNumbers before opinions, then iterate.
what I watchTTFTITLend-to-end latencytask successtool-call errorsguardrail hit rate
mentored an internknowledge-sharing sessions for teammatescoached kids' football
always exploringSystem One models, like Jev by TypeSafe AIgame developmentAI for engine diagnostics and repair
Always learning. When a new idea shows up, like non-autoregressive System One models, I read the paper, try it, and work out where it would actually fit.
I try to be a considerate peer: clear updates, credit where it's due, and no surprises in review. Async written updates, calls or pairing all work for me. I'm also actively looking for a mentor, so if you've built platforms like these for longer than I have, I'd like to learn from you.
Off the clock · floor plan, not to scale
Where the hours go.
The rest of me, one spot at a time. Pick one.
06 · Response
Ask, then reach the real me.
The terminal answers questions about my work. It's an AI stand-in, so for anything that matters, email me.
twin@shriyans:~local notes
Roles
Not looking for a new role right now
Happy to stay in touch for later. My focus: AI Engineer and Forward Deployed work on agentic systems. Email is the best way to reach me.
Projects
Make your agent hold up in production
Agent workflows, tool design, model routing with fallbacks, guardrails and evals. Scoped around your stack.
Early AI engineering
The layer between your model and your product
I've split a monolith into services before. Let's talk about getting yours right the first time.
Notes and mentorship
Compare notes, or mentor me
Agent orchestration, MCP tooling, Claude Code, model serving. I'm looking for a mentor who has built this at scale.
Football
Training sessions
I play every week and run training sessions. Tell me your group size and level.