Bhargava Sai Deepak Mathukumalli — Full-stack Applied-AI & Core Systems Engineer

I ship full-stack AI products in production — and understand them down to the metal.

Software engineer (SDE-2) working end to end: LLM agents and their evals, Next.js dashboards, Go/gRPC services — and, on my own time, the layer underneath it all, rebuilt from first principles: inference servers, trained models, CUDA kernels.

See the work GitHub

▹~2.5 yrs full-time · SDE-2 ▹production full-stack AI · Next.js · Go · FastAPI ▹LLM agents · evals · MCP ▹from CUDA kernels to served models

In production — professional work

Shipped, full-stack, and used by customers.

Two systems I built end to end at Amnic — the layer where real users, real data and real deadlines live. (Company work: described at a high level, no code shared.)

In production Amnic · full-stack

Prism by Amnic

Built end to end — frontend, backend, services & infra

A pre-deploy cost-intelligence SaaS: it estimates the cloud-cost impact of infrastructure-as-code changes before they ship, and exposes it to AI clients through a Model Context Protocol server so Claude or Cursor can query scans in natural language.

Next.js · TypeScript FastAPI · GraphQL Go · gRPC MCP server Celery · Postgres Fly.io

Multi-service monorepo · React dashboard → Python API → Go scanner engine.

In production Amnic · applied AI + evals

Amnic AI — agent & eval harness

LLM agent plus the framework that keeps it honest

A production LLM agent that answers cloud-cost questions over customer data via tool calls — and the LLM-as-judge evaluation framework I built to guard it: acceptance-criteria and golden-reference scoring, plus regression and investigation modes that replay historical tool-call traces.

LLM agent · tool-use LLM-as-judge evals regression + golden refs Anthropic · OpenAI · Bedrock

Model development and evaluation — the full loop.

Built from scratch — the depth

And the same stack, rebuilt from the metal up.

Shipping on top of the ML stack is the day job. On my own time I rebuild it: three personal projects that fit together — the kernels power the server, the server serves the model, and I wrote all three, from the GPU up.

L2 · The Modelabstraction

scribe

A ~30M-parameter GPT trained from scratch on TinyStories — kept GPT-2-compatible on purpose, so my own inference server can load it and stream stories.

GitHub ↗ Hugging Face ↗

val loss 1.74
~3.4h on MPS
TinyStories · GPT-2 compat

L1 · The Serversystems

ember

A mini-vLLM inference server: paged KV cache, continuous batching and an OpenAI-compatible streaming API — with GPT-2 itself rebuilt from scratch inside it.

GitHub ↗

paged KV cache
continuous batching
13 paging tests

L0 · The Metalhardware

cuda-kernels

Hand-written CUDA, escalating in difficulty: vector-add → tiled GEMM → fused softmax → FlashAttention with online softmax — benchmarked against cuBLAS and PyTorch.

private · available on request

FlashAttention
online softmax
vs cuBLAS / torch

Selected work — breadth

Beyond the ML stack.

The same instinct, pointed at distributed systems, developer tooling, generative models and data engineering.

Distributed Systems · Go

quorum

A distributed key-value store with a Raft consensus layer written by hand — leader election, log replication and snapshots — verified under Go's race detector.

GitHub ↗ LSM engine

Developer Tooling · Python

athena-cost-guard

A pre-flight cost estimator for AWS Athena with a @cost_guard budget decorator — sqlglot + Glue + S3. Published and installable from PyPI.

PyPI ↗ GitHub ↗ shipped

Generative ML · PyTorch

text-diffusion

A tiny text-conditioned diffusion model — a mini Stable Diffusion — built from scratch and trained on Apple-silicon MPS, with a generated sample grid in the README.

GitHub ↗ Hugging Face ↗

Numerical Methods · NumPy

kalman-timeseries

An adaptive, robust Kalman filter in pure NumPy, written up as an academic-style LaTeX paper. The self-tuning variant cut error ~51% over the baseline.

GitHub ↗ + paper

Data Engineering · PySpark

clickstream-sessionizer

A medallion lakehouse (bronze → silver → gold) with gap-based sessionization, skew-salting, broadcast joins and Structured Streaming — validated on ~400k events.

private · available on request

Everything here is open, personal work — read the code, the tests and the commit history.

All repositories ↗

About

Two and a half years shipping. The rest of the time, building the hard parts.

I'm a Software Engineer (SDE-2) with roughly 2.5 years of full-time experience, currently at Amnic, where I work across the whole stack — a Next.js dashboard, Python and Go services, an LLM agent and the eval framework that keeps it correct.

The through-line is that I don't like treating layers as magic. So on my own time I rebuild the ones I depend on — inference servers, consensus protocols, GPU kernels, trained models — from first principles, because reimplementing something is the only way I fully trust that I understand it. Those projects are open source: real code, real training runs, real benchmarks.

Put together: I can ship a full-stack AI product and also reason about what happens inside it, down to the kernel. If you're hiring for applied AI, ML, or backend that values both — let's talk.

roleSDE-2

experience~2.5 yrs FT

focusapplied AI · full-stack

alsoML · systems · data

languagesPython · Go · TS · CUDA

publishedPyPI · Hugging Face

github@mbsdeepak