Available: Internship now · Full-time Dec 2026 · Open to relocate / remote

Shubham Bhattacharya

$ Software Engineer — Backend (TypeScript) · GenAI & RAG|

Engineering grounded RAG systems, agentic LLM backends, and evaluation harnesses with TypeScript, pgvector, and multi-model failover.

Live ContractLens Demo Live RAG Demo GitHub
01 / About

Engineering Background

Shubham Bhattacharya

BCA student, graduating Dec 2026 at Parul University (CGPA 8.13) specializing in GenAI and LLM application engineering with TypeScript, Node.js, Express, Fastify, PostgreSQL, and Redis. I build resilient AI architectures guided by a strict design law: "The LLM proposes; deterministic code disposes." I benchmark retrieval accuracy, deterministic quote verifiers, and failover behavior with 560+ automated tests. Based in Vadodara, Gujarat.

Open to Software Engineer — GenAI / LLM Applications (TypeScript) roles where I can ship high-impact, resilient AI infrastructure. Engineering philosophy: zero-trust schema validation, deterministic grounding verification, structured telemetry, and graceful failure handling.

Deterministic Quote Verifier
Pure code matches LLM citations against canonical document byte offsets; invented or paraphrased quotes are 100% rejected.
Multi-Model Resilience
Automatic waterfall failover across LLM providers (Groq, Gemini) keeps sessions alive through provider outages.
Hybrid Retrieval (BM25 + Vector)
BM25 + Gemini embeddings fused with RRF (Reciprocal Rank Fusion) and strict evidence gating thresholds.
Sandboxed ReAct Agents
Custom agent loops with server-side trusted tool execution, strict round caps, and Langfuse tracing.
02 / Architecture

GenAI & Backend Architecture Patterns

Production patterns I use to build reliable, observable LLM applications in TypeScript.

The LLM Proposes, Deterministic Code Disposes

Model output is untrusted end to end. Every quote is verified against canonical document text by pure deterministic code before display — paraphrased or fabricated quotes are rejected.

Deterministic VerificationZero HallucinationPure Math

Multi-Model Waterfall Routing

Resilient orchestration across Groq and Gemini. If primary models hit rate limits or provider outages, automatic failover routes traffic to secondary models.

GeminiGroqOpenRouterFailover

Sandboxed ReAct Agent Execution

Custom ReAct research loop with server-side trusted tool executor injecting session/document scopes to prevent jailbreaks, with a strict 4-round safety cap.

ReAct LoopTrusted Executor4-Round Cap

Hybrid Retrieval & Evidence Gating

BM25 lexical search combined with dense vector embeddings fused via Reciprocal Rank Fusion (RRF). Gated by strict evidence thresholds: weak signals escalate to agentic research or refusal.

BM25Gemini VectorsRRF Fusion

Automated Evals & Test Invariants

560+ automated tests across systems covering chunker invariants, quote verification, refusal gates, session isolation, and rate-limiting window semantics in CI.

560+ TestsCI/CDGolden-Set Evals

Telemetry & Rate Limiting

Redis-backed fixed-window rate limiter on LLM endpoints failing open + structured Langfuse and Pino tracing for complete latency breakdown.

Redis LimitsLangfusePino
03 / Projects

Featured Work

Production GenAI & distributed systems I've architected and shipped solo.

ContractLens Agentic Legal Contract Analyser interface
Oct 2026 · Personal Project (Solo) · Flagship

ContractLens — Agentic Legal Contract Analyser

  • The LLM proposes; deterministic code disposes: Pure whitespace-tolerant verification matching model quotes against canonical document byte offsets and page numbers. Every displayed quote is verified before render — unverified text cannot reach the UI.
  • Interactive citation highlighting: Clicking any citation instantly scrolls the viewer to and highlights the exact verified passage (page-correct, cross-page capable).
  • Agentic research loop: ReAct loop with a hard 4-round cap and trusted server-side tool executor (search_document, get_section, list_clauses) injecting session/document scopes to prevent jailbreaks.
  • Clause comparison & numeric severity: Clause-level alignment flagging numeric changes (e.g. liability cap moving $100,000 → $1,000,000 is CRITICAL regardless of identical phrasing); handles up to 5 documents simultaneously.
  • 318 automated tests: 301 backend unit/integration + 17 frontend tests validating chunker invariants, quote-verification guarantees, and rate limits with Langfuse tracing.
Next.js 14TypeScript (strict)Express 5RedisBM25GeminiGroqLangfuseTailwindSSE
Source Code Live Demo
RagChatBot chat interface preview
Jul–Sep 2026 · Personal Project (Solo)

RAG Chat — Grounded Document Intelligence

  • Refusal over hallucination: Cosine-gated hybrid retrieval (distance threshold ≤ 0.5) — ungrounded evidence triggers automatic refusal instead of a hallucinated answer.
  • 14/14 golden-set eval passed (Sep 3, 2026 committed run): Automated harness testing answer grounding against adversarial trap prompts (knowledge in parametric memory but absent in the corpus); every run is committed as JSON in-repo — dated and auditable, not a standing guarantee.
  • Async job queue & SSE streaming: Offloaded 10–30s document parsing & vector embedding generation via BullMQ + Redis, streaming real-time tokens over SSE.
BunFastifyTypeScriptpgvectorGeminiGroqSSEBullMQRedis
Source Code Live Demo
AI Ops project preview
May–Aug 2026 · Personal Project (Solo)

AI Ops — AI Customer Support Backend

  • Custom ReAct agent loop (no LangChain): Autonomous customer support agent orchestrating multi-step order tracking, ticket routing, and refund draft workflows via auditable decision steps.
  • 100% Zod-validated tool registry: Strict runtime schema verification on every LLM parameter before executing Prisma ORM writes to PostgreSQL, eliminating hallucinated function calls and malformed parameters.
  • Resilient model failover: Automated Llama 70B → 8B routing through provider rate-limits and outages, with structured Pino and Sentry error telemetry.
Node.jsTypeScriptExpressGroqPostgreSQLPrismaZodSentry
Source Code Live Demo
04 / Stack

Technical Skills & Technologies

Technologies I use to build and ship GenAI applications.

Languages

TypeScript JavaScript (Node.js, Bun) SQL

Backend

Express.js Fastify REST API Design SSE Streaming

Databases & Caching

PostgreSQL pgvector Prisma ORM Redis (ioredis & Upstash REST) BullMQ Vercel Blob

AI & LLMs

Deterministic Quote Verification Hybrid Search (BM25 + Vector + RRF) ReAct Agents (Sandboxed Tools) Evals & Golden Sets (560+ tests) Multi-provider failover (Groq/Gemini) SSE Streaming Zod Runtime Validation

DevOps & Observability

Langfuse Tracing GitHub Actions CI/CD Docker Compose Vercel & Render Sentry Pino Structured Logging

Fundamentals (BCA)

Data Structures & Algorithms DBMS Operating Systems Computer Networks
05 / Education

Education

Bachelor of Computer Applications (BCA)

Expected Dec 2026

Parul University · Vadodara, Gujarat

  • CGPA: 8.13 | First Class with Distinction · Relevant: DBMS, OS, Networks, Software Engineering
06 / Contact

Get In Touch

Open to Software Engineer — GenAI / LLM Applications (TypeScript) roles. Feel free to reach out through any channel below.

Available: Internship now · Full-time Dec 2026 · Open to relocate / remote

Email shubhambhattacharya107@gmail.com Phone +91-9155252394 GitHub github.com/shubhambhattacharya-dev LinkedIn linkedin.com/in/shubhambhattadev