// vineethavc25@gmail.com · Los Angeles, CA

Vineetha
Vallish Kumar

|

ML engineer building AI systems that work in production, in regulated environments, and under real constraints — from biosignal reconstruction in MRI scanners to edge model compression and on-device inference benchmarking.

0.00 ECG Reconstruction Correlation
0.00x INT4 Model Size Cut, Measured on a Raspberry Pi 4
0% Faster Training Convergence
0M+ Samples in Stress Detection Model

01 — About

Who I am

I'm an M.S. student at USC (ECE, Machine Learning & Data Science, Dec 2026) building at the intersection of production ML, biosignal research, and embedded systems. My work spans clinical AI collaborations with CHLA, Verizon-funded wireless research, and enterprise LLM infrastructure at SHash AI, and my recent open-source work is on honest, reproducible edge-AI benchmarking.

Currently a Data Scientist (co-op) on the R&D data science team at Alcon, where I lead a GenAI tool on Amazon Bedrock that compares a new project's FMEA with its predecessor's, in an FDA-regulated medical-device environment. This summer I was a data science intern on supply chain analytics there: I built a GenAI data assistant that answers leadership's ad hoc data questions in seconds instead of up to a day, and forecasting and inventory optimization work that surfaced $3.7M in potential savings.

Before grad school, I spent a year and a half at Tata Consultancy Services writing production embedded C for a General Motors Cadillac program on a Snapdragon SoC running QNX, under AUTOSAR and ISO 26262, and building a defect-diagnosis service that ran in production. That hardware-first background is what makes my ML work different.

PyTorch / TensorFlow LLM Fine-tuning RAG / Agentic AI AWS SageMaker Federated Learning Edge AI / QAT Biosignal Processing C / C++ Firmware Linear Programming Monte Carlo Sim XGBoost / Prophet Tableau / Power BI MCP / Amazon Bedrock ONNX Runtime Bare-metal ARM / STM32 Kernel Optimization

02 — Research

What I've built in the lab

Biosignal AI · USC Khan Lab / CHLA · Advanced Intelligent Systems (in prep)

ECG Reconstruction During MRI Scanning

Built a BiLSTM + transfer learning pipeline to reconstruct 12-lead ECG from a wearable patch device during low-field MRI — a hard problem because MRI introduces severe electromagnetic noise that corrupts cardiac signals. Cross-functional collaboration with electrical and RF engineers at CHLA. Manuscript in preparation.

r = 0.94 at rest r = 0.89 active MRI +20% over baseline
Multimodal AI · USC Khan Lab

Stress Detection from Wearable Biosignals

ResNet–BiLSTM model with attention for stress-level inference from simultaneous PPG, GSR, and HRV signals. Trained on 13M+ samples.

~0.87 correlation ~85% accuracy 13M+ samples
Wireless AI · USC WiDeS Lab · Verizon-Funded

PMNet — Radio Pathloss Prediction

Improving training efficiency and cross-city generalization of PMNet, a published deep learning model that replaces slow ray-tracing simulations for wireless network planning. Proposed the elevation input behind PMNet 2.5 and trained the 33M-parameter CNN to SSIM above 0.90 on 8x A40 GPUs (PyTorch DDP, SLURM). On-the-fly augmentation cut epochs to convergence 44% (110 to 62) and training time ~50% (25h to 12h) at equivalent accuracy.

44% fewer epochs ~50% less training time SSIM > 0.90 SSIM ~0.91 Multi-node DDP

03 — Projects

Things I've shipped

🏠

Hearth — A Voice Diary That Remembers What You Promised

An Alexa+ add-on built as a self-hosted MCP server (Streamable HTTP, six tools) for people who live alone and lose track of what they said they would do. It acknowledges your day without giving advice, remembers your commitments, and tells you what matters most and why. Amazon Bedrock extracts commitments and tone; dates are resolved in code, not by the model, and prioritisation is deterministic. Tested end to end against an Alexa+ simulator; registering it with real Alexa+ is not done yet.

