Skip to content
View yadava5's full-sized avatar
2 followers · 1 following

Highlights

  • Pro

Block or report yadava5

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
yadava5/README.md
Ayush Yadav — the estate drawn as one circuit board. Six product marks sit on a shared bus and every printed edge is a checkable claim: Cadence and Applied share forced row-level security, Glyph and jetpack share SIMD, AutoML feeds Postgres. VisualAssist hangs on an open stub — no URL, it needs an iPhone with LiDAR in your hand. B.S. Computer Science, Miami University, May 2026; open to full-time software engineering roles; aesh at gmail. Port one of four on the link bus: ayush-yadav.com, the portfolio, wearing the Waymark. The three page buses run whole through the row — every port taps off its own lane, and no lane leaves the bundle.Port two of four: the résumé, as a PDF. Its mark is drawn from the Waymark's own construction — the same bar, with a sheet standing astride it — on a tap off the rust lane.Port three of four: the LinkedIn profile, carrying LinkedIn's own mark. The rust tap runs on through the résumé's package to reach it.Port four of four: email — opens a mail draft. The mail mark is drawn in the board's own palette; the zinc tap reaches it past every other package. The three page buses continue past the link ports, whole, into the evidence — sections two to seven. The three page buses dive to an inner layer at via pads and run beneath the prose; the silkscreen legend says what the drawn links are: test points — probe the claim.

C++ · TypeScript · Python · Java · Swift — B.S. Computer Science, Miami University (May 2026). Cincinnati, Ohio. Open to full-time software engineering roles.

Six systems, and a year of paid data work. Every number is drawn on the plate beside it and re-derived in CI from a pinned commit, except §I and §VII's grant, which are my word and say so where they stand. Four of them ship a system card, a print-format walkthrough of the architecture and the evidence.

The three buses resurface from the inner layer at their via pads and enter the next plate. I · Work — the one section drawn as an empty footprint: a component the board expects that is not fitted, because the year of paid work is off-repo. A 57.8M-row field-usage table, distilled from 1.6M Oracle Analytics query logs at Miami University. None of these numbers can be re-derived by a reader — the data belongs to Miami University and to a competition; this section is attested, not derived. The three page buses dive to an inner layer at via pads and run beneath the prose; the silkscreen legend says what the drawn links are: test points — probe the claim.

ITSM Data Integration Intern at Miami University, June 2025 to May 2026, and team lead of three at DataFest 2026. This section and §VII's grant are the two places you take my word for a number, and the plate says so on its face.

The three buses resurface from the inner layer at their via pads and enter the next plate. II · jetpack — parallel gzip on JDK 25, drawn as the SIMD bus fanning into four deflate blocks stitched back to one gzip member: 422 against 66.2 MB per second single-threaded, 6.4 times, on an M1 Pro from a 3-fork JMH run. The against-self result is printed on the plate: the JDK's own Adler-32 intrinsic does 14.06 GB per second and is not beaten; the hand-vectorised checksum reaches 4.26, bit-identical to java.util.zip. The three page buses dive to an inner layer at via pads beneath the prose, and the section's test-point rail taps off the rust lane.Test point one — probe the section live, in the browser.Test point two — the system card, a print-format walkthrough of the architecture and the evidence.Test point three — the source repository.

Parallel gzip on JDK 25 — one virtual thread per block, a bounded in-flight window so peak memory is independent of file size when streaming — the in-memory API buffers the whole result — and a byte-valid single gzip member rather than the concatenated members parallel compressors usually emit.

The three buses resurface from the inner layer at their via pads and enter the next plate. III · Glyph — SIMD kernels over a course-provided MNIST net, drawn as one register-lane bundle carrying current and two unpopulated footprints beside it. Kernels are written for NEON, AVX2 and AVX-512, but the family is chosen at compile time, there is no runtime dispatch, and the reference machine is arm64 — so only the NEON bundle was built, and only its lanes run to benchDot/256: 3.5× from OpenMP threading over the same kernels, its own single-threaded build, from committed runs. The AVX2 and AVX-512 slots are printed with their pads bare and their leads left open, silkscreened: written, not measured. The package that ships is WASM_SIMD128, 43,751 bytes, byte-identical to main and checked daily by this page's own CI. The against-self results are printed: the same flags lose 8 of 12 matrix-op cases and buy nothing end-to-end — worst is benchAxpy/128, 10.7× slower, memory-bandwidth-bound, where axpy never pays at any size measured — and 97.01% is a training-time number — the test set that graded the net also picked its checkpoint. The three page buses dive to an inner layer at via pads beneath the prose, and the section's test-point rail taps off the rust lane.Test point one — probe the section live, in the browser.Test point two — the system card, a print-format walkthrough of the architecture and the evidence.Test point three — the source repository.

The network is not mine: a course-provided C++ MNIST net. The work is what happened to it — hand-written SIMD kernels over a scalar fallback, with Shree Chaturvedi credited for kernel contributions. The React browser application is mine.

The three buses resurface from the inner layer at their via pads and enter the next plate. IV · AutoML — the tool library drawn as one registry strip: a 44-pin comb, one pin per tool schema, with seven tool-set taps below it — no phase is handed all 44. The against-self line is the sandbox: it guards against accidents, not adversaries — the network is an env var the beta template renders as bridge, with no cap-drop, no pids-limit, no seccomp. The three page buses dive to an inner layer at via pads beneath the prose, and the section's test-point rail taps off the zinc lane.Test point one — probe the section live, in the browser.Test point two — the system card's source. AutoML is the one system here whose card is written but not deployed.Test point three — the source repository.

