// vineethavc25@gmail.com · Los Angeles, CA
|
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.
01 — About
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.
02 — Research
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.
ResNet–BiLSTM model with attention for stress-level inference from simultaneous PPG, GSR, and HRV signals. Trained on 13M+ samples.
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.
03 — Projects
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.
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.
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.
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.
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.
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.
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.
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.
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).
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.
04 — Learnings
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.
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.
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.
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
SEP 2026 – PRESENT
Alcon · Johns Creek, GA (Remote)
MAY 2026 – AUG 2026
Alcon · Supply Chain NA · Fort Worth, TX
NOV 2025 – SEP 2026
SHash AI · Remote
JUN 2024 – DEC 2024
Analogica Software · Bangalore, India
OCT 2022 – MAY 2024
Tata Consultancy Services (client: General Motors) · Bangalore, India
06 — Contact
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.