Dev Patel
Space mode
Portrait of Dev Patel

Software Engineer · Pittsburgh

Dev Patel

Builds AI systems. Reports the honest number.

See the workGet in touch

Now: Technical Solutions Engineer at YinzCam, building Twinky.

Space mode

I’m a huge space nerd, so I built a version of this site you can fly through.

Enter space mode

I build systems at the edge of what AI can be trusted with: agents that read the open web, a radio that hears the sky, Wi-Fi that notices you breathing. Each one ships with its tests, its benchmark, and a plain list of what it hasn’t proven yet.

Experience

Where I’ve shipped.

Seven roles across a sports-technology product team, a simulation-software company, a research lab and a consultancy. Current one first.

Now · YinzCam, Inc. · since Feb 2026

Technical Solutions Engineer at YinzCam.

Shipping websites for Burnley FC, Edgbaston, IMG Dubai and PJL — and building the internal service that took a manual step out of every ticket.

See the role

YinzCam — the official apps and websites behind professional sports teams and venues.
4major site launches
8client sites supported
1internal service shipped

YinzCam builds the official digital products for professional sports teams and venues. I build internal automation and production web systems for them — Node.js integrations, event-driven workflows, data-feed routing, website delivery, and the production debugging that happens across eight live client environments.

The part I care most about is the tooling. A lot of delivery friction is really tooling friction — a hand-off living in someone's DMs, a status that exists only in one person's head. Where I could delete that with code, I did.

  • Built a Node.js Slack-to-Jira service that turns structured Slack requests into tickets and subtasks, maps Slack users to Jira reporters, carries attachments across, syncs status back through webhooks, and recovers failed requests with retry and fail-safe logic.
  • Engineered and launched the platforms for Burnley FC, Edgbaston, IMG Dubai and PJL — implementing the site experiences, configuring data-feed integrations, establishing frontend routing and content mappings, and taking each one through production readiness.
  • Traced and resolved production failures across eight professional-sports websites, following each one through data feeds, routing rules, rendering behaviour and third-party integrations, and coordinating the release back out across engineering and client-facing teams.
  • Built a Confluence/JQL delivery tracker and ran engineering status reviews on it with the VP and CTO, coordinating dependencies across product managers, developers, designers and account managers.
Slackstructured requestNode.js servicewebhooksJiraticket · status sync
  1. 01

    Slackstructured request

  2. 02

    Node.js servicewebhooks

  3. 03

    Jiraticket · status sync

The Slack-to-Jira service, end to end — a Slack request in, a Jira ticket out, and status synced back.

Node.jsREST APIsWebhooksJira APISlack APIJQLConfluencePython

Before YinzCam

Six roles, to scale.

Pick one. The chart puts it next to school, the current role and every project.

Projects 20
Roles 7
YinzCam
School
B.S. Computer Science · Texas A&M
202120222023202420252026Now

Ansys · Austin, TX

Software Engineering Intern

May 2024 – Dec 2024 · 7 mo

Web APIs for a system migration, and 12+ bugs hunted across four languages and three architecture layers.

12+bugs closed
4languages debugged
3architecture layers
  • Implemented .NET and C# Web APIs for system migrations, using gRPC and COM interfaces to move data between microservices.
  • Closed 12+ bugs spanning C#, C++, Python, JavaScript and Angular, using dynamic tracing and memory profiling to locate faults across architecture layers.
  • Built REST/MVC request paths for asynchronous work, batch-script automation for file deployment, and IronPython scripting hooks for simulations.

Full write-upC#.NETC++gRPCCOMAngularIronPython

The Austin office — simulation software whose customers are the people who build the cars.

Texas A&M University · College Station, TX

Student Data Researcher

Nov 2023 – Feb 2024 · 3 mo

Five inconsistent disaster-response datasets turned into something a model could actually learn from.