A dataset and a sentence in, a trained model out, over a LangGraph state machine and an MCP tool server. GPL-3.0 at the commit this page pins; a relicence to PolyForm Noncommercial is proposed in PR #5, which still targets the pre-repin history, and needs my co-author's review.

The three buses resurface from the inner layer at their via pads and enter the next plate. V · Cadence — the whole API drawn as one part: 36 route handlers enter one serverless function, and its single output passes through the one door in a wall — row-level security, FORCEd — before reaching the 7 tables behind it. The against-self line: every one of the six services carries a conditional owner guard; a caller that forgets the identity still sends the query, and the database's refusal is the one that cannot be forgotten. The three page buses dive to an inner layer at via pads beneath the prose, and the section's test-point rail taps off the verdigris lane.Test point one — probe the section live, in the browser.Test point two — the system card, a print-format walkthrough of the architecture and the evidence.Test point three — the source repository.Test point four — the row-level-security isolation suite, run against real Postgres.

Type lunch with sam friday 1pm and get a task dated Friday 1pm; drag it onto the calendar to make it an event. The engineering is what the audit turned up: application code that filters by user is code that has to remember to filter, so the database enforces it instead.

The three buses resurface from the inner layer at their via pads and enter the next plate. VI · Applied — the mail triage drawn as the part that ships: the rules layer grades a labelled strip of 96 messages and gets 2 wrong — 0.979 macro-F1, rules layer alone. Below it, drawn dashed with its lead left open — it is not what ships — is the full cascade, all three layers, and it scored lower: 0.9583, a number a tracker line records. What is deployed runs only that first layer. The three page buses dive to an inner layer at via pads beneath the prose, and the section's test-point rail taps off the verdigris lane.Test point one — probe the section live, in the browser.Test point two — the system card, a print-format walkthrough of the architecture and the evidence.Test point three — the source repository.Test point four, the browser build of the mail classifier, its page and its quantized weights.

Your inbox already holds the verdict on most applications you have sent. Three layers read it locally, cheapest first; the hosted app runs the first alone. The band just under the confidence gate is walked to a human, and anything below the floor is dropped. The model is allowed to say it does not know.

The three buses resurface from the inner layer at their via pads and enter the next plate. VII · VisualAssist — the alert policy drawn to scale from the phone in your hand: inside 0.5 m a critical buzz and a spoken Stop, out to 1.0 m a warning pulse and a spoken Caution, and past 1.0 m the automatic check does nothing at all — silence begins there, not at the widest ring. The 2.0 m arc is the reach of the manual Describe call: zones farther out are not mentioned. Below it, the settings slider the app actually shows — labelled in metres, its value read aloud — drawn with its lead ending open before LiDARService, the threshold it never consults. A live control that does nothing lies to the person the app is for. And a three-second announcement cooldown sits above the critical branch, so the Stop the code comments call always-announce can be swallowed for as long as that. The three page buses dive to an inner layer at via pads beneath the prose, and the section's test-point rail taps off the rust lane.Test point one — the source repository, the one probe this section has.The next test point is an unpopulated footprint, drawn dashed with its lead open — there is no live deployment to probe; VisualAssist needs an iPhone with LiDAR in your hand.

An iPhone app for low-vision users, in Swift: ARKit LiDAR depth becomes speech and haptics, so the phone tells you what is in front of you before you reach it. MUCAT Design Innovation finalist, with a $2,500 prototyping grant — attested, like §I. It is the one system here you cannot click into; it needs an iPhone with a lidar sensor in your hand.

The three buses resurface from the inner layer at their via pads and enter the next plate. Colophon — the board's edge connector: the three buses land on their pads and leave the page. 31 claims, 31 commands: 29 re-run in CI from pinned commits and 2 against the live deployment; what is my word rather than a derivation says so where it stands.

build/claims.json ties every derived number to a commit SHA and to the shell command that re-derives it from the blob at that SHA. The gate runs those commands on every push to main, on pull requests, and once a day, rejects any number no row accounts for, and closes with a negative test that breaks the plates on purpose and fails if a check sleeps through it.

The guarantee is narrower than it sounds: numbers are gated, prose is not. Every false claim an audit has turned up on this page has been a noun or a verb, not a digit.

Everything above, with LifeQuest — a quest tracker for people rebuilding after a layoff or retirement — is collected at ayush-yadav.com.

Pinned Loading

  1. glyph glyph Public

    High-performance C++ MNIST neural network, 784-100-10, at 97.01% test accuracy. The hot loop is hand-written four times — AVX-512, AVX2, NEON and wasm simd128 — over a scalar fallback, selected at …

    TypeScript

  2. applied applied Public

    Email-powered job application tracker — syncs Gmail & iCloud, classifies emails with ML, tracks your job search pipeline

    Python

  3. lifequest lifequest Public

    LifeQuest turns real-world routines into map-based missions for laid-off professionals and retirees, rewarding progress with Quest Coins and perks to rebuild structure, community, and job reentry. …

    TypeScript

  4. cadence cadence Public

    Production-ready full-stack calendar & task management platform with NLP-powered smart input, multi-calendar support, and real-time conflict detection. Built with React 19, TypeScript, PostgreSQL, …

    TypeScript

  5. ai-augmented-auto-ml-toolchain ai-augmented-auto-ml-toolchain Public

    Agentic AutoML platform — a dataset and a sentence in, a trained model out. A LangGraph state machine runs the pipeline over an MCP tool server, dealing 44 tool schemas as seven phase-scoped sets s…

    TypeScript

  6. jetpack-compress jetpack-compress Public

    High-throughput, gzip-compatible parallel compression engine on JDK 25: virtual-thread block framing stitched into one byte-valid gzip member, a hand-vectorized SIMD Adler-32 (Vector API), and FFM …

    TypeScript