# Shivam Patil — Software Engineer > Full text of https://www.shivamsfolio.com, as the site's own retrieval index holds it. > Generated from the same corpus that serves https://www.shivamsfolio.com/ask, so this and the site cannot disagree. ## Low-Latency Market Data & Order Entry Stack Source: https://www.shivamsfolio.com/projects/low-latency-market-data-order-entry ### Summary Low-Latency Market Data & Order Entry Stack. Price-time-priority matching engine publishing a binary feed over redundant A/B UDP multicast, with a Rust feed handler that arbitrates the two and rebuilds MBP/MBO books without allocating. Outcome: 1M+ msg/s · ~100ns decode. Built with Rust, C++, FIX 4.4, UDP multicast, SBE-style encoding, Docker. Topics: Rust, C++, FIX 4.4, UDP multicast, SBE-style encoding, Docker. ### Use case Consume an exchange feed and place orders against it without the decode path or the book update ever touching the heap — and recover cleanly when packets are lost rather than resynchronising by restart. ### What was built Matching engine and binary feed: Price-time-priority matching publishing a binary market-data feed over redundant A/B UDP multicast channels, with configurable packet-loss injection, a 2-second snapshot cycle, and a TCP replay service for recovery. Allocation-free feed handler: A/B feed arbitration, sequence-gap detection and snapshot-based recovery into MBP/MBO order books — sustaining 1M+ messages/sec at ~100ns decode and ~200ns book update, with zero heap allocations per message verified by a counting allocator. FIX 4.4 order gateway: A full session layer — logon, heartbeats, resend/gap-fill, durable sequence persistence — reconciling order state across a hard process restart, plus a risk service enforcing pre-trade limits on an allocation-free path. ### Architecture FIX 4.4 gateway: session · resend · gap-fill Risk service: pre-trade limits Matching engine: price-time priority Snapshot + replay: 2s cycle · TCP recovery Rust feed handler: A/B arbitration · MBP/MBO ## Distributed Rate Limiter & API Gateway Source: https://www.shivamsfolio.com/projects/distributed-rate-limiter-api-gateway ### Summary Distributed Rate Limiter & API Gateway. Multi-tenant API gateway: token-bucket and sliding-window quotas across a consistent-hash shard ring, with Raft election covering shard failure. Outcome: 45K req/s · <8ms p99. Built with Go, gRPC / REST, Redis, PostgreSQL, AKS, Prometheus, React. Topics: Go, gRPC / REST, Redis, PostgreSQL, AKS, Prometheus, React. ### Use case Keep per-tenant quotas accurate across replicas and regions while a noisy neighbour, a lost shard or a region blip is in progress. ### What was built Dual limiting strategies: Token-bucket for smooth burst control and sliding-window counters for stricter endpoint policies. Resilient control plane: Consistent-hash sharding keeps a tenant on a stable limiter shard; Raft election promotes a replacement leader on failure. Observable delivery: gRPC and REST APIs expose decisions and quota status, while Prometheus metrics feed Grafana and a live React dashboard. ### Architecture Client traffic: REST / gRPC Gateway replicas: auth · routing · policy Limiter shard ring: hashing · Raft Redis + Postgres: counters · policies Grafana + React: metrics · quotas ## Agentic AI Orchestration Platform Source: https://www.shivamsfolio.com/projects/agentic-ai-orchestration-platform ### Summary Agentic AI Orchestration Platform. Multi-agent runtime on a plan-retrieve-act-critique loop, emitting a full decision and tool-call trace on every run. Outcome: 87% success · 150 eval cases. Built with Python, FastAPI, Tool calling, Pinecone / Weaviate, PostgreSQL, WebSockets, React. Topics: Python, FastAPI, Tool calling, Pinecone / Weaviate, PostgreSQL, WebSockets, React. ### Use case Make knowledge-work requests repeatable and auditable — the reasoning trace and tool activity stay visible rather than collapsing into an answer. ### What was built Specialised agent loop: Planner, retriever, executor, and critic agents share task state and take turns through explicit tool-calling contracts. Safe recovery: A sandboxed execution loop captures failed actions, lets the critic propose corrections, and bounds retries with policy checks. Governance surface: WebSocket events stream live traces to React; PostgreSQL persists inputs, tools, outputs, and checkpoints for replay. ### Architecture Operator: React workspace FastAPI runtime: tasks · WebSockets Agent team: plan · retrieve · act · critique Knowledge + audit: Vector DB · Postgres Evaluation harness: quality · replay ## High-Performance LLM Inference Server Source: https://www.shivamsfolio.com/projects/high-performance-llm-inference-server ### Summary High-Performance LLM Inference Server. Rust inference runtime built on continuous batching and explicit