Harsh Shroff HarshShroff

Harsh Shroff

I build agents and edge models, then try to break them. Most AI demos stop at "it works on my prompt." I keep going until there's a number, and I put the number where you can check it.

Portfolio · LinkedIn · Email · Open to AI/ML engineering roles, especially forward-deployed and applied-agent work.

Receipts

Claim	Number	Check it
vLLM beats a naive Hugging Face loop	1.6x to 2.2x throughput, one T4, Qwen2.5-3B	results
Agent finds cell-level drug effects	0.92 segmentation F1, 98% outlier precision on BBBC021	Bio-Oracle
A percentile bug in a benchmarking tool	wrong in 286 of 4,444 cases on main, 0 on the fix	my review of guidellm #1194
A scoring engine you can audit	15 factors, 118+ tests	Silicon Oracle

Featured Projects

MARS

Research reports that send weak drafts back

Eleven LangGraph agents turn a topic into a citation-grounded report. A QC agent scores every draft and loops weak ones back through a planner instead of shipping them. Runs as a Streamlit app and as an MCP server other clients can call.

Stack: Python · LangGraph · Pydantic · MCP · Gemini · Streamlit

#agents #evals #mcp

vLLM benchmark

Testing the vendor claim instead of repeating it

vLLM's continuous batching against a naive Hugging Face generate loop, one T4, Qwen2.5-3B. Harness, fixed prompts, raw results and plots are in the repo so anyone can rerun it.

Stack: Python · vLLM · PyTorch · Colab

#inference #benchmarks

Bio-Oracle

A reasoning agent on top of cell segmentation

Cellpose segments microscopy images, then a PydanticAI agent answers screening questions using outlier statistics over the extracted features. Validated on the public BBBC021 drug-screen dataset.

Stack: Python · Cellpose · PydanticAI · Gemini · Docker

#agents #vision #biotech

Silicon Oracle

Stock analysis with a scoring engine you can audit

A 15-factor scoring engine, a paper-trading tracker and Gemini-written alerts, with encrypted bring-your-own-key access. Educational, not investment advice.

Stack: Flask · PostgreSQL · Supabase · Gemini · pytest

#fullstack #llm

How MARS routes a draft (simplified from graph.py)

flowchart LR
    T[Topic] --> R[Researcher]
    R --> A[Analyst]
    A --> C[Critics in parallel]
    C --> S[Synthesizer]
    S --> Q{QC score}
    Q -- weak --> P[Planner]
    P --> R
    Q -- passes --> F[Formatter]

On the edge side I work with offline vision-language models, speech pipelines and NVIDIA Jetson hardware.

How I work

Measure first. A claim without a number is a guess.
Check the checker. A validator that can't be wrong is a validator nobody tested.
Offline by default. If it only works with a good connection, it isn't done.

Lately

Reading and reviewing code in the vLLM ecosystem, and looking for the next thing worth measuring.

_{M.S. Data Science, UMBC · AWS Certified Machine Learning · Python, LangGraph, PydanticAI, MCP, vLLM, Jetson, YOLOv8, CLIP, AWS, Docker}

Navigation Menu