Manideep Gouraram

Résumé

Manideep Gouraram · Applied AI Engineer & Systems Builder

Applied AI and systems engineer. I build products where a model, a rules engine or a data pipeline has to be right, not just fluent. Recent work covers tool-grounded conversational ordering, deterministic incident investigation and the evaluations that keep both honest.

LinkedIn GitHub

Selected projects

  1. Tableside

    tableside.menu

    Live at tableside.menu · private repository · demo captures use a synthetic restaurant

    A diner wants useful menu guidance. A language model left unchecked can invent dishes, prices or allergens, or claim cart changes that never happened.

    Role
    Sole builder: product design, database and security model, AI tool layer, interface and deployment. Developed with AI coding assistance.
    Verified finding
    In a captured session on a synthetic demo restaurant, asking the waiter to add a dish and place the order produced a server-priced summary and a confirmation question. The conversation log shows no order-submission call, and no order was created.

    Next.js (App Router) · React · TypeScript · Supabase (PostgreSQL, Auth, Realtime, row-level security) · Stripe Connect · Zod · Tailwind CSS

  2. Tableside waiter evaluation

    Part of the Tableside repository (private) · fictional restaurant · free model via OpenRouter

    Unit tests prove the code paths. They can't say whether a model, given a real diner's sentence, ends up doing the right thing.

    Role
    Sole builder: scenarios, simulated restaurant, graders, runner and report. Developed with AI coding assistance.
    Verified finding
    The first full run passed 41 of 47 judged visits and exposed two production bugs. After those and four grader mistakes were fixed, a full run passed 46 of 47.

    TypeScript · Vitest · OpenRouter · Tableside's production prompt, tool schemas and agent loop

  3. Fulfillment Exception Causality Engine

    Independent prototype · synthetic data only · runs locally

    A fulfillment exception reports what failed. It rarely says which upstream condition caused it, or which delivery promises are now at risk.

    Role
    Sole builder: problem framing, simulator, candidate-ranking engine, interface and evaluation. Developed with AI coding assistance.
    Verified finding
    On 15 fresh held-out synthetic datasets, the top-ranked candidate matched the simulator's injected condition for all 166 exceptions that had one, and 2 of 168 asserted candidates were wrong (1.2%). The benchmark is synthetic; real-world accuracy is unknown.

    Next.js 16 · React 19 · TypeScript · PGlite (PostgreSQL in WebAssembly) · Zod · Vitest · Playwright

Education

  • Master's in Management Science (Analytics & Systems)

    University of Bridgeport

  • Bachelor's in Electronics and Communication Engineering

    J.B. Institute of Engineering and Technology

Skills, as used in the projects above

Applied AI
Tool-calling assistants · Schema-validated tools · Grounding checks · Model vetting and fallback · Anthropic and OpenRouter APIs
Systems and data
PostgreSQL · Row-level security · Event normalization · Deterministic scoring engines · Simulation for evaluation
Product engineering
TypeScript · React · Next.js · Supabase · Stripe Connect · Zod
Quality
Vitest · Playwright · GitHub Actions CI · Held-out evaluation design

Contact

gmdg4444@gmail.com · LinkedIn · GitHub