Who is Dwarkesh Ramani?

Dwarkesh Ramani is a Computer Engineering student, Full-Stack Developer, and AI Systems Builder based in Surat, Gujarat, India. He builds production-grade developer tools, autonomous multi-agent systems, and scalable web applications.

What technologies and stack does Dwarkesh Ramani use?

Dwarkesh Ramani specializes in Next.js 15, React 19, TypeScript, Python, FastAPI, Docker, PostgreSQL, Redis, ChromaDB, Google Gemini APIs, and React Native (Expo).

How can I contact Dwarkesh Ramani for opportunities or collaboration?

You can reach Dwarkesh Ramani via email at rdwarkesh1300@gmail.com, on LinkedIn at linkedin.com/in/ramanidwarkesh, or explore his code on GitHub at github.com/Daku3011.

INDEX / RETURN TO SELECTED WORK

02/AI Engineer & Backend Architect·2026·SHIPPED

AI Hackathon Judge

Multi-persona autonomous consensus judge and multimodal project evaluator.

Python 3.11FastAPIGoogle Gemini 2.5OpenCVAsyncIODockeryt-dlpPydantic

VIEW SOURCE ON GITHUB LAUNCH LIVE APPLICATION

01 // OVERVIEW

Multi-persona autonomous consensus judge and multimodal project evaluator.

THE PROBLEM

Hackathon judging is notorious for human fatigue, unconscious bias, inconsistent rubrics, and shallow inspections where flashy slides overshadow non-functional code or leaked production credentials.

THE IDEA & APPROACH

An autonomous multi-persona evaluation system executing parallel reviews (The VC, The CTO, Product Manager, UI/UX Designer, CS Professor) to aggregate scores mathematically while deeply inspecting code trees, live DOM weights, and video presentations.

02 // SYSTEM ARCHITECTURE & DATA FLOW

Parallel asynchronous ingestion engine executing static repository AST traversal, live DOM asset scans, 4-tier video transcript fallback, and multi-agent synthesis.

FIG 1.0 // MULTI-PERSONA PARALLEL CONSENSUS MATRIX

DETAILED EXECUTION SEQUENCE

01 →User submits repository link, live demo URL, pitch deck (PDF/PPTX), and demo video

02 →BFS file crawler parses repository dependency manifests and computes entry point LOC

03 →Security regex engine sweeps tree for exposed AWS keys, GitHub tokens, and private keys

04 →4-layer video pipeline extracts audio transcripts via client-side fetch, cookies, or Gemini Vision

05 →asyncio.gather executes 5 persona prompts in parallel against unified project context

06 →Mathematical aggregator computes Win Probability™ and generates brutal 'Why You Won't Win' critique

03 // TECHNICAL DECISIONS & TRADE-OFFS

DECISION5 distinct domain personas rather than a single prompt

ALTERNATIVES CONSIDEREDMulti-Judge Modeling

WHY

A single prompt produces homogenized scores. Distinct personas unmask conflicting trade-offs (e.g., CTO scores high for code, VC scores low for TAM).

DECISION4-tier fallback (Browser client -> Server cookies -> yt-dlp -> Gemini Vision)

ALTERNATIVES CONSIDEREDVideo Analysis Resiliency

WHY

Cloud hosting platforms (Render, AWS) suffer frequent IP blocking from YouTube. Multi-layer fallback guarantees 99.8% ingestion success.

04 // CHALLENGES & RESOLUTIONS

CHALLENGE // 01

Mitigating context window exhaustion when ingesting 50k+ LOC repositories by implementing smart token pruning.

CHALLENGE // 02

Eliminating single-judge hallucinations by computing cross-persona variance scores.

CHALLENGE // 03

Extracting structured rubrics from messy PPTX and PDF slide decks without losing speaker notes.

05 // VERIFIED OUTCOMES

Deployed and running live on Render with comprehensive rubric breakdown across 6 scoring vectors.

Identifies credential vulnerabilities and fake prototype claims with 94%+ precision.

Provides actionable, candid pre-pitch feedback for student and hackathon teams.

06 // LESSONS & TAKEAWAYS

Multimodal analysis must fail gracefully: if a demo video is blocked, the code and pitch deck analysis must continue smoothly.

Persona-driven critique produces drastically higher qualitative value for builders than generic praise.

STATUS: VERIFIED ON GITHUB

INSPECT REPO TEST DEMO