Journey  ·  Projects  ·  Systems Reference  ·  Stack  ·  Resume  ·  Contact

I build AI systems that hold up under enterprise scrutiny — retrieval pipelines whose answers trace back to a source passage, and multi-agent workflows that pause for a human before they commit.

Computer Engineering at Thapar Institute of Engineering & Technology. Over a six-month AI engineering internship at Ernst & Young (Technology Consulting, CNS — Technology Strategy & Transformation), I went from validating someone else's enterprise AI system to designing and building three of my own.

One standard runs through all of it: an AI output is not correct because it is plausible. It is correct because it is verifiable against a source, consistent across repeated runs, and fails safely to a human rather than quietly producing something wrong.

Four projects. Each one was motivated by the outcome of the one before it.

Validator → validated a live government AI portal, and learned what "correct" has to mean in an enterprise setting. Builder → applied that standard to a system of my own, starting at the simplest retrieval architecture that could work. Architect → measured where that baseline broke and redesigned its retrieval to fix exactly those failures. Platform Engineer → generalised the pattern from answering questions about documents to generating the documents themselves.

Document intelligence with no vector database — BM25 lexical retrieval and a locally hosted LLM.

Problem. Government departments hold large policy repositories that keyword tools can match on text but not on meaning. An officer answering one question may search the whole repository by hand — slow, inconsistent between reviewers, and impossible to audit, because a manual answer carries no trace back to the passage that justified it.

Why not a vector database. The conventional answer is embeddings plus a vector store. For air-gapped government deployments that means embedding infrastructure, index maintenance and recurring cost. This project tested whether that burden was necessary before proving it necessary — starting at the keyword-only end of the spectrum on purpose.

Key decisions. Structure ingestion to preserve traceability from answer to source passage from the outset, rather than retrofitting citations later. Constrain the prompt so responses are generated strictly from retrieved content, not model knowledge. Rerank with a cross-encoder so lexical recall is traded back for precision before context is assembled.

Technology — Python · FastAPI · BM25 · Cross-Encoder Reranking · Ollama · Streamlit · Tesseract OCR

_{A document-grounded answer, with retrieval and generation timing and the full citation trace of source chunks used.}

Retrieval pipeline

Outcome. It worked — and its limits are the point. Retrieval degraded exactly where keyword matching has no notion of meaning: paraphrased queries, indirect references, non-English content. That finding motivated the next system.

Repository →

Semantic and lexical retrieval run in parallel and fuse into one ranking.

Problem. The three failure modes measured in the vectorless baseline, in a corpus that also contains scanned pages with no machine-readable text layer and terminology in more than one language.

Why not just switch to vector search. Keyword retrieval is precise on policy numbers and regulation codes; vector retrieval handles paraphrase and crosses languages but can under-rank an exact-but-non-obvious match. Choosing either means accepting the other's weakness. Fusing both rankings — rather than replacing one with the other — was driven by an observed failure pattern in a working system, not a generic preference for more advanced retrieval.

Key decisions. Treat semantic and keyword retrieval as complementary signals and merge them with Reciprocal Rank Fusion. Make citation generation a first-class pipeline output rather than an afterthought. Use multilingual embeddings so one index serves cross-lingual search instead of one store per language. Keep inference on-premise so document content never leaves the host.

Technology — Python · FastAPI · FAISS · ChromaDB · BM25 · Reciprocal Rank Fusion · Multilingual Embeddings · React · Ollama · OCR

_{Retrieval transparency — every retrieved chunk with its source document and relevance score.}

Retrieval pipeline

Outcome. Better semantic understanding, more consistent multilingual handling, and explainability through citations — a grounding-plus-transparency design now proven across two independent systems. That consolidated pattern became the foundation for the platform below.

Repository →

Nine specialised AI agents turn a business requirement into engineering artifacts, gated by mandatory human approval.

Problem. Enterprise delivery depends on a sequence of specialists — business analysts, architects, database designers, security reviewers — each producing documentation that must stay consistent with everyone else's. Coordinating that manually produces incomplete requirements at initiation, architecture quality that varies with individual experience, artifacts that drift out of consistency, and slow knowledge transfer.

Why not a single assistant. Language models draft structured technical content competently, but are typically deployed as isolated assistants rather than coordinated collaborators. An assistant answering one prompt at a time cannot hold a project's context across disciplines — which is precisely where the inconsistency comes from.

Key decisions. Centralise project memory in a Memory Agent so each agent reads shared context instead of re-deriving requirements — this is what prevents cross-artifact drift. Require explicit human sign-off at two checkpoints before output influences the next stage. Structure the schema around traceability first, which is what makes the Documentation and Temporal Replay Centers possible at all. Keep the reasoning layer configurable so a project can supply its own provider credentials.

Technology — Python · TypeScript · FastAPI · React · PostgreSQL · LangGraph · Multi-Agent Orchestration · BYOK Provider Routing

_{Orchestration dashboard, then the Requirements and Architecture workspaces generated by their agents.}

Agent pipeline — two mandatory human approval checkpoints

Outcome. A coordinated agent set producing consistent requirements, business analysis, architecture, database design, UI/UX and compliance artifacts — governed by approval checkpoints and kept auditable by the Documentation and Temporal Replay Centers.

Repository →

Supporting designs that span the projects above.

Retrieval trade-off — the comparison that drove the hybrid redesign

Data model — traceability designed in from the schema up

BYOK provider resolution — graceful degradation across AI providers

Media generation pipeline — artifacts to narrated presentation

Maharashtra AI Portal — the twelve-stage clearance workflow I validated

Interface stills — full resolution

Vectorless RAG — conversational interface	Hybrid RAG — conversational interface
Hybrid RAG — citation trace	SDLC Platform — orchestration dashboard
SDLC Platform — requirements workspace	SDLC Platform — architecture workspace

AI Engineer — Retrieval-Augmented Generation & Backend Systems
_{Single page, ATS-friendly. B.E. Computer Engineering, Thapar Institute of Engineering & Technology, 2022–2026.}

Download PDF   ·   Preview in browser