Shaurya Singh

Shaurya Singh

Software EngineerReal-time voice systems, backend infrastructure, ML evaluation

Download two-page PDF Engineering record

Location
Seattle, WA

Education

University of Washington, Seattle, WA

Expected 2027

Pursuing a B.S. in Computer Science. Coursework: data structures and algorithms, computer systems (C, x86-64), machine learning, computer vision, NLP, statistics.

Experience

Confidential B2B Software Company — Software Engineer Intern

2025 – Present

Real-time voice and recruiting platform connecting browser WebRTC calls to carrier telephony and streaming voice models; built the browser app, Node.js/TypeScript APIs, live-call service, background worker and deployment.

342 merged pull requests, Jul–Oct 2026 · TypeScript, Node.js, React, WebRTC, Twilio, PostgreSQL/Prisma, Redis, Railway

Real-time calls

  • Replaced immediate WebRTC teardown on disconnect with a 15-second reconnection window, covered by browser connection-state regression tests.
  • Rebuilt turn ordering and interruption handling: serialized model turns, discarded stale responses after a caller interrupted, held speech until the caller finished and kept retracted facts out of conversation state and call recaps.
  • Capped audio buffered before the upstream provider connects at 250 frames / 64 KiB, set a 10-second readiness deadline and handled duplicate carrier start events.

Security, consent and data integrity

  • Replaced long-lived media-stream credentials with tickets that expire after 180 seconds; blocked unauthorized call events, cross-tenant caller lookups and credential leaks through third-party redirects.
  • Prevented duplicate sends under concurrent requests with in-flight idempotency reservations; reconciled voice, SMS and email opt-outs with transactional audit records; made call windows DST-safe.
  • Kept untrusted candidate résumés separate from agent instructions; documented 68 findings from an adversarial review of the voice system.

Platform, releases and matching

  • Split production on Railway into web/API, worker, real-time call, Redis (persistent storage) and monitoring services; built candidate and recruiter flows across UI, database, messaging, telephony and scheduling.
  • Added release gates: CI fails when database migrations drift from the schema, the voice runtime runs its own CI, commit identities are checked and deploys build from source.
  • Made match scores traceable by storing their evidence, contradiction records, append-only history, candidate explanations and rescore deltas.
  • Added pagination, person-name scrubbing and explicit error handling to the job-ingestion code for 69 configured job boards.

Open-source contributions

Designed and directed an automated audit of evaluation, grading, metric and reward code in open-source projects (Sep–Oct 2026): 238 issues filed in 229 repositories, 28 since closed by maintainer-merged PRs; 33 fix PRs opened, 24 with tests.

  • Data-Juicer Merged PR #1080 Sep 2026

    In a pairwise LLM-judge evaluator, failed judge calls were scored [0, 0] and pulled both models' averages toward zero; they are now left out of the averages and counted.

  • ir_measures Merged PR #84 Sep 2026

    Depending on request order, a plain nDCG requested beside an nDCG with custom gains ran on the remapped relevance labels (0.50 instead of 0.76); it now uses the original labels, tested in both orders.

  • Otter-Grader Merged PR #1021 Sep 2026

    The otter assign command logged "All autograder tests passed." when grading the solutions notebook produced 0 of 0 results; it now fails with the underlying error, covered by 3 tests.

  • ai-gateway (Go) Shipped in v1.5.10 Oct 2026

    A malformed word-filter blocklist loaded as empty and passed config validation; startup and validation now reject it. Both commits shipped in release PR #455 with changelog credit.

  • open-instruct Approved PR #1917 Oct 2026

    With the non-additive R1 format gate on, correctly formatted responses lost their verifiable reward unless the format reward equaled 1.0; the open PR checks format validity instead.

  • OSWorld Revised fix merged (PR #614) Oct 2026

    A numeric result check matched digit substrings, so 125.27 passed against a gold answer of 25.27; the maintainers merged a revised fix that includes the commit from PR #598.

Selected projects

  • From-Scratch Language Model (PyTorch)

    Trained a 10.7M-parameter decoder-only transformer with a byte-level BPE tokenizer from random initialization on Apple Silicon; validation loss fell from 8.45 to 3.54 over 400 steps.

  • Global Signals Observatory (Python, SQLite)

    Python service built to ingest 16 public feeds (space weather, earthquakes, weather alerts, orbital data, river and tide readings) every 15 minutes into SQLite and serve them through 22 read-only endpoints; a health endpoint reports failure when data goes stale.

  • Private Infrastructure Control Plane (Python, Docker, systemd, SQLite)

    Self-hosted service host for software agents: role-scoped MCP and API gateways; services reachable only on the private network; health probes that distinguish stale, disconnected and unmeasured services; SHA-verified off-machine backups; a prototype OAuth 2.1/PKCE gateway for read-only remote access.

  • AI Builder (Python, pymoo)

    Cabin-design prototype that runs locally: Seattle residential-code rule engine, NSGA-II search over cost, space and comfort, and IFC/DXF/STL/glTF export.

  • Automated Bug Detection

    Prototype VS Code extension that flags 20+ categories of bugs and security vulnerabilities in Python and JavaScript as code is typed, using syntax-tree traversal and regex.

  • CiteGuard

    Prototype that checks each legal citation in a language model's answer against an external source and grades it as fabricated, misattributed, stale or unsupporting, so the model cannot approve its own citations.

  • shauryasingh.com (Astro)

    Static site with one build dependency that ships no JavaScript and makes no third-party requests of its own; the build recomputes the numbers the site states about itself, fails if they change and checks every internal link.

Founder experience

TurtleShell — Founder

Nov 2022 – Oct 2023
  • Clustered 10,000+ LAPD-reported incidents into 40 risk zones with K-means in PySpark.
  • Built a Swift/CoreLocation iOS prototype with one-tap SOS alerts and route choice; accepted into Microsoft for Startups Founders Hub.

SkillTern — Co-Founder & CTO

May 2022 – Aug 2023
  • Built the curriculum, teaching website and certification workflow for a CS and AI education nonprofit that taught 350+ students through 4+ partner organizations.
  • Wrote the instructor and certification runbooks so volunteers could run courses without the founders.

Recognition and writing

  • WatGov Hackathon, University of Waterloo. Runner-up, Jul 2023: co-authored policy paper on reducing AI-driven phishing.
  • IKEA × The Knowledge Society. Team strategy project on Gen-Z retail (2023): primary consumer research, practitioner interviews and a 19-page recommendation for a browser-based room-layout product.
  • Writing. Four essays on optogenetics, brain-computer interfaces, connectomics and convolutional neural networks, published on Medium.

Technical skills

Languages
Python, TypeScript/JavaScript, Go, SQL, Swift, C, HTML/CSS
Web, backend and infrastructure
Node.js, React, REST APIs, WebRTC, Twilio, PostgreSQL/Prisma, Redis, SQLite, Docker, Linux, systemd, Cloudflare Workers, Railway, CI/CD, Git
ML and evaluation
PyTorch, NumPy, pandas, scikit-learn, PySpark, pytest