5+source datasets
25%over baseline
2BI surfaces
  • Architected and deployed Python workflows that automated cleansing and integration across 5+ multi-source datasets.
  • Lifted validation accuracy 25% over the first working baseline through feature engineering and hyperparameter tuning, on models built over Pandas, SciPy and SQL.
  • Refined reporting with complex SQL — CTEs and views — surfacing results through Power BI and Tableau for decision-makers.

Full write-upPythonPandasSciPySQLETLPower BITableau

5+ sourcescsv · sql · apiCleansingPython · PandasFeature storenormalised SQLModelsSciPyReportingPower BI · Tableau
  1. 01

    5+ sourcescsv · sql · api

  2. 02

    CleansingPython · Pandas

  3. 03

    Feature storenormalised SQL

  4. 04

    ModelsSciPy

    ReportingPower BI · Tableau

Ingest, reconcile, model, report — the reconciliation stage is where the accuracy came from.

4C Consulting — Technology Division · On-site

Software Engineering Intern

May 2023 – Jul 2023 · 2 mo

RESTful services in Python over PostgreSQL, and a deployment path that stopped being manual.

REST+ GraphQL surface
CImanual → Jenkins
  • Employed RESTful APIs for server-side data handling and storage in Python and PostgreSQL, focused on system efficiency.
  • Designed responsive web interfaces with HTML, CSS and JavaScript; improved deployment using Jenkins, Agile, Git, Jira, Postman and GraphQL.

Full write-upPythonPostgreSQLRESTGraphQLJenkinsGitJira

Web clientHTML · CSS · JSPython APIREST + GraphQLPostgreSQLstorageJenkinsbuild + deploy
  1. 01

    Web clientHTML · CSS · JS

  2. 02

    Python APIREST + GraphQL

  3. 03

    PostgreSQLstorage

  4. 04

    Jenkinsbuild + deploy

A conventional shape, done properly — the win was automating the right-hand column.

Texas A&M University · College Station, TX

Full-Stack Developer

Aug 2022 – Dec 2022 · 4 mo

An Express and Node application where the performance work happened in the database, not the code.

20%backend performance
Redisread cache
AWShosting
  • Modeled a full-stack web application using Express.js and Node for server-side processing, integrating React and Xamarin for client-side rendering.
  • Advanced SQL database management through query structuring, indexing and Redis caching to reduce API latency.
  • Hosted the application securely on AWS and lifted backend performance by 20%, integrating client requirements through Jira and Git.

Full write-upNode.jsExpressReactXamarinRedisAWSSQL

The desk on day one, before anything was on it.

Texas A&M University — CSE · College Station, TX

Machine Learning Research Assistant

May 2022 – Aug 2022 · 3 mo

Decision trees and random forests over reverse-engineered malware features, with a dashboard for 30k+ datapoints.

5+models piloted
30k+datapoints visualised
Ghidrafeature extraction
  • Piloted 5+ predictive machine learning models using decision trees and random forests for advanced malware threat analysis.
  • Implemented a React analytics dashboard visualising 30,000+ malware datapoints, with features extracted through Python libraries and Ghidra reverse engineering.

Full write-upPythonscikit-learnGhidraReactSQLPandas

SamplesbinariesGhidrareverse engineeringFeature setPython extractionClassifierstrees · forestsDashboardReact · 30k pts
  1. 01

    Samplesbinaries

  2. 02

    Ghidrareverse engineering

  3. 03

    Feature setPython extraction

  4. 04

    Classifierstrees · forests

    DashboardReact · 30k pts

Feature extraction is the expensive stage — the classifier at the end was the easy part.

Sigma InfoSolutions · On-site

Software Engineering Intern

May 2022 – Jul 2022 · 2 mo

An e-commerce transaction path across React and Spring Boot, with a 30% query-performance win underneath it.

30%query performance
2databases via Hibernate
S3 · EC2 · RDSAWS surface
  • Improved an e-commerce transaction feature across React and Spring Boot, enhancing the interface and the backend APIs behind it.
  • Elevated PostgreSQL and MongoDB access through Hibernate, optimising data transactions and improving query performance by 30% for financial analytics.
  • Extended AWS infrastructure across S3, EC2 and RDS for new features and monitoring; maintained CI/CD through Jenkins and Kubernetes.

