Mahir Dalal

Builds systemsthat haveto be fast.

Third-year CS at IIIT Surat. Competitive programmer. I build agentic backends and distributed systems — and I plot the record rather than describe it.

See the work
  • C++
  • TypeScript
  • Python
  • Java
  • Next.js
  • React
  • Node.js
  • Express
  • NestJS
  • PostgreSQL
  • Prisma
  • MongoDB
  • Docker
  • Socket.IO
  • WebSockets
  • Event-Driven Architecture
  • Microservices
  • Distributed Systems
  • Concurrency
  • Multithreading
  • pgvector
  • Groq
  • Gemini

JAN-AI

768-dim embeddings, HNSW index, 5 report categories

The AI cannot break a report

Civic issue reporting — people photograph a pothole or a broken streetlight, mark it on a map, and the report enters a tracked dashboard.

Hard part. Two people photographing the same pothole must not become two reports, and an embedding comparison against every prior report does not scale. Dedup runs as an agent: Gemini embeddings at 768 dimensions over a pgvector HNSW index narrow the field, Groq-hosted reasoning judges the survivors, and a human confirms. A rejection-memory constraint records every dismissed match so the same false positive is never raised twice.

Stack
Next.js, Express, Prisma, PostgreSQL, pgvector, pg-boss, Gemini, Groq
Role
Solo — backend first, then the frontend on it
Status
Deployed

DSA/CP Trainer

160 unit tests, 32 topics, 14 tables

Deterministic by construction

Pulls your solve history from LeetCode and Codeforces, works out which topics you are actually weak at, and ranks what to solve next.

Hard part. Mastery could not be solved/total, which hands you 1.0 for a single lucky solve. It is a Laplace-smoothed rate pulled toward 0.5 by a confidence term that scales with sample size, so one solve out of one reads as "no evidence" rather than "mastered". The selection is ordinary arithmetic end to end — no model chooses, ranks or filters anything, so identical inputs always give identical output. The LLM only writes hints on top of decisions already made.

Stack
Node, Express, TypeScript, Prisma, PostgreSQL, Next.js 16, Groq
Role
Solo
Status
Deployed

Self-Healing Network Simulator

15 nodes, 21 edges, 4 attack models

Routes around what it cannot trust

A distributed real-time network simulation where the routing layer reroutes around nodes it no longer trusts.

Hard part. Two routing algorithms run concurrently over the same live topology — a naive OSPF-style shortest path against a health-aware Dijkstra — so the cost of trust-blind routing is visible rather than argued. Node health feeds back into the cost function, letting the health-aware router drop compromised nodes mid-transit without any manual intervention.

Stack
Node, Express, Socket.IO, React 19, Three.js
Role
Solo
Status
Not public
Source not public

TripSync

No cloud database at all

Group trip planning — shared trips, destination tabs, and hotel links collected in one place instead of scattered across chat threads.

Hard part. Built deliberately as a weekend project, and the interesting call was dropping the cloud database entirely. It runs on a local SQLite file through better-sqlite3, with passwordless name-plus-email sessions, which removes the whole OAuth and connection-pooling surface from a tool that only ever needs to serve one group at a time.

Stack
Next.js 15, better-sqlite3, NextAuth v5, Zod
Role
Solo — weekend build
Status
Source available

Every rated contest, 71 in total, as each platform reports it. Ratings are on their own platform’s scale and are never converted.
LC WC509LeetCode2026-07-052,6401963+7
LC BW186LeetCode2026-07-049061956+42
LC WC508LeetCode2026-06-285271914+66
LC WC507LeetCode2026-06-219,0891848-25
LC BW185LeetCode2026-06-205,2691873-7
CF Order Capital Round 2Codeforces2026-06-182,5911454+50
CF Edu 191Codeforces2026-06-092,0401404+78
CF R1102 D2Codeforces2026-06-075,9561326-18
LC WC505LeetCode2026-06-071,5201879+39
LC BW184LeetCode2026-06-069,1691840-23
CF R1099 D2Codeforces2026-05-212,7691344+77
LC BW182LeetCode2026-05-093,3471864+6
CF R1093 D2Codeforces2026-04-135,5131267-19
CF R1091 D2 and CodeCraft 26Codeforces2026-04-075,3491286+19
LC WC494LeetCode2026-03-222,5931857+29
CF R1087 D2Codeforces2026-03-213,4261267+78
LC WC493LeetCode2026-03-153,1521828+28
LC BW178LeetCode2026-03-145,6841801+6
LC WC492LeetCode2026-03-083,6661794+25
LC WC491LeetCode2026-03-012,7841769+39

  • Top 3%
    LeetCode, globally
    Peak contest rating 1963 — Knight. 27,337 of 882,088 rated.
  • Top 15,000
    Google BigCode 2026
    Among more than 100,000 participants.
  • 1,000+
    DSA problems solved
    Across Codeforces, LeetCode, CodeChef and AtCoder.
  • Top 2%
    JEE Main
    98.2 percentile, of more than 1.4 million candidates.
  • Qualified for Round 2 of Amazon ML Summer School 2026.
  • Global rank 527 in LeetCode Weekly Contest 508 — all four problems solved.
  • Global rank 134 in CodeChef Starters 218.

Education

Indian Institute of Information Technology, Surat

B.Tech, Computer Science and Engineering

2024 – 2028

CGPA
8.05 / 10
JEE Main
98.2 percentile
Experience

Lead, Learn Code Solve

Aug 2026 — present

  • Design, review and validate algorithmic problems for contests and practice sets, against a structured quality standard.
  • Mentor students in DSA and competitive programming — problem-solving approach and contest strategy.