Backend, distributed systems,
and agentic AI in production.

• 01
  Iolite SoftwaresSoftware Developer

  Jan 2025 — Present

  Backend, AI infra, and platform work on a multi-tenant SaaS — primarily a LangGraph-based chatbot, the C#/ASP.NET API behind it, and the trademark-automation product line.

  • 01

    Agentic NL→SQL chatbot

    Architected a multi-tenant agentic NL→SQL chatbot on LangGraph with locally-served Gemma. Built a 3-tier follow-up engine, a destructive-SQL gate, and a binary-encoded RBAC whitelist across 50+ modules.

    ~60% drop in follow-up LLM calls
    why / how

    Picked LangGraph over plain LangChain so state, retries, and conditional routing live inside one explicit state graph rather than chained prompts — easier to reason about, easier to debug. Gemma runs on-prem because tenant schemas travel through the agent and we couldn't ship that metadata to a public API. The 3-tier follow-up engine resolves the easy half of repeat questions from a hashed-prompt + last-execution cache (L1), kicks the next slice to a tiny rewrite agent that prompts only with the diff against the previous turn (L2), and only falls through to a full re-plan (L3) when the question genuinely shifts intent. Authorization sits in front of generation, not behind it: the RBAC whitelist is encoded as packed bitmasks per (role × table × column), so checking access on a generated query is one AND, not a join.

    LangGraphGemmaPythonRBACPostgres
  • 02

    FastSQLDatabase + RAG retrieval

    Built a custom FastSQLDatabase wrapper with cached INFORMATION_SCHEMA snapshots and multi-tier TTL+LRU caches. Added RAG few-shot retrieval with hybrid embedding + lexical fallback for tenant priming.

    ~70% faster tenant cold-start
    why / how

    Each new tenant was paying a multi-second schema-introspection penalty on cold queries; pre-computing a versioned schema snapshot per tenant and layering an in-process LRU on top of a cross-process TTL store collapsed that to roughly a quarter of the original. Retrieval is hybrid on purpose: dense embeddings catch paraphrase and intent, BM25 catches the exact column names and identifiers that embeddings smooth over. Few-shot beat fine-tuning here because tenant schemas evolve weekly and re-tuning is operationally wrong for that rate of change.

    FastAPIVector DBHybrid retrievalTTL/LRU
  • 03

    Hybrid distributed caching

    Built ETag-based hybrid HTTP caching layered with IMemoryCache and Microsoft Garnet on the ASP.NET / C# / Angular 17 stack.

    ~70% DB-load reduction · ~60% faster responses
    why / how

    Three layers, each doing one job. Weak ETags at the edge let Angular skip large list payloads on revalidation — the cheapest cache hit is the one that never crosses the network. IMemoryCache short-circuits hot reads inside each app instance for sub-millisecond wins on N+1-style hotspots. Garnet (Microsoft's Redis-protocol drop-in, lower latency than Redis on our access pattern) is the shared L2 across the cluster so a cold instance still gets warm data. Invalidation is event-driven over a thin pub/sub bus rather than TTL-only, which keeps stale reads bounded after writes.

    ASP.NETC#GarnetETagAngular
  • 04

    Trademark automation at scale

    Shipped scrapers for 200+ trademark offices with custom CAPTCHA-solving models and PDF/image+text similarity pipelines (ResNet, CLIP, RapidFuzz) running across 250+ regional journals.

    200+ jurisdictions · 250+ journals
    why / how

    Every jurisdiction is its own scraping problem — sessions, JS-rendered pages, IP rate limits, and a different CAPTCHA per office. Commercial CAPTCHA APIs were uneconomic at this volume so we trained per-style solvers offline. Conflict ranking blends three signals rather than picking one: RapidFuzz on normalised marks for textual proximity, ResNet feature distance on logo crops for visual similarity, and CLIP for the cross-modal cases (text mark vs logo, or vice versa). Scoring is calibrated per jurisdiction because filing standards and similarity tolerances genuinely differ.

    ScrapersCVCLIPResNetRapidFuzz
• 02
  TTFA AcademySoftware Developer (part-time, during B.Tech)

  Jan 2023 — Dec 2024

• 01

  agentic-ai · production

  Agentic NL→SQL Chatbot

  Multi-tenant LangGraph chatbot with locally-served Gemma.

  ~60% LLM-call reduction · 50+ modules · binary RBAC

  A 3-tier follow-up engine routes queries through a hashed-prompt cache (L1), a small rewrite agent that prompts only with the diff against the previous turn (L2), and a full re-plan only when intent shifts (L3). A destructive-SQL gate blocks DDL/DML on read-only tenants before generation lands. The RBAC whitelist is encoded as packed bitmasks per (role × table × column) so authorization is a single AND, not a join — checked in front of generation, not behind it.

  key decision

  LangGraph over plain LangChain — explicit state graph beats chained prompts when retries, branching, and human-in-the-loop all touch the same state.

  LangGraphGemmaPythonPostgresRBAC

  view readmelive demo
• 02

  infrastructure · multi-cloud

  Active-Passive Disaster Recovery

  Two clouds, three regions, canary failback gated on health checks.

  Cloudflare → AWS Mumbai (active) · AWS SG (hot) · Azure (failover)

  Cloudflare's load-balancer routes primary traffic to AWS Mumbai with hot-standby on AWS Singapore and Azure failover for MongoDB Atlas. Failback ramps 5%→25%→50%→100% gated on rolling error rate and p99 latency — fail any gate and traffic snaps back automatically. Health checks chain edge → app health → DB write probe so a single failing layer can't drag the others. Terraform owns the topology end-to-end and Jenkins runs weekly drills so the failback path doesn't bit-rot.

  key decision

  Active-passive over active-active — read-heavy traffic, small team, and a data tier that couldn't safely tolerate concurrent multi-region writes.

  AWSAzureCloudflareTerraformJenkins

  view readme
• 03

  full-stack · led

  TeamSync — Project Management Platform

  Backend + integrations lead, Infosys Springboard 5.0.

  Role-based access · task tracking · team workflows

  Led the backend and integration tracks of a collaborative project-management platform built on Node.js, Express, React, and MongoDB with role-based access, task tracking, and end-to-end team workflows. Owned the API design, auth model, and the integration surface; coordinated four contributors across the build.

  scope

  Backend + integrations lead — owned API design, auth model, and the cross-team integration surface.

  Node.jsExpressReactMongoDB

  view readmelive demo

— rest of the work lives on github.com/ShubhamPatel2305