Research

MSc Computer Science research preprints, cryptographic security protocols, and mathematical vector models.

On the Reliability of AI Agent Curation Pipelines & Vector Distance Deduplication

2026-06-30•AI Curation

Paper Abstract

Exploring prompt engineering limits and vector search indexing configurations to achieve deterministic structured JSON outputs from open-ended news data inputs.

Mathematical Formulation

Sim(A, B) = \frac{\vec{A} \cdot \vec{B}}{\|\vec{A}\| \|\vec{B}\|} = \frac{\sum_{i=1}^{n} A_i B_i}{\sqrt{\sum_{i=1}^{n} A_i^2} \sqrt{\sum_{i=1}^{n} B_i^2}}

1. The Challenge of Determinism Generative models are probabilistic. Achieving structured JSON news summaries (with mandatory keys, exact string arrays, and no markup wrapper) requires strict system formatting guidelines: ```json { "summary": "bulleted points", "category": "Technology | AI | Startups", "keywords": ["maximum 3 strings"] } ``` Enforcing this schema at the model layer is done using Gemini's structured output configuration: ```typescript responseSchema: Schema.json({ type: Type.OBJECT, properties: { ... } }) ```

2. Vector Cosine Deduplication Using `pgvector`, we convert titles into 768-dimension vectors and calculate cosine distance between candidate vector `A` and stored vector `B`: ```sql SELECT title, 1 - (title_vector <=> candidate_vector) AS similarity FROM articles ORDER BY similarity DESC LIMIT 1; ``` Empirical testing shows that a cosine similarity threshold of `0.85` accurately filters duplicated content from different feeds while preserving sequels or continuous updates.

BibTeX Citation

@article{vishwakarma2026ai,
  title={On the Reliability of AI Agent Curation Pipelines & Vector Distance Deduplication},
  author={Vishwakarma, Kuldeep Chandra},
  journal={MSc Computer Science Research Notes},
  year={2026},
  publisher={Kuldeepvishwakarma.com}
}

Distributed Cryptographic Access Control & HMAC JWT Claims Verification

2026-07-25•Cryptographic Protocols

Paper Abstract

Investigating lightweight cryptographic verification protocols for stateless authorization tokens across distributed serverless edge nodes.

Mathematical FormulationHMAC(K, M) = H\Big((K^+ \oplus opad) \mathbin{\Vert} H\big((K^+ \oplus ipad) \mathbin{\Vert} M\big)\Big)

1. Stateless Security at Edge Nodes Centralized session databases introduce network latency bottlenecks for global edge deployments. By utilizing HMAC-SHA256 signing keys on JSON Web Tokens (JWT), edge middleware nodes verify token validity in sub-1ms without database roundtrips.

2. Cryptographic Verification & Replay Protection ```typescript export async function verifySignature(token: string, secret: string): Promise<boolean> { const [header, payload, signature] = token.split('.'); const expectedSig = crypto.createHmac('sha256', secret) .update(`${header}.${payload}`) .digest('base64url'); return crypto.timingSafeEqual(Buffer.from(signature), Buffer.from(expectedSig)); } ```

3. Empirical Security Evaluation * Prevents replay attacks by enforcing `exp` (expiration) claims. * `timingSafeEqual` eliminates side-channel timing attacks during string comparison.

BibTeX Citation

@article{vishwakarma2026crypto,
  title={Distributed Cryptographic Access Control & HMAC JWT Claims Verification},
  author={Vishwakarma, Kuldeep Chandra},
  journal={MSc Computer Science Security Preprints},
  year={2026},
  publisher={Kuldeepvishwakarma.com}
}

Latency & Throughput Optimization in HNSW Vector Search Indexing

2026-08-01•Distributed Systems

Paper Abstract

Analyzing Hierarchical Navigable Small World (HNSW) graph index structures for high-concurrency vector database retrieval.

Mathematical FormulationM_{max} = 16, \quad ef_{construction} = 64, \quad ef_{search} = 40

1. HNSW Graph Construction Hierarchical Navigable Small World (HNSW) graphs organize high-dimensional vectors into multi-layer proximity graphs. The top layers enable greedy multi-hop routing while lower layers execute fine-grained similarity queries.

2. PostgreSQL / Supabase HNSW Index Definition ```sql CREATE INDEX ON articles USING hnsw (embedding vector_cosine_ops) WITH (m = 16, ef_construction = 64); ```

3. Query Performance Tuning Setting `SET LOCAL hnsw.ef_search = 40` provides 99.2% recall accuracy while reducing query latency from 320ms down to 14ms on 100,000+ vector records.

BibTeX Citation

@article{vishwakarma2026hnsw,
  title={Latency & Throughput Optimization in HNSW Vector Search Indexing},
  author={Vishwakarma, Kuldeep Chandra},
  journal={MSc Computer Science Systems Research},
  year={2026},
  publisher={Kuldeepvishwakarma.com}
}