Mohar Das.

Production-grade LLM inference router in Go, sitting between apps and OpenAI-compatible/Anthropic backends. Adaptive routing (EWMA latency / weighted / priority), sliding-window circuit breaking with a zero-allocation hot path (4.8 ns/op), health-gated failover with backoff, multi-dimension token-bucket rate limiting, TTL response caching, SSE streaming that never hangs, config hot-reload preserving breaker state, and dependency-free Prometheus metrics — ~5k stdlib lines, one external dependency (yaml.v3).

Go / Circuit Breaking / Rate Limiting / Streaming / Prometheus

8-stage hybrid retrieval pipeline fusing dense HNSW, BM25 and knowledge-graph traversal with Reciprocal Rank Fusion + cross-encoder reranking. HotpotQA ablation: KG traversal lifted Exact Match 29% → 52%; full pipeline at 66% F1, 80% answer recall, 82% Hit@5, 90% answer-in-context.

Python / PostgreSQL/pgvector / KuzuDB / FastAPI / Docker

End-to-end pipeline: 3,350 instruction-tuning samples across 20 relation types (MinHash dedup, hard negatives, curriculum ordering) + Qwen3-0.6B fine-tune with 4-bit LoRA → 100% schema adherence, 0.6850 composite on multi-axis eval.

Qwen3 / LoRA / Unsloth / Hugging Face

From-scratch reimplementation of Logic-LM (Findings of EMNLP 2023), the neuro-symbolic framework that pairs LLMs with symbolic solvers for logical reasoning tasks.

Python / Neuro-symbolic / LLM Reasoning

Agentic research system based on bisociation. 9-agent LangGraph pipeline with qualifier-driven routing and automated critique loops; persistent workflow state, per-node error boundaries, and a metrics-first evaluation suite.

Python / LangGraph / FastAPI / Anthropic / Groq / Tavily

Rediscover physics equations from noisy data with a from-scratch Kolmogorov–Arnold Network (KAN).

Python / KAN / Symbolic Regression

Scripts to train PunyPunk and bestow great capabilities from larger models.

Python / Distillation / Fine-tuning