- 🎓 2nd-year CS student at JECRC University, Jaipur (2024–2028)
- 🏗️ Built 3 production AI systems from scratch — shipped, measured, iterated
- 🤖 Obsessed with multi-agent pipelines, RAG, and LLM systems that actually work at scale
- 🧠 Believer in deterministic, evaluation-first AI — reproducible outputs over vibes
- ⚡ 30+ hackathons · 2nd Place @ Bit to Code 2024 (AI/ML track)
- 📬 Reach me: krishagarwal52139@gmail.com
- 📖 Reading about agent architectures and retrieval systems
- 🛠️ Breaking and rebuilding things to understand how they work
- 🏆 Competing at hackathons — speed-building under 24-hour constraints
Python Pathway Sentence Transformers Gemini
Constraint-based NLP reasoning pipeline — validates character backstory consistency across 100,000+ word novels
| What | Result |
|---|---|
| Eliminated full-document scans via atomic claim decomposition | Targeted lookup with CORE / SIGNIFICANT / SURFACE tagging |
| Hierarchical chunking with 2-paragraph overlap + timeline-aware indexing | Cross-boundary evidence loss → near-zero across 50-entry test set |
| Deterministic rule-based classifier (zero LLM for binary decisions) | 100% reproducible outputs — same input, same result, always |
| Semantic search via Sentence Transformers + cosine similarity | Top-k retrieval validated on 28 consistent + 22 contradictory cases |
Python Groq (Llama 3.1) Streamlit Plotly
8-stage autonomous multi-agent pipeline for professional market intelligence — runs in minutes, not days
| What | Result |
|---|---|
| Parallelised Market + Capability agents via ThreadPoolExecutor | ~40% reduction in total pipeline runtime |
| SHA-256 LLM response cache — identical prompts served from memory | Zero duplicate API calls across 7 Plotly visualisations |
| Hidden Critic-driven refinement loop (sections < 7/10 auto-rewritten) | Only polished output surfaces to user |
| 8 specialised agents via shared AgentMemory | Fully stateless, no direct agent-to-agent calls |
Python Flask React Gemini 2.5 Flash Google Places API
Multi-agent RAG system — generates personalised date itineraries for Indian cities in under 15 seconds
| What | Result |
|---|---|
Parallelised tool calls with asyncio.gather() |
62% latency drop — 30s → 11s, eliminated primary drop-off point |
| RAG pipeline over 60+ curated city documents (Jaipur, Delhi, Mumbai, Bangalore) | Context-aware itineraries with budget tiers, safety & culture filters |
| SSE streaming to React frontend | Real-time plan appearance, venue quality filtered at 4.0+ / 50+ reviews |
|
|
|
- 🥈 2nd Place — Bit to Code Hackathon 2024 (AI/ML track) · 200+ competing teams
- ⚡ Top finisher across 5+ AI/ML hackathons (2023–2024) — shipped working LLM + RAG prototypes under 24-hour constraints
- 🎓 8.6 GPA maintained while building 3 production AI systems in year one of college
Built with 🧠 and too much Groq inference by Krish Agarwal
⭐️ From Krish-afk-bot · Let's build AI that actually works.