Python MCP Amazon Bedrock SQLite Alexa+

SynapNet-Edge — Edge Model Compression, Measured Honestly

An 8.7M-parameter hybrid SSM-attention-memory model with component-aware quantization and a 3.3K-parameter learned memory-eviction policy (71% of critical context retained under 90% memory pressure, versus 0% for FIFO/LRU). On a Raspberry Pi 4, bit-packed size fell from 16.6 MB to 10.2 MB (INT8) and 8.4 MB (INT4), within about 2 points of FP16 on retrieval up to 4K tokens, but latency rose 2–18% because the weights were dequantized on every pass. Real INT8 ONNX kernels then ran 1.54x faster single-threaded on an Apple M3, and the KV-cache eviction policy was benchmarked against H2O, SnapKV and PyramidKV. A matched-control study (11 variants, 3 seeds, paired bootstrap CIs) showed an earlier QAT gain was a fine-tuning effect, not a quantization effect.

PyTorch QAT SSM / Mamba ONNX Runtime Raspberry Pi 4 Edge AI
👁️

VisionFlow — On-Device Vision-Language Pipeline

An edge-first quantized VLM pipeline (SmolVLM) that detects your hardware, picks a model size that fits, and constrains decoding so the JSON output is valid by construction. On 150 real DocVQA documents, grammar-constrained decoding took usable output from 2% to 79% and was faster. INT4 runs in 1.85 GB of VRAM at 47.7 tok/s on an L40S, and the ONNX vision encoder runs at 11.90 ms p50 on TensorRT, 109.8x faster than CPU and 1.62x faster than the CUDA execution provider. Every benchmark row runs in a fresh subprocess and records the provider the live session actually used, which caught two silent CPU fallbacks that advertised themselves as TensorRT before the real number was earned. These are encoder-only numbers; the decoder is still PyTorch.

SmolVLM bitsandbytes ONNX Runtime Constrained decoding Python

flashattn-amx — FlashAttention in C on Apple's AMX

The FlashAttention forward pass in about 130 lines of C, running its matmuls on Apple's AMX matrix coprocessor through Accelerate SGEMM, with vForce for the exponentials and GCD for parallelism. Online softmax avoids ever storing the full score matrix, causal masking skips whole blocks, and the 256x256 tiles are sized to fit the M3's 128 KiB L1 cache. On an M3 it is 1.2 to 2.2x faster than PyTorch's own flash-attention CPU kernel on every tested shape (peak 634 GFLOP/s) and 1.19 to 1.28x faster end to end in a GPT-2-small-sized model, with a max logit difference of 4e-6. A100 and M3 GPU tables are included to show where it does not win. Published on PyPI as flashattn-amx.

C Apple AMX Accelerate FlashAttention PyTorch
🔌

renode-fw-tests — Bare-Metal STM32 Firmware, Tested in CI

Register-level C firmware for an STM32F407 (Cortex-M4) with no vendor HAL, no CMSIS and no C library: a hand-written vector table and linker script, RCC clock gating, GPIO alternate-function (AF7) pin muxing and a polled USART2 driver at 115200 baud. A Robot Framework suite runs in GitHub Actions on every push: it boots the firmware in the Renode emulator, checks the UART output, and reads the GPIO and USART registers back to confirm the hardware was configured as intended.

C ARM Cortex-M4 STM32 Renode Robot Framework GitHub Actions
🎙️

MnemoSense — A Local Memory Assistant You Can Talk To

🏆 2nd place, Hardware Track · LA Tech Week AI/ML Buildathon 2025

A team project: an always-on assistant that logs audio around the clock but stores only summaries, and answers spoken questions about your day through a voice UI (the answer is spoken and shown). It also captions video, from a URL or a live camera every 10 seconds, into the same retrieval index. Runs entirely on a free local LLM through Ollama, with Whisper for speech-to-text.