KV-cache management, with live throughput and tail-latency signals. Outcome: +230% throughput · −42% p99. Built with Rust, Python / PyO3, gRPC, Continuous batching, INT8 / FP16, Docker, React. Topics: Rust, Python / PyO3, gRPC, Continuous batching, INT8 / FP16, Docker, React. ### Use case Serve concurrent LLM requests without the throughput collapse and p99 spikes that single-request inference hits under load. ### What was built Continuous scheduler: A Rust scheduler admits compatible requests at decode boundaries instead of waiting for a full batch to finish. Memory-aware serving: KV-cache slots are reserved, reused, and released predictably; model loading supports INT8 and FP16 variants. Inspectable performance: gRPC exposes generation controls, while React visualises token rate, batch fill, and memory pressure in real time. ### Architecture Apps + dashboard: gRPC · React Inference gateway: routing · streaming Rust scheduler: continuous batch Model + KV cache: INT8 / FP16 Telemetry: tokens · memory · batches ## Secure Banking System Source: https://www.shivamsfolio.com/projects/secure-banking-system ### Summary Secure Banking System. Banking platform on a Hyperledger Fabric ledger, with Kafka-decoupled transaction events and Vault-managed secrets behind OAuth2 Django APIs. Outcome: Fabric ledger · Kafka · Vault. Built with Python, Django, Hyperledger Fabric, Web3.py, Kafka, Vault, Kubernetes, Grafana. Topics: Python, Django, Hyperledger Fabric, Web3.py, Kafka, Vault, Kubernetes, Grafana. ### Use case Process financial operations against a verifiable ledger, with asynchronous enrichment that survives a downstream outage. ### What was built Ledger-backed transactions: Hyperledger Fabric records authorised state changes while Django APIs coordinate application-level workflows. Event-driven processing: Kafka decouples high-volume transaction events from downstream enrichment and notification services. Defense in depth: OAuth2, Redis caching, Vault-managed secrets, container deployment, and Prometheus/Grafana support security and operations. ### Architecture Banking clients: web · mobile Django API: OAuth2 · policy Fabric network: endorsers · ledger Kafka + Postgres: events · operations Vault + Grafana: secrets · observability ## Online Examination System Source: https://www.shivamsfolio.com/projects/online-examination-system ### Summary Online Examination System. Assessment platform: NGINX across FastAPI workers, Redis-held session and scoring state, JWT-scoped endpoints, server-side validation throughout. Outcome: NGINX → FastAPI · Redis · JWT. Built with Python, FastAPI, PostgreSQL, Redis, Docker, NGINX, AWS, JWT. Topics: Python, FastAPI, PostgreSQL, Redis, Docker, NGINX, AWS, JWT. ### Use case Hold a consistent assessment session when hundreds of candidates start, autosave and submit inside the same few seconds. ### What was built Fast, balanced API layer: NGINX routes requests across FastAPI workers and Redis keeps hot session and scoring state close to the application. Secure exam sessions: JWT-protected endpoints, encrypted data handling, access controls, and server-side validation protect candidate interactions. Fault-tolerant scoring: Submission events are safely persisted and a resilient scoring path uses caching to keep results responsive under load. ### Architecture Candidates: exam workspace NGINX edge: TLS · load balancing FastAPI workers: auth · answers · score Redis + Postgres: sessions · durable data Scoring worker: retries · outcomes ## Secure RAG with RBAC, Guardrails & Monitoring Source: https://www.shivamsfolio.com/projects/secure-rag-with-rbac-guardrails-monitoring ### Summary Secure RAG with RBAC, Guardrails & Monitoring. Enterprise RAG with RBAC metadata pushed into the vector-search filter, PII masking before generation, and Ragas scoring on every change. Outcome: RBAC-filtered retrieval · Ragas. Built with Python, Qdrant / Milvus, Streamlit, Ragas, Docling, AWS, RBAC, LLMs. Topics: Python, Qdrant / Milvus, Streamlit, Ragas, Docling, AWS, RBAC, LLMs. ### Use case Answer from private corpora while guaranteeing a caller can retrieve only what their role permits — enforced at the retrieval filter, not the prompt. ### What was built Permission-aware retrieval: RBAC metadata is applied to the retrieval filter so vector search respects the caller's organisational scope. Ingestion and protection: Docling prepares source documents; guardrails mask PII and detect out-of-scope or unsafe requests before generation. Quality feedback loop: Ragas-based monitoring evaluates retrieval relevance and response quality so regressions are visible and actionable. ### Architecture Knowledge user: Streamlit workspace RBAC + guardrails: policy · PII checks RAG service: retrieve · ground · answer Vector knowledge: Docling · Qdrant / Milvus