Jay
Hemnani.

currently positioning asforward-deployed engineer

A data and AI engineer with a stubborn preference for systems that survive the Monday morning after the demo. Spent the last four years shipping production computer-vision pipelines, real-time data warehouses, and the kind of agentic AI plumbing that gets talked about more often than it gets built. Two IEEE papers as an undergrad. Three patches landed in vLLM, the Anthropic MCP SDK, and Google's A2UI. One personal site that an LLM can read directly.

§ 02Featured work, in six acts.the catalogue
ID 01ID 07ID 10BALLComputer Vision · MLOps
fig. 01, tactical interaction graph, 22 fps/work/fifa
№ 01
Computer Vision · Sport Analytics

FIFA Soccer DS.

YOLOv8 detects, ByteTrack persists, a GraphSAGE scaffold builds the tactical graph, all wrapped in DVC, MLflow, and a FastAPI shipping container.

A production-shaped computer-vision pipeline for soccer video, end-to-end and reproducible. Seven-stage DVC pipeline, MLflow experiment tracking, twenty-two frames per second. The thing actually runs.

YOLOv8ByteTrackPyTorch GeometricMLflowDVCFastAPI
Read the dossier
auditor · depth/loop capscausalplanneredgeselectoradversaryred teamsupporterblue teamdialecticjudgewriterVERIFIED · FALSIFIEDAgentic AI
fig. 02, causal-adversarial graph/work/cag
№ 02
Agentic AI · Causal Reasoning

CAG Deep Research.

LangGraph plans a causal graph of a question; red and blue agents fight over each edge; a dialectical judge rules.

A system built to falsify its own claims. Each causal link is attacked by a paired adversary and supporter before a judge returns verified, falsified, or unclear. Hexagonal architecture, swappable search and LLM providers, Ollama with a Groq fallback.

LangGraphLangChainOllamaGroqPydantic
Read the dossier
DIM_SYMBOLtickernamesectorDIM_DATEdate_idyearquarterDIM_SOURCEsrc_idfeedlatencyDIM_EXCHANGEmictzregionFCT_TICKStsopenhighlowclosevolumeData Engineering
fig. 04, star schema, 25 yrs of ticks/work/stock
№ 04
Streaming · Warehouse

Stock Data Platform.

Kafka, eighteen Airflow DAGs, a TimescaleDB star schema, and a dashboard that wants to be Bloomberg.

Ten tickers, twenty-five years of history, four upstream sources, a unit test suite, seven services in a single Docker compose. Built the way a production data platform ought to be built, which is to say, with relentless skepticism toward the data.

KafkaAirflowTimescaleDBDockerDash
Read the dossier
WEBMCP · TOOL REGISTRY❯browseragentMCPhubsearch{ }JSONresume{ }JSONskills{ }JSONcontact{ }JSONtheme{ }JSON8 tools · JSON Schema in, typed results outAI-Native Web
fig. 05, webmcp tool registry/work/webmcp
№ 05
AI-Native Web · WebMCP

WebMCP, on this very site.

An LLM can browse this portfolio the way humans do, only faster, and without scrolling.

Eight structured tools exposed via the WebMCP browser API. Project search, resume, skills, contact, theme, all callable by an agent with JSON Schema inputs and typed results, instead of scraping the DOM.

TypeScriptNext.js 16WebMCPJSON Schema
Read the dossier
VDOS SPECTRUMω →CONTACT GRAPHChemBERTa · ligandTri-modal
fig. 06, vdos spectrum/work/nobel
№ 06
Research · Scientific ML

Nobel Data Intelligence.

Tri-modal deep learning for protein stability: sequence, structure, and a vibrational signal nobody else is using.

ProtT5 for sequence. GATv2 over structural graphs. VDOS spectra from normal-mode analysis, the novel bit. ChemBERTa for chemistry, fused through learned gated attention. Unit-tested across the fusion stack, eight notebooks, Colab-ready.

PyTorch GeometricProDyBioPythonRDKitTransformers
Read the dossier
See all 27 projects in the index
§ 03In this issue.contents

Everything else on the site, indexed the way a magazine indexes itself.

  1. AThe Work27 entries/work →Computer vision, agentic AI, a Rust stock platform, a Go voice-to-text tool. The full archive, with deep-dives on most of it.
  2. BThe Writing2 essays · 27 writeups/writing →Project post-mortems and short essays on building in 2026. Unflattering details left in. Still under construction, in the honest sense of the phrase.
  3. CThe ParticularsCV · OSS · Pubs/about →Four years of work, two IEEE papers, three merged pull requests into ecosystem repositories, and a Rubik's cube record that will not make the resume.
Correspondence · 04

Hiring? Building? Stuck?

I am looking for forward-deployed and data-engineering roles, though senior software is also on the table. Staff-shaped problems welcome.

Open the linejayhemnani992000@gmail.com
GitHub ↗LinkedIn ↗Twitter ↗Résumé →