Kartikeya Mandhar

Blindspot

Per-room LangGraph game engine with fail-closed AI players

Server-authoritative LangGraph state machine with Postgres checkpointing. AI players whose clues must clear a 4-channel fail-closed audit, gated by a zero-leak eval suite and 117 tests in CI.

10-agent Claude pipeline for strategy-case decomposition

Turns an industry, company, and strategic question into a MECE hypothesis tree, a prioritized analysis plan, a red-team report, and a PowerPoint workplan, grounded in live SEC EDGAR and Yahoo Finance data.

Multi-agent dispatch with self-correction and human-in-the-loop

An 11-node LangGraph pipeline over 7 agents with a cyclic audit self-correction loop, interrupt-based human approval, and deterministic Python guardrails, with about 140 offline mocked-LLM tests.

The Engine That Would Rather Stay Silent Than Lie

Grounded SEC-filings RAG with in-code citation enforcement and a refusal gate

Answers questions on SEC filings and cites the exact source span, refuses when evidence is missing, pulls figures from structured XBRL rather than scraping HTML, and blocks any release that fails a golden-set quality gate.

Faithfulness enforced as code, not prompt hope.

Reranking lab: BM25 vs BERT cross-encoder vs LLM strategies

Benchmarks BM25, a BERT cross-encoder, and three GPT-4o-mini reranking strategies on SciFact and TREC-DL with MRR@10 and NDCG@10 via pytrec_eval, plus latency and cost tracking and a live React and D3 dashboard.

Small query counts shown honestly. Cost column is an estimate.

GraphRAG schema-enforcement study on clean vs noisy corpora

A 1,200-query comparison across two GraphRAG frameworks and two corpora where schema-free GraphRAG scored 90.0% on clean Wikipedia and 79.5% on noisy SEC filings, beating ontology-guided and vector RAG.

Counterintuitive result: ontology-guided degraded more on noisy data.

Neo4j GraphRAG example, merged into Apache Hamilton

A 4-module GraphRAG pipeline (ingest, embed, retrieve, generate) contributed to Apache Hamilton and merged upstream as PR #1532.

Full-stack knowledge-graph RAG for technical manuals

Turns PDF manuals into a Neo4j knowledge graph and answers questions with intent-routed hybrid retrieval: LLM query expansion, LLM re-ranking, text-to-Cypher with a document-scope guard, and a web-search fallback.

Hybrid LLM + vector-search recommender

End-to-end recommender pairing GPT-4o-mini query parsing with Pinecone semantic search and a constraint-boost re-ranker, scoring 0.85 mean NDCG on the Starbucks challenge's 100 training queries.

Training-set self-computed NDCG, not a leaderboard score.

Multimodal knowledge-graph builder with hybrid recommendations

Turns text, PDF, image, and audio into a live NetworkX knowledge graph and returns explainable recommendations by blending graph BFS paths with FAISS vector similarity, covered by 36 mocked tests.

The Alibi That Checked Out

LedgerBench: are analytics agents business-correct, not just runnable?

An open-source benchmark scoring analytics agents on adversarial traps compiled from a semantic layer. It found that more business context raised average accuracy but worsened the tail, with more double counting and fewer refusals.

sliceval: a published library for slice-based model evaluation

Pip-installable library that finds underperforming data subgroups with bootstrap confidence intervals and permutation significance tests, backed by 162 tests spanning 13 sklearn model types and 3 task types.

KV cache mechanics: latency, memory, and GQA vs MHA

Single-GPU benchmarks of transformer decode caching showing GQA cuts the per-token KV cache 7x and cache-on decode stays flat while cache-off rises about 1.6x over 32 to 1024 tokens.

Connected to agent prefix-caching economics.

Fact-Faithful Fine-Tuned Commentator

Fine-tuned 1.5B commentator engineered so it cannot state a wrong fact

Solves hallucination in small language models: an end-to-end fine-tuning and serving architecture (Qwen2.5-1.5B, QLoRA) that separates deterministic facts from generated voice, with validate-or-fallback serving, deployed live on FastAPI.

Causal RDD study of Steam review labels on player demand

A reproducible regression-discontinuity pipeline on 1,282 real Steam games estimating the causal demand effect of review-summary labels, finding an adjusted peak-CCU jump of +0.96 (p=0.03) at the 40% cutoff with a 70% falsification check.

Industry-standard rdrobust + rddensity, placebos, donut RDD.

RL dynamic pricing (PPO) on 693K real ride records

Trains a PPO surge-pricing agent over 693,071 real ride records with a FastAPI backend and web demo.