Full write-upReactSpring BootPostgreSQLMongoDBHibernateAWSJenkins

ReactstorefrontSpring Boottransaction APIHibernateORM layerPostgreSQLtransactionalMongoDBdocuments
  1. 01

    Reactstorefront

  2. 02

    Spring Boottransaction API

  3. 03

    HibernateORM layer

  4. 04

    PostgreSQLtransactional

    MongoDBdocuments

Two stores behind one ORM — the analytics reads were the ones worth tuning.

The work · recent first

Twenty projects.

Built on my own time. Each one says exactly what it has measured, and what it hasn’t. Every page links its repository; the newest are private for now and open on request.

01The People's Supercomputer

Kill a laptop mid-sentence.
Not one token changes.

The coordinator's dashboard after a node stopped answering mid-request: twelve layers re-planned across three survivors in 0.02 s, the failover in the event log, and a ledger that sums to zero. The nodes are local processes named like devices, running the bundled 1.9M-parameter model.

The People's Supercomputer. Pool idle laptops and gaming PCs into one network that runs open LLMs too big for any of them — and when a node dies mid-sentence, the stream carries on with output token-for-token identical to an uninterrupted run.

140 / 1401tokens identical after a kill -9
98tests
3adversarial review rounds
StackPython · PyTorch · FastAPI · WebSockets · Ed25519
Statusv0.1.0 · verified on CPU
BuiltNov 2026
Learn moreView on GitHub

02Rune Duel

Draw the rune.
Cast the spell.

A practice duel in the arena — health bars, the five-spell hand, and the rune you draw to cast.

Rune Duel. A real-time 3D wizard duel played by drawing runes with a finger or a mouse — in any browser, with the server re-drawing every stroke so a client cannot lie about it.

30spells, five to a deck
Serverre-scores every stroke
1codebase, web + iOS
StackJavaScript · React · Three.js · React Three Fiber · Node.js
StatusLive on web
BuiltJul 2026 – Present
Learn moreView on GitHub

03InjectionWatch

Seven real injections
in 426,318 pages.

The feed's dashboard: seven human-reviewed findings and where on each page the text hid. Hosts and names are blurred here — the point is the technique, not the site.

InjectionWatch. A human-reviewed feed prototype of prompt injections hidden in real web pages and aimed at AI agents — measured on 426,318 Common Crawl pages, where the honest count was seven.

426,3182pages scanned
0.44precision, up from 0.027
1.64injections per 100k pages
StackPython · Common Crawl · WARC · SQLite · FastAPI
StatusEvaluated prototype · v0.3
BuiltOct 2026
Learn moreView on GitHub

04Spectrum

A $40 radio.
Every plane, ship and satellite overhead.

The live map on the built-in simulated sky over San Francisco: eight aircraft and six ships decoded by Spectrum's own receivers from generated radio samples, with the AIS waterfall below.

Spectrum. Plug a $40 USB radio into a laptop and see the invisible world: planes and ships on a live map from my own ADS-B and AIS decoders, the next weather satellites overhead, and plain-English answers about mystery signals.

203 / 2043ADS-B frames decoded, synthetic
±7 kHzAIS tuning error tolerated
101tests
StackPython · NumPy · SciPy · Skyfield · FastAPI
Statusv0.1.0 · synthetic-IQ verified
BuiltNov 2026
Learn moreView on GitHub

Every project

Fifteen more, same rules.

Filter by what it is, sort by when. Open any row for the numbers; every one has a full write-up with the problem, decisions and diagram.

  1. The review screen after a synthetic week: three repeated workflows mined from 28 episodes and 341 recorded events, each waiting for approval before it can run.

    A local-first recorder that notices the workflows you keep repeating, turns them into reviewable skills with inputs, and runs them for you — asking before anything irreversible.