Ragas monitor: quality signals ## AI/ML & Generative AI skills Source: https://www.shivamsfolio.com/skills ### AI/ML & Generative AI Technologies and tools Shivam Patil works with. AI/ML & Generative AI skills in his stack: AI Agents, Prompt Engineering, RAG (Retrieval-Augmented Generation), Vector Databases. ### Backend Technologies and tools Shivam Patil works with. Backend skills in his stack: FastAPI, REST APIs, Microservices, Distributed Systems. ### Cloud & DevOps Technologies and tools Shivam Patil works with. Cloud & DevOps skills in his stack: Azure Kubernetes Service (AKS), Docker, Kubernetes, Prometheus, Grafana. ### Core CS Technologies and tools Shivam Patil works with. Core CS skills in his stack: System Programming, Multithreading, Concurrency. ### Frontend Technologies and tools Shivam Patil works with. Frontend skills in his stack: React.js, Next.js. ### Languages Technologies and tools Shivam Patil works with. Languages skills in his stack: C++, Python, Go, Rust, JavaScript, TypeScript. ## Systems consultation Source: https://www.shivamsfolio.com/services ### Systems consultation Hire Shivam Patil for consulting: Systems consultation. A focused hour on one hard problem — throughput ceilings, queue design, caching strategy, or a failure mode you can't reproduce. You get the whiteboard session plus a written summary of the decisions and trade-offs. A paid engagement, booked through this site, running 60 minutes. ### Architecture review Hire Shivam Patil for consulting: Architecture review. A deep read of an existing design or codebase, followed by 90 minutes together. Delivered as a written review: what will break first, what is over-built, and the three changes with the best return. A paid engagement, booked through this site, running 90 minutes. ### AI systems deep-dive Hire Shivam Patil for consulting: AI systems deep-dive. For RAG pipelines and agent runtimes: retrieval quality, evaluation harnesses, guardrails, and cost control. Two hours, plus an eval plan you can actually run against your own data. A paid engagement, booked through this site, running 120 minutes. ## Software Development Engineer (SDE), Tata Consultancy Services Source: https://www.shivamsfolio.com/experience ### Tata Consultancy Services Where Shivam Patil has worked. Software Development Engineer (SDE) at Tata Consultancy Services, Jan 2025 — Present, India. Shipped 13+ production REST APIs behind a system serving 10,000+ requests/day. Response latency down 33%. Rewrote hot-path SQL and data-processing pipelines: +29% application performance. Raised RAG answer accuracy 19% on a production RBI chatbot — vector-search relevance tuning plus an automated evaluation harness. Built authentication and role-based authorization against enterprise access-control standards. Reviewed peer changes across the team's services, gating on testability and maintainability. ### Education Bachelor of Technology, Artificial Intelligence and Data Science, Pune University, July 2020 — June 2024. GPA: 8.88 / 10 ## Achievements Source: https://www.shivamsfolio.com/achievements ### Achievements 45K req/s — sustained throughput at sub-8ms p99 latency (Distributed Rate Limiter & API Gateway). 87% — agent task success across 150 evaluation cases (Agentic AI Orchestration Platform). +230% — inference throughput gain from continuous batching (High-Performance LLM Inference Server). −42% — p99 latency cut versus naive single-request serving (High-Performance LLM Inference Server). 13+ — production REST APIs shipped, cutting response latency 33% (Tata Consultancy Services). +29% — application performance gain from query & pipeline optimization (Tata Consultancy Services). +19% — RAG answer-accuracy improvement for a production chatbot (Tata Consultancy Services). ## Certifications Source: https://www.shivamsfolio.com/certifications ### Certifications Microsoft Certified: Azure Data Scientist Associate from Microsoft, 2026 (in progress). AWS Certified Developer – Associate from Amazon Web Services, 2026 (in progress). ## Contact Source: https://www.shivamsfolio.com/contact ### Contact How to get in touch with Shivam Patil. Email shivampatilinfo@gmail.com for anything — work, questions, or a conversation about a role. There is also a phone number and WhatsApp for anything time-sensitive. The contact form on this page reaches the same inbox and is the fastest route; a message written offline is queued in the browser and sent when the connection returns. ## Reach out Source: https://www.shivamsfolio.com/reach-out ### Reach out Every channel Shivam Patil can be reached on, in one place: email, LinkedIn, GitHub, and a map of where he is based. Use this page rather than the contact form when you would rather message on a platform you already use. ## About Source: https://www.shivamsfolio.com/about ### About Who Shivam Patil is: a software engineer working on backend systems, distributed infrastructure and developer tooling. This page covers his background, how he approaches building software, and what he is looking for in the work he takes on. ## Résumé Source: https://www.shivamsfolio.com/resume ### Résumé The résumé, as a PDF you can read in the browser or download. It covers work history, education, the technologies used in each role, and the outcomes each piece of work produced — the same material as the experience page, in the format a recruiter or hiring manager expects. ## Projects Source: https://www.shivamsfolio.com/projects ### Projects The index of every project written up on this site, each with the problem it solved, the architecture behind it, and what the outcome was. Filterable by the technologies and topics involved. ## Ask this site Source: https://www.shivamsfolio.com/ask ### Ask this site Ask a question about this work in plain English and get an answer drawn from the site's own content. Search here is hybrid retrieval: a vector index over embedded passages and a full-text lexical index, fused with reciprocal rank fusion so a result found by both ranks above one found by either alone. The answer is extractive — sentences are selected from the retrieved passages and scored against the question, so nothing is paraphrased and nothing can be invented. ## Compute lab Source: https://www.shivamsfolio.com/compute ### Compute lab Real computation running in your browser: a numerical kernel compiled to WebAssembly, the same work spread across Web Workers with SharedArrayBuffer, and a WebGPU implementation where the hardware allows it. Each backend is benchmarked live against the others, and the benchmark argues against itself where the result is misleading. ## Data pipeline Source: https://www.shivamsfolio.com/data ### Data pipeline Where a page view goes after it is recorded: through the ingest, into two ledgers and two rolled-up aggregates, out to a warehouse export, and into a DuckDB instance compiled to WebAssembly running in your browser, where you can write your own SQL against it. Includes a validated lineage DAG and tumbling-window aggregation. ## System design Source: https://www.shivamsfolio.com/system-design ### System design How this site is actually built, and why each decision was made: the caching strategy, the write path through a command boundary with idempotency keys, the event outbox and its dead-letter queue, and the trade-offs accepted at each layer. ## Reliability Source: https://www.shivamsfolio.com/reliability ### Reliability The site's own RED metrics — rate, errors and duration — measured by the code that serves each request and aggregated in shared state so the numbers survive running on serverless functions. Also the circuit breakers guarding the GitHub and LeetCode integrations, distributed tracing, and a chaos experiment that drives the production breaker through a simulated outage in your browser. ## Security Source: https://www.shivamsfolio.com/security ### Security A tamper-evident audit ledger that recomputes its own hash chain on every page load, envelope encryption where rotating a key never touches the encrypted payload, a committed software bill of materials, a static analysis ruleset, and a SAML validator demonstrating the signature-wrapping attack that defeats an otherwise valid signature. ## Edge and offline Source: https://www.shivamsfolio.com/edge ### Edge and offline Request filtering that runs before the application does, sharing one ruleset between the live edge proxy and a deployable Cloudflare Worker. Also an offline-capable progressive web app: a cached shell, and a queue that holds a message written without signal and sends it when the connection returns. ## API lab Source: https://www.shivamsfolio.com/api-lab ### API lab The same data served four ways: a hand-written OpenAPI 3.1 specification, a GraphQL subgraph with depth and complexity limits, protocol buffers encoded and decoded by hand on the wire, and a console that builds itself from the specification so the two cannot drift apart. ## Retrieval quality Source: https://www.shivamsfolio.com/mlops ### Retrieval quality How well search on this site actually works, measured rather than asserted: recall, mean reciprocal rank and normalised discounted cumulative gain over a labelled set of questions, run live against the index. Also index drift detection, and what generating an answer would cost on a quantized model running on your device versus a hosted one. ## Terminal Source: https://www.shivamsfolio.com/terminal ### Terminal A working command-line interface to this site. Navigate, search, ask questions and change the theme without touching the mouse; the same command registry backs the Cmd+K palette.