AI Engineer · Project 4

Domain-Expert Research Engine

Where ChatGPT is wrong, this is right. The Harvey-clone for your niche.

Duration: ~10 daysDeliverable: An advanced-RAG service replacing P1's retrieval layer. Public eval scorecard comparing every variant. Tech-blog post comparing approaches. 4 checkpoint PRs reviewed.

Build a research tool a domain expert would actually trust to do their job. Flagship pitch: legal precedent search for solo and small-firm lawyers. Westlaw is $400/month, lawyers want a cheaper option that actually works. The real market is $9B raised in 2025 alone (Harvey: $300M, Glean: $260M, Hebbia: $130M, Elicit, Casetext) — and there's massive room for niche verticals. Hybrid retrieval (BM25 + dense), cross-encoder reranking, query rewriting / HyDE, GraphRAG. Every improvement quantified with Project 3's eval harness — you ship with numbers, not vibes. The portfolio outcome: 3 real domain experts from your network using it to do their actual job.

The product in detail

The product solves the "ChatGPT is wrong about my domain" problem. The user is a working professional in a high-stakes field — law, medicine, finance, regulatory compliance — who can't afford to have an AI assistant make things up. They upload (or your product ingests) the authoritative documents for their domain, and from then on every answer is grounded, cited, and accurate enough to trust in their actual daily work. The flagship pitch: legal precedent search for solo and small-firm lawyers. Westlaw is $400/month and the interface hasn't changed since 2005. Lawyers want a cheaper, modern alternative.

The target user persona: a solo or small-firm lawyer who handles 5–15 active cases and needs to find relevant precedent fast. They're currently paying $400+/month for Westlaw or LexisNexis, or doing manual research that takes hours. Adjacent personas include nurse practitioners querying clinical literature for patient care decisions, retail investors analyzing 10-Ks and earnings calls, and compliance teams answering "is this configuration HIPAA-compliant" questions 20 times a week.

Real market: Harvey AI ($300M raised, Sequoia-led), Hebbia ($130M, used by hedge funds and law firms), Casetext (acquired by Thomson Reuters for $650M), Elicit ($9M for scientific research papers), PaperPal. Collectively over $1B raised in 2025 for domain-specific AI research tools. The pattern: pick a niche where accuracy and citations matter more than breadth. The winners are vertical, not horizontal.

Technically you upgrade Project 1's vanilla RAG into a production-accuracy retrieval system: hybrid retrieval (BM25 sparse + dense vector, fused with reciprocal rank fusion), cross-encoder reranking (two-stage: top-50 from hybrid, reranked to top-5 for the LLM), query rewriting and HyDE for ambiguous or jargon-heavy queries, and optionally GraphRAG for domains with strong entity relationships (case citations, scientific paper references, drug interaction graphs). Every improvement is quantified against Project 3's eval harness — you ship with metrics, not vibes.

What you ship: a research tool over a real domain corpus, used by 3+ real domain experts from your network to do their actual job. A public scorecard comparing every RAG variant you tried (vanilla → hybrid → reranked → HyDE → GraphRAG) with precision/recall/faithfulness numbers. A tech-blog post documenting when each approach wins and when it loses. When a recruiter for Harvey, Hebbia, or any vertical-AI startup sees this portfolio, they see someone who has already shipped what they're hiring for.

Pick what you'll build

This is one project with one skill graph — but you pick the product framing that fits your interests or career goals. Each option below is a viable portfolio piece. We'll narrow it together in the kickoff call.

  • Legal case research — precedent search for solo/small-firm lawyers (Westlaw is $400/mo; market is wide open below that)
  • Scientific paper discovery — your field's arXiv with citation-graph navigation (Elicit-clone for one niche)
  • Medical literature Q&A — PubMed search for a specific clinical question (audience: nurse practitioners, junior clinicians)
  • Financial report analyst — Q&A across 10-Ks, earnings calls, analyst reports (audience: retail investors)
  • Regulatory compliance Q&A — GDPR, HIPAA, SOC2 question answering (audience: compliance teams at small SaaS)
  • Production-accuracy codebase Q&A — when Project 1's vanilla approach isn't good enough for shipping
Why this project
  • Advanced RAG is a tacit must-know — Δ+3 with interviews 14% > posts 11%.
  • The specific tested concepts (reranking, hybrid search, HyDE, GraphRAG, RAG-vs-CAG, 'lost in the middle') appear in 7+ interview questions but rarely as explicit JD bullets.
  • The opportunity: build this and you can pass any RAG senior-engineer interview.

Research backing this project

Every project on projeckt.ai is calibrated against real job posts and real interview questions, not opinion. Here's the data behind this one.

27
Job posts citing concepts in this project
45
Interview questions this project prepares you for
4
Tacit must-knows drilled — concepts interviewers test but JDs rarely list
Tacit must-knows you'll drill
  • Hybrid retrieval (interview-heavy, post-light)
  • Cross-encoder reranking
  • HyDE / query rewriting
  • GraphRAG / RAG-vs-CAG decision-making

Themes covered

Each chip shows how many of the 46 analysed job posts cite this theme.

LLM & Prompt Foundations · 43 postsRetrieval & Knowledge Systems · 27 postsAdvanced RAG (hybrid, rerank, HyDE) · 5 postsEvals (drives every change) · 21 posts

Justified by real job posts

Verbatim quotes from the postings that drove the design of this project.

Develop and optimize RAG pipelines, agent architectures, and LLM-powered systems

LlamaIndex — AI Engineer

How do you handle the challenge of updating a RAG model's knowledge base without retraining the entire model?

ProjectPro — RAG Interview Questions (interview signal)

You're building a customer support chatbot for a global company, and user queries are often multi-lingual. How would you use an embedding model to retrieve relevant documents in various languages…

ProjectPro — RAG Interview Questions (interview signal)

What you'll be able to do

  • 1Implement BM25 sparse retrieval alongside dense vector retrieval and fuse rankings (reciprocal rank fusion)
  • 2Choose between bi-encoder and cross-encoder rerankers and explain the quality / latency tradeoff
  • 3Implement HyDE (Hypothetical Document Embeddings) query transformation
  • 4Implement query rewriting for multi-turn