    34skills mined from 28 episodes
    107tests
    131domains never recorded

    PythonPlaywrightClaudeFastAPISQLiteTesseract

    Read the case studyGitHub

  2. The dashboard replaying a bundled synthetic living-room recording: occupied, still, breathing at 21.5 a minute, with the simulator's ground truth in yellow beside every estimate.

    Camera-free presence, motion and breathing-rate sensing from the Wi-Fi channel between three cheap ESP32-S3 boards — a deliberately small scope, a simulator with a written-down physical model, and a plain list of what isn't validated.

    93.7%5presence accuracy, synthetic
    0.61 /minbreathing error, synthetic
    82tests

    CESP-IDFESP32-S3PythonNumPySciPy

    Read the case studyGitHub

  3. The local web UI on a simulated scene: per-class keep and mute toggles, the scene tagger's readout, frame time against the 16 ms budget, and focus mode keeping one enrolled voice.

    Choose which sounds exist: a CPU-only, real-time engine that mutes or keeps nine classes of sound through your headphones, with a benchmark that scores every claim against ground truth — including the case where it fails.

    +5.3 dB6mute the leaf blower
    −4.5 dBon held-out sounds
    1.5 msper 16 ms frame

    PythonONNX RuntimeNumPySciPysherpa-onnxPortAudio

    Read the case studyGitHub

  4. The approval screen: a drafted request to reverse $142 in bank fees, waiting on one button. Approval is tied to a hash of this exact text. Fictional demo case.

    An open-source agent that disputes bills, appeals denials and reverses fees for you — it reads the document, finds the errors, drafts the letter and chases the deadline, and sends nothing without your approval of the exact text.

    67editable playbooks
    12deterministic checks
    111tests

    PythonClaudeFastAPIJinjapydanticSQLite

    Read the case studyGitHub

  5. What the compiler saw: eleven key frames picked from a 44-second screen-recorded tutorial at the moments the UI changed.

    Show it a tutorial video and it compiles the skill — steps, UI targets, variables and success checks — then runs it on your computer with your inputs, and only calls a run a success when the app's end state is checked.

    44 s8tutorial → 9-step skill
    7variables learned
    62tests

    PythonClaudePlaywrightOpenCVfaster-whisperFFmpeg

    Read the case studyGitHub

  6. The local web UI after the bundled scripted conversation: Spanish out in a synthetic demo speaker's cloned voice, English back as Spanish captions, and per-stage latency against the budget — 6.8 s median against 3.7 s.

    Real-time translation for calls and videos, spoken in your own cloned voice, entirely on your machine — where cloning requires a live, spoken, liveness-checked consent on a real microphone.

    39timed liveness prompts to enroll
    6.8 smedian latency vs 3.7 s budget
    138tests

    PythonWhispersherpa-onnxONNX RuntimeFastAPIPulseAudio

    Read the case studyGitHub

  7. The wiring guide on step 20 of 26: one jumper, exact holes, and the breadboard with this step drawn in yellow.

    An AI hardware engineer that uses a non-expert human as its hands — the model proposes a circuit, deterministic code decides whether it is safe, and a person wires it one jumper at a time.

    17810tests passing
    7 / 7injected faults caught
    20deterministic safety rules

    PythonMicroPythonRaspberry Pi Pico WMCPpytestmypy

    Read the case studyGitHub

  8. The real-trace replay over 8,754 public coding-agent trajectories: 5.0% exact hits, 0.2% of wall time saved, and a 1.8% ceiling for a perfect predictor.

    Speculative tool execution for AI agents — a lossless TypeScript library, and a measured negative result: on 8,754 real trajectories it saves 0.1–0.2% of wall time, and a perfect predictor could save about 1–2%.

    8,75411real trajectories replayed
    0.1–0.2%of wall time saved
    ~1–2%perfect-predictor ceiling

    TypeScriptNode.jsMCPVitesttsupDocker

    Read the case studyGitHub

  9. The editor — a recorded workflow as ordered steps, each with its own screenshot, title and note.

    A capture extension and editor that turn a workflow you perform once into a step-by-step guide — masked before it is saved, and stored on your own device.

