Bengaluru, India · Open to AI & Agentic AI roles

I find problems people have quietly accepted — and build the systems that solve them.

I'm Tejasgouda, an AI engineer. The interesting part of this work isn't the model it's the moment it stops being a demo and becomes something people depend on: retrieval that's actually grounded, agents that know when to stop, guardrails that catch the confident wrong answer. Every system below started as somebody's daily friction.

Day job — 4 years at Accenture, Bengaluru: three on PCI-DSS payment infrastructure in Java, the last as an AI Engineer shipping multi-agent workflows and RAG pipelines to production .

How I build

Anyone can get a model to answer once. The engineering is making it answer the same way tomorrow.

Ground it Retrieve over the real source — a repo index, a live catalog, actual telemetry — never the prompt's summary of it.
Check it Deterministic guardrails that verify what the model claimed against something incapable of hallucinating.
Measure it Golden datasets and LLM-as-judge harnesses, so a prompt tweak that quietly degrades output gets caught.
Bound it Agent loops that terminate, humans in the loop on risky actions, token budgets that survive real traffic.
Agile delivery · Live

Agiler AI

An autonomous multi-agent platform that turns a raw PRD into a sprint backlog the codebase actually agrees with.

Try the live demo agilerai.space
The problem

A product doc becomes a backlog through three days of meetings, and the story points are guesses wearing the costume of estimates.

Hand it to an LLM and you get faster guesses — confident stories citing files that don't exist, sized against nothing real.

What I built
  1. A five-agent CrewAI pipeline that drafts epics, user stories and Gherkin acceptance criteria, streaming each agent's output to the browser live over Server-Sent Events instead of a spinner.
  2. A hybrid retrieval engine — BM25 sparse search fused with dense embeddings via Reciprocal Rank Fusion — so every story is written against the indexed repository, not the PRD's description of it.
  3. A deterministic hallucination guardrail that cross-checks every file path the model cites against the real codebase index. Fabricated references never reach the user.
  4. A bounded Reflexion loop: a QA agent critiques the draft and triggers exactly one revision pass — enough to fix, not enough to spiral.
  5. An LLM-as-judge evaluation harness over a golden dataset, catching prompt changes that degrade non-deterministic output before they ship.
CrewAIHybrid RAGGPT-4o FastAPIPythonBM25 scikit-learnSSEReact PostgreSQLDocker
Open agilerai.space and try it
Conversational commerce · Live

ChatCart

Multi-tenant SaaS that gives any small merchant an AI storefront — one that answers, recommends and checks out inside the conversation.

Try the live demo chatcart.space
The problem

A small store gets a search box. A shopper who doesn't know the exact product name types two words, finds nothing, and leaves.

The fix is a conversational storefront — which normally means hiring an AI team the store will never afford.

What I built
  1. A conversational RAG assistant retrieving over an in-memory index of catalog metadata — grounded recommendations, live stock answers and dynamic prompt chips in < 50 ms, fast enough to read as typing rather than loading.
  2. Chat-to-cart: intent classification turns a sentence into a cart action, and checkout completes in-conversation across Razorpay and UPI Reserve Pay. The shopper never leaves the thread.
  3. Tenant isolation by construction — store-scoped queries, JWT auth and FastAPI dependency injection, so one merchant's catalog can't surface inside another's assistant.
  4. Merchant payment secrets held under Fernet envelope encryption at rest, decrypted in memory only long enough to authorize a charge.
Conversational RAGMulti-tenant FastAPIPythonTF-IDF PyJWTFernetRazorpay React 18TypeScript
Open chatcart.space and try it
Retail intelligence · Live

Comparo

An agentic research engine that aggregates listings across Indian retailers and answers the question comparison sites don't: should you buy today?

Try the live demo comparo.space
The problem

Price comparison tells you what something costs right now. It never tells you the sale that halves it is eleven days away.

And the same product appears four times under four slightly different titles, so the cheapest result is often just the worst-named one.

What I built
  1. A LangGraph-orchestrated aggregation pipeline firing concurrent SerpAPI and Tavily calls across Amazon, Flipkart and Croma — 65% lower end-to-end latency than fetching in sequence.
  2. Variant resolution using RapidFuzz token matching at 92% precision, collapsing four listings of the same phone into one row you can genuinely compare.
  3. A Decision Coach grounded in an Indian sale-calendar matrix, returning an explainable buy vs. wait call with its reasoning attached — not just a verdict.
  4. Checkout exposed as agent-callable tools through a FastMCP server, so an autonomous agent can complete a purchase, secured with HMAC-SHA256 signature verification.
  5. A background price radar daemon scanning for drops and dispatching alerts, so the user doesn't have to come back and check.
LangGraphMCPGPT-4o-mini FastAPISerpAPITavily RapidFuzzRazorpayReactTypeScript
Open comparo.space and try it
Observability · Open source

AEGIS War Room

An incident platform where an AI agent reads live topology, metrics and logs, then hands the on-call engineer a ranked list of what probably broke.

Read the source code GitHub
The problem

It's 3 a.m. Something is down. The engineer has twelve dashboards, four log streams, and no hypothesis.

Triage becomes pattern-matching under sleep deprivation, while the clock is the incident.

What I built
  1. A root-cause investigation agent on the Anthropic Claude API that reasons across service topology, metrics and logs together, returning ranked hypotheses with the evidence behind each one.
  2. A WebSocket telemetry pipeline (Fastify WS + Redis + BullMQ) ingesting live signals from instrumented services and batching them into a real-time dashboard.
  3. A published npm SDK with three integration paths — Express middleware, programmatic client, raw WebSocket — so any external service starts reporting in a single line.
  4. Org-level multi-tenancy across 30+ endpoints through centralized query-scoping middleware, blocking cross-tenant leakage at the database layer rather than route by route.
  5. Hardened and proven: JWT auth, RBAC, per-route rate limiting, OpenAPI 3.1 docs and a 65-test Vitest suite running in GitHub Actions.
Claude APIRoot-cause agent TypeScriptFastifyNode.js RedisBullMQWebSocket PostgreSQLDrizzleKubernetes
Browse the repository on GitHub
Vertical SaaS · Live

DeskMedia

One multi-tenant platform, three front doors — gyms, clinics and schools each get a portal shaped around how they actually run the day.

Try the live demo deskmedia.online
The problem

A gym runs on a WhatsApp group and a paper register. A clinic's queue is a name shouted down a corridor. A coaching institute takes attendance in a notebook.

Underneath, they all need the same spine — members, schedules, roles, billing — yet every vendor rebuilds it from scratch and prices it for enterprises.

What I built
  1. A multi-tenant SaaS core — tenants, roles, members, scheduling and billing — shared across every vertical, so adding a domain is configuration rather than a rewrite.
  2. Three domain-specific portals on that core, each carrying its own workflows and vocabulary instead of one generic admin panel bent to fit three businesses.
  3. Role-scoped access throughout, so an owner, a trainer, a front-desk operator and a member each land in a portal built for their job.
Fitness IronDesk

Gym and club portals — memberships, plans, attendance and the daily front-desk workflow.

Healthcare DocDesk

Clinical queue management and EMR records for hospitals and clinics.

Education ClassDesk

Class scheduling and coaching-institute ERP for schools and tutoring centres.

Multi-tenant SaaSRole-based access Gym portalClinic EMRClass scheduling Coaching ERPReactTypeScript
Open deskmedia.online and try it
Contact

If you're building something in this space, I'd like to hear about it.

⌘ + Enter to send