AI Engineer from Pithoragarh, India — building LLM systems that actually work.
LinkedIn · kunalbisht909@gmail.com · JLPT N5 🇯🇵
"LLMs are great at extraction. They're terrible at cultural nuance."
I built a meeting analysis system and hit a wall at 22% accuracy. The problem wasn't the LLM — it was that no amount of prompting teaches a model what nemawashi means in a real boardroom.
So I went and learned it myself. Two months of Japanese business linguistics, 16 soft-rejection patterns, MeCab morphological analysis for keigo detection. Then wired it as a post-processing NLP layer on top of the LLM output.
93% accuracy. 5 iterations. Every drop in accuracy had a specific root cause and a specific fix.
What's actually interesting here:
- Hallucination guard is rule-based token overlap — the LLM never grades its own output
- PII masking runs locally before the transcript reaches any API. Zero data leaves the machine unmasked.
- Same person written as 田中, Tanaka, and "the Director" resolves to one speaker identity
- 3-tier fallback: Groq → Ollama → Mock. Explicit UX signal at each tier, not silent degradation
github.com/aiKunalBisht/Transcript-ai · Live on HuggingFace
FastAPI Groq MeCab FAISS Streamlit GitHub Actions 21 tests passing
Built to answer a specific question: how bad does RAG actually get at scale?
- 241k Wikipedia articles → ~39k semantic chunks with FAISS
- Sub-50ms vector search on CPU
- 100% faithfulness on DeepEval test set
- FastAPI endpoint
github.com/aiKunalBisht/rag-wikipedia
Backend engineering on a live college review platform.
The interesting parts: brought response time from ~300ms → ~85ms through query profiling and indexing, not rewrites. JWT RBAC across user roles. Monolith → microservices migration that actually reduced deploy complexity instead of adding it.
Node.js MongoDB TypeScript Next.js
End-to-end: raw data → feature engineering → XGBoost → prediction API → 4 Plotly dashboards → Gunicorn + Nginx. First time I owned the full MLOps cycle.
github.com/aiKunalBisht/TransitIntel
AI/ML — Python, scikit-learn, PyTorch, Pandas, NumPy, LangChain, FAISS, Groq, Whisper, MeCab, Sentence-Transformers, RAG, LLM eval (DeepEval, RAGAS), Prompt Engineering
Backend — FastAPI, Node.js, Express, Flask, PostgreSQL, MongoDB, Redis, Docker, JWT
Frontend — React, Next.js, Tailwind CSS, TypeScript
Master's in Data Science & Business Analytics — Simplilearn x IBM
B.Tech Computer Science & Engineering
Statistics · Probability · Model training and evaluation · Data analytics
Open to AI/ML Engineer and GenAI Developer roles — India and Japanese MNCs.