Python Ollama Whisper RAG Voice UI

Edge Deploy Agent — An Agent That Tunes Models on a Real Raspberry Pi

Nebius x NVIDIA Global AI Hackathon · Coding & Agentic Engineering track

A Nemotron agent (LangGraph) that finds the fastest way to run a speech model (wav2vec2) on a Raspberry Pi 4 within latency, memory and accuracy budgets, by benchmarking real configurations on the device. It picked dynamic INT8 with 4 threads and extended graph optimizations: 1932 ms p95, 26% faster and 39% less RAM than the same-settings FP32 baseline, for +0.23 points of WER. The winner is exported to Triton Inference Server on Kubernetes, provisioned with Terraform. Because the agent reads web docs, it sits behind layered prompt-injection defenses (regex sanitizer, NeMo Guardrails input and output rails, and a deterministic allowlist the LLM cannot widen); in 60 attack attempts, none reached execution on the Pi. Every trial is tracked in MLflow and traced with OpenTelemetry.

Nemotron LangGraph NeMo Guardrails Triton Kubernetes Terraform Raspberry Pi 4
📱

On-Device AI on Snapdragon NPUs

Two models taken from PyTorch to Qualcomm's Hexagon NPU through Qualcomm AI Hub, with every number linked to its AI Hub job. Qwen3-0.6B (W4A16) compiled into QNN context binaries for a Galaxy S25 (Snapdragon 8 Elite): 97 tok/s decode at 1K context and 40 tok/s at 4K, 29 ms to first token, and all 27.6K ops on the NPU. These are NPU graph latencies, so they exclude runtime overhead like sampling and tokenization. And INT8 MobileNetV2 on a Galaxy S24 (Snapdragon 8 Gen 3): 0.37 to 0.28 ms with 105 of 105 layers on the NPU, at a 0.18 point top-1 cost measured only after verifying FP32 parity with PyTorch.

Qualcomm AI Hub Hexagon NPU INT8 QNN On-device LLM
📦

PromptOps — Git for LLM Prompts

Open-source CLI bringing Git-style version control to LLM prompts. Semantic diffing via sentence-transformers, Thompson Sampling A/B testing, LLM-as-judge regression eval, an MCP server, and a GitHub Actions regression gate with three triggers (paired t-test, per-case collapse, per-dimension floor) calibrated on 90 run-to-run comparisons, because prompt regressions break a few inputs badly rather than all of them slightly. Install from the v0.3.0 GitHub release (package promptops-git).

Python Thompson Sampling MCP GitHub Actions

AgentAudit — Multi-Agent Evaluation

Open-source Python library that instruments multi-agent AI pipelines and evaluates the workflow, not just the final output. Scores inter-agent handoff quality via LLM-as-judge, attributes failures to the earliest responsible agent (bisecting instead of scanning), and detects degenerate retry loops with cost guardrails. The judge is validated on a 60-trace benchmark with planted failures: Claude Sonnet reaches 86.7% attribution accuracy (Cohen kappa 0.92), and a local llama3.1-8B fails outright, an honest negative result. Published on PyPI as agentaudit-eval.

LangGraph CrewAI LLM-as-judge CI gates Python

04 — Learnings

What the benchmarks taught me

Edge inference · SynapNet-Edge

Smaller is not automatically faster

INT8 and INT4 models were up to 2x smaller on disk but 2–18% slower on a Raspberry Pi 4 with RAM unchanged. The weights are dequantized to float on every pass, the Cortex-A72 has no INT8 dot-product path, and the runtime and activations dwarf 8–17 MB of weights. Switching to real ONNX Runtime INT8 kernels fixed it: 1.54x faster single-threaded on an Apple M3. On a Hexagon NPU, INT8 MobileNetV2 went from 0.37 to 0.28 ms with all 105 layers on the NPU.

Experiment design · SynapNet-Edge

Control for the fine-tune, not just the quantizer

A quantization-aware-training gain over FP16 vanished once FP16 got the same 200 fine-tuning steps; it was slightly ahead at 4K context. The fix was a matched control, 11 variants, 3 seeds and paired bootstrap confidence intervals instead of single-run comparisons.

