About

I work on the back end at Level AI: data models, APIs, background jobs, search, and the alerting and dashboards in Datadog.

Selected work

Projects in progress

01

atropos

A small Rust proxy that keeps API keys split: the server stores only ciphertext, my machine stores only a random token, and a request has to pass through the proxy to use the key.

The protocol spec and the Cargo workspace are in. The core crypto crate is next.

RustRust · Spec stage

Authorization: Bearerat1_7h2kq4m6xw3b5n7d_3vq7…

  1. 1kid → rowfound
  2. 2HKDF(secret, kid) → verifiermatches
  3. 3HKDF(secret, kid) → enc_keyin memory
  4. 4open(ciphertext, enc_key)provider key
  5. 5forward to the row’s hostapi.anthropic.com

02

mi8ess

A chess engine written in Rust.Reads any position. Can’t move yet.

RustRust · In progress · 2 of 8

03

netwatch

A network analyzer. Starts small, with DNS queries and spotting tracker domains.DNS queries in, tracker domains out.

24.2.0.1 → www.example.comAllowed

242.8.0.1 → www.test.comAllowed

44.42.41.24 → ads.google.comBlocked

12.1.11.12 → telemetry.cloudflare.comBlocked

RustRust · In progress

04

loba

A Discord bot that announces Codeforces contests. It posts when a contest is listed, then 24 hours, 2 hours and 15 minutes before it starts.Codeforces contests, announced on Discord.

  1. New contestlisted
  2. Starts in 24 hours24h
  3. Starts in 2 hours2h
  4. Starts in 15 minutes@contest-ping

PythonPython · In progress

Everything else →

Experience

  1. Software Engineer · Level AI

    Oct 2024 — now

    PythonDjangoCeleryPostgreSQLSnowflakeElasticsearchRedisFastAPIMongoDBPub/SubReactDatadog

    • Coaching analytics, built end to end. A new product area that shows whether agent coaching works: a KPI catalogue, KPI goals with baselines and targets on coaching action items, APIs for the dashboard and coaching impact, and the React and TypeScript dashboards for coaches and agents. Metrics come from daily Celery rollups over Snowflake and PostgreSQL, bulk-upserted under a per-tenant distributed lock, instead of being computed on each request. Removed N+1 queries from the list and dashboard endpoints.
    • Bulk actions on up to 20,000 QA evaluations. Accept, undo, change evaluator, change due date and clear, run as background jobs that report progress and pick up again after a page refresh. Select-all-matching with per-row exclusions, eligibility checks that match the single-item rules, stale-job recovery, a failure reason for each item, and before-and-after audit rows.
    • AI coaching simulator. FastAPI services (Pydantic v2, MongoDB) for practice scenarios and personas inside a coaching session, with practice progress. Practice calls are scored by the platform’s AI scoring through a new synchronous transcript API, with service-to-service auth.
    • QA form import and export. Customers export a QA form, edit it offline and import it across teams and tenants. It started as CSV and moved to XLSX; exports know the form’s version and imports land as drafts.
    • Measuring AI scoring. Most of the backend that measures AI QA scoring against human-labelled datasets: dataset upload and validation through signed URLs, accuracy APIs, and Celery runs with their failure states. Also human overrides of AI answers, and customer tags as scoring inputs with a backfill of past data.
    • SLA tracking. On Time and Breached status for evaluations, from the Django API through Elasticsearch filters and hourly Celery reminders to a Pub/Sub change-data-capture consumer that works out SLA time when a record is indexed.
    • Case assignment. Weighted round-robin that respects evaluator time off, workload limits per goal period, and a lock that fixed a race in rule-based assignment.
    • Search and filters. Starts-with and does-not-start-with operators on reusable Elasticsearch prefix-match helpers, NULLs-last sorting, and input checks that turned server errors into clear 400s.
    • Monitoring and on-call. Wrote and set up the alerting, dashboards and metrics in Datadog, with PagerDuty for on-call alerts and ELK for logs. Was on PagerDuty on-call for over two months in total, and documented the issues that came up.
    • Tests and quality. 136 merged changes across 5 repositories, and about half the hand-written code in them is tests. Selenium and Cucumber regression suites for five product areas, with test-only APIs that make them repeatable. Fixed data-isolation and production issues raised by enterprise customers.
  2. Software Engineer Intern · BYJU'S

    Jan — Jun 2023

    PythonDjango REST FrameworkPostgreSQLAWS S3, SQS, LambdaSnowflake

    • Built downloadable analytics reports with SQS and Lambda writing to S3.
    • Added device last-sync time end to end: Snowflake updates with the data team, cron jobs reading from S3, models and APIs.
    • Wrote leaderboard APIs and backfill scripts, and ran staging and production migrations.
  3. B.Tech, Information Technology · NIT Kurukshetra

    2020 — 2024

    • 9.0 CGPA. Got in with a 99.44 percentile in JEE Main.
    • 300+ problems solved on Codeforces and LeetCode.

Everything else

NIT Kurukshetra Website

The institute's new website, front end and back end in one repo. I worked on the back end as a member of ISAC, the student club that builds it.

TypeScriptNext.jsDrizzle ORMTypesense

Image Flipper

Flips bitmap images upside down, reading the file format by hand.

C++

Toolbox

Languages
PythonRustJavaGoTypeScriptC++SQL
Backend
DjangoDjango REST FrameworkCeleryFastAPIPydanticSpring Boot
Data
PostgreSQLSnowflakeElasticsearchMongoDBRedisPub/Sub
Cloud
AWS S3, SQS, LambdaDockerGitLab CIGitHub Actions
Monitoring
DatadogPagerDutyELKSentry
Testing
pytestSeleniumCucumberTestNGJest
Daily driver
Fedora Linux

Contact

How this was made

Built with Claude, by Anthropic.

My work is on the back end, so I used Claude to get a clean page up quickly and keep the attention on the work it describes. All the content here is my own and factually accurate.

Jump to

↑ ↓ move↵ openesc close

You found the 8.

mi8ess can read this board but cannot generate moves yet, so both sides are moved by hand.

Play e4 to start

Open mi8ess

Click anywhere to close