aayush.
Avatar: anime-style portrait with glowing blue eyes on a red sky

Aayush Sharma

Backend Python developer · --:--

I build backend systems in Python: APIs, async pipelines and LLM apps. And I check that they actually work. I compare against a baseline before claiming accuracy, and I test the AI grader against my own ratings before trusting its scores.

how to read this page

Every number links to its source. For example, my support-ticket classifier hit 92.1%[1] accuracy. Hover or tap the [1] to see where that number comes from.

Open to SWE internships · 2027 batchIndia · open to Bangalore / relocation
Resume
01 / work#

P1AI Support Agent

2026

Classifies Apple support tweets, drafts grounded replies, and decides when a human should take over. Then I graded my own grader.

PythonGroqgpt-oss-20bChromaDBpandasSciPy
GitHubReport
Intent accuracy · 6 classesn = 239
keyword rules41.8%[1]
always predict majority55.2%[1]
few-shot LLM92.1%[1]

One class is 55% of the data, so 55.2% is the floor. Any accuracy claim has to beat that first.

Data funnel · log scale
  • raw support tweets2.8M[2]
  • AppleSupport rows~213k[2]
  • issue → reply pairs~106k[2]
  • known-good resolutions (RAG index)~3,000[2]
  • hand-labeled golden set239[2]
Auto-handle vs escalate · golden set
escalate to a human 185[1]auto-handle 54[1]

Conservative on purpose. Billing, account access and hardware-safety issues always go to a person. A wrong auto-reply costs more than a slow one.

The part most demos leave out

I graded the grader, and it failed.

I scored 10 generated replies myself, blind, and compared them with the LLM judge's scores. The judge was lenient and couldn't rank quality, so its scores are reported as unvalidated instead of being passed off as results.

20%[1]
exact agreement
ρ −0.46[1]
spearman vs me
6/10[1]
judge scored higher

Also documented: 5[1] failure modes and 15[3] non-obvious design decisions.

P2Second Brain

2026

A memory layer for my PC. It indexes files, screenshots, browser history and clipboard, then lets me search all of it in plain English. Nothing leaves the machine.

FastAPIReactSQLite FTS5ChromaDBOllamasentence-transformerswatchdogWebSockets
GitHub
Second Brain · Your PC's Memory Layer
Second Brain dashboard showing indexed documents, browser logs and a recent activity timeline
271[4*]
documents indexed
2,405[4*]
browser history logs
0[5]
cloud calls
indexing + search pipelineclick a stage
A background tray daemon watches Desktop, Documents and Downloads in real time and pulls browser history from Chrome, Edge, Firefox and Brave, plus the clipboard. A single-instance lockfile stops two daemons fighting over one index, and CPU throttling keeps it out of the way.

P3DayLog

2026

Write a messy #100DaysOfCode log and get back tidy bullets, a tweet-ready draft, a streak and a heatmap. It's deployed and people can use it.

FastAPISQLAlchemy 2.0Postgres (Neon)ReactTypeScriptTailwind v4Gemini / Groq
LiveGitHub
Which LLM key gets usedbring-your-own-key[6]
  1. try 1your keystays in the browser, sent per request, never stored or logged
  2. try 2shared keyserver's own key, capped at 20 uses per user per day
  3. try 3no-AI fallbackplain-code tidy-up, so the app still works

The LLM sits behind one function. Streaks, heatmap levels and counts are plain code, so they're right even when the model is down.

  • Own auth. email + password with argon2, Google sign-in, hashed verification tokens, rate-limited login
  • Same-origin by design. FastAPI serves the build; Vercel rewrites proxy to Render so the session cookie never goes cross-site
  • Shipped. multi-stage Dockerfile, Postgres on Neon in prod, SQLite locally
  • Tested. pytest with network mocked; Vitest for the heatmap and form validation

P4Valorant Tracker

2025

A Flask + MongoDB stats tracker for me and my friends: match history, per-player stats, win rates. Designed, modelled and built solo, with API responses under 200 ms[7*] in local benchmarks.

FlaskMongoDBJavaScriptRESTMVCGitHub
02 / open source#

03 / activity#

78 active days in the last year
NovDecJanFebMarAprMayJunJulAugSepOct
lessmore
133[9]
LeetCode solved
76E48M9H
74[9]
LeetCode active days
29[9]
max streak (days)
41[10]
GitHub contributions
04 / stack#

Languages
PythonJavaJavaScriptTypeScriptSQL
Backend
FastAPIDjangoFlaskNode.jsExpressRESTWebSocketsasync
Data
PostgreSQLMySQLMongoDBSQLite FTS5ChromaDBSQLAlchemy
AI / LLM
RAGfew-shot promptingembeddingsLLM-as-judgeGroqOllamasentence-transformers
Tools
GitDockerPostmanAWS (EC2, S3, IAM, VPC)VercelRenderpytestReactTailwind
05 / education#

Parul University

Expected 2027

B.Tech, Computer Science and Engineering

CGPA 7.09 · Class 12: 81.25% · Class 10: 100%

  • —NPTEL · Computer Networking (OSI, TCP/IP, routing, network security)
  • —AWS Fundamentals · EC2, S3, IAM, VPC
06 / contact#

Hiring a backend intern who checks their own work?

I'm looking for a software engineering internship (2027 batch), in Bangalore or anywhere I can relocate to. The fastest way to reach me is email.

Email meResume
07 / receipts#

Where every number above comes from. Items marked * are self-reported, with no public source yet, and are labelled that way wherever they appear. Activity stats were snapshotted on 21 Oct 2026.

  1. [1]AI Support Agent · report.md (results, baselines, judge check)
  2. [2]AI Support Agent · README (pipeline, dataset sizes)
  3. [3]AI Support Agent · decision_log.md (15 design decisions)
  4. [4*]Second Brain · dashboard screenshot from my machine (shown above)
  5. [5]Second Brain · repository
  6. [6]DayLog · README (architecture, key priority, auth)
  7. [7*]Valorant Tracker · local benchmark, not published
  8. [8]GitHub · merged pull requests by Aayush-shrma
  9. [9]LeetCode · profile Aayush9139 (snapshot at build time)
  10. [10]GitHub · contribution calendar (snapshot at build time)