Measurement hygiene · SynapNet-Edge · VisionFlow

Audit the benchmark before you trust it

I caught three of my own errors before reporting: a benchmark run on the wrong architecture, a 3 MB positional table missing from the size estimate, and a 48-question eval too noisy to rank variants. In VisionFlow, two tiers measured in one process reported memory figures 0.01% apart because the second read the first one's allocations, so every configuration now runs in a fresh subprocess.

Product scope · Hearth

Write the limits next to the results

Each of my recent READMEs has an honest-limits section: Hearth is single-user and untested against real Alexa+, VisionFlow reports 21 fields over three images, and a TensorRT row that fell back to CPU is labelled as such. Claims that survive a skeptical reader are the only ones worth putting on a resume.

05 — Experience

Where I've worked

SEP 2026 – PRESENT

Data Scientist, R&D Data Science (Co-op)

Alcon · Johns Creek, GA (Remote)

  • Lead a GenAI tool for medical-device R&D on FDA-regulated data, scoped with SMEs, that compares a new project's FMEA with its predecessor's so engineers triage only what changed.
  • Build the APIS/IQ-FMEA XML to JSON parser preserving structure-tree, failure-net and risk relationships; evaluate Amazon Bedrock Knowledge Bases on grounded answers and source traceability.

MAY 2026 – AUG 2026

Data Science Intern

Alcon · Supply Chain NA · Fort Worth, TX

  • Built a GenAI data assistant (ForeSight Pro agent, Amazon Bedrock, DynamoDB) that answers leadership's ad hoc data questions in plain English in seconds instead of up to a day of analyst wait; it shows the source data per answer, asks clarifying questions first, and guards against out-of-scope pulls. Handed off to the data science team.
  • Forecasting and inventory optimization (K-Means, XGBoost, Prophet) on AWS SageMaker and MLflow surfaced $3.7M in potential savings; moved capacity planning to p50/p95 reporting.
  • Worked cross-functionally with Transportation, Customer Service, Data Science and S&OP; insights went to leadership for Q3 consideration. Scoped a shipment-level anomaly early-warning system.

NOV 2025 – SEP 2026

LLM Engineer Intern

SHash AI · Remote

  • Architected an enterprise LLM system with privacy-preserving GenAI and federated learning (Flower); patent application filed.
  • Deployed a 3-client Flower setup (TinyLlama + LoRA) across two machines and a MacBook Air, with client-side backpropagation independent of server updates so gradients never propagate to the server.
  • Designed LLM evaluation harnesses for response quality, factual grounding, and safety across multi-turn dialogue; began extending to Mamba-based architectures for long-context efficiency.

JUN 2024 – DEC 2024

Machine Learning Engineer Intern

Analogica Software · Bangalore, India

  • Fine-tuned BERT for healthcare NLP; improved F1 by 15%, cut inference latency ~25% via distillation.
  • Built computer-vision QC pipeline (~92% accuracy); productionized 5+ models in ONNX/TorchScript.

OCT 2022 – MAY 2024

Embedded Software Developer

Tata Consultancy Services (client: General Motors) · Bangalore, India

  • Owned three production C components (Functional Safety, Vehicle Data, Display Brightness) on a Snapdragon SoC running QNX for a Cadillac program, shipping across 2 releases under AUTOSAR and ISO 26262.
  • Proposed and built a defect-diagnosis service (React, FastAPI, XGBoost tracked in MLflow, PostgreSQL, Kubernetes), run in production for 4 months before handoff; auto-triaged ~80% of incoming Brightness defects and cut analysis time ~50%.
  • Built the Vehicle Data unit-test suite from scratch; resolved 100+ Jira defects with root-cause analysis. Reproduced a defect on the bench from CAN signals in Vector CANoe with no ECU access; the fix cut latency 50% (200 ms to under 100 ms).

06 — Contact

Let's build something together

Open to full-time roles in R&D, medical AI, and production ML systems.
Graduating December 2026, available January 2027. Based in Los Angeles, CA.