    0typed values logged
    AES-GCMencrypted in the browser
    20checks on every push

    TypeScriptReactViteChrome MV3IndexedDBWeb Crypto

    Read the case studyGitHub

  10. ShakeRage — CoreMotion turns shake intensity and duration into a score.

    Two SwiftUI apps built back to back — one widget-first capture tool, one motion-driven SceneKit game.

    2apps shipped
    <100msleaderboard queries
    100+curated entries

    SwiftUISwiftDataWidgetKitSceneKitCoreMotionStoreKit 2

    Read the case studyGitHub

  11. The live interface — a question, a model-generated probability estimate, and the reasoning behind it.

    A probability engine for arbitrary real-world questions — five free-tier providers behind one fallback chain, and a JSON contract that decides which answers count.

    5providers, one chain
    4-keyJSON contract
    0paid tiers used

    FlutterDartSQLiteRESTFirebase HostingGitHub Actions

    Read the case studyGitHub

  12. The activities calendar — what replaced the group chat, rendered in the member's own timezone.

    A Rails application for a quantum computing club, built test-first and shipped through CI to 300+ members.

    300+users
    TDDRSpec throughout
    CI/CDActions → Heroku

    Ruby on RailsPostgreSQLDockerRSpecGitHub ActionsHeroku

    Read the case studyGitHub

  13. React UIRedux storeExpress API

    A point-of-sale front end in React and Redux over a Node REST API, with Postgres and maps integration.

    Reduxsingle order store
    RESTasync/await + axios
    Dockerreproducible

    ReactReduxNode.jsExpressPostgreSQLGoogle Maps API

    Read the case studyGitHub

  14. The browser running: tab strip, address bar, and the Chromium widget doing the rendering.

    A multi-tab desktop browser in Python and PyQt5, with SQLite history and AES-encrypted profiles.

    Multi-tablifecycle by hand
    SQLitehistory + bookmarks
    AESencrypted profiles

    PythonPyQt5SQLiteOpenSSLAES

    Read the case studyGitHub

  15. Close price, RSI and MACD over the same window — the features the model actually reads.

    An LSTM in TensorFlow and Keras for price movement, combined with mean reversion and backtested honestly.

    LSTMsequence model
    Backtestedheld-out history
    2signals combined

    PythonTensorFlowKerasLSTMPandasNumPy

    Read the case studyGitHub

Materials

What I actually work in.

Ranked by how many of the 27 pieces of work above each one appears in. Counted, not claimed.

Python16
GitHub Actions14
Docker6
FastAPI6
PostgreSQL6
React6
SQLite6
Claude5
Node.js5
NumPy4
SciPy4
MCP3

Languages

Python, Java, C#, C++, JavaScript, TypeScript, Swift, Dart, Ruby, SQL

Backend & APIs

Node.js, Express, .NET, Ruby on Rails, REST, Webhooks, MCP

Data

PostgreSQL, SQLite, Redis, Supabase, Firebase, Pandas, NumPy

Cloud & DevOps

Docker, AWS, GitHub Actions, CI/CD

Frameworks & ML

React, Flutter, SwiftUI, TensorFlow, scikit-learn

One more thing.

Twinky.

In build

A palm-sized, sassy open-source desk robot: it rolls around, tracks you with a tilting head and two glowing eyes, roasts you out loud, remembers what you told it, and never drives off the desk.

Concept drawings of Twinky: a big rounded head with a black visor and two glowing eyes, swept-back ear fins, and a small wheeled base
~150 mm12tall
2round LCD eyes
In buildparts ordered Oct 2026
Learn moreGitHub

Let's talk.

I'm looking for backend and full-stack engineering roles. Write here, or email me directly.

dev.patel2436@gmail.com
LinkedIn GitHub Résumé

Goes straight to my inbox. Used only to reply to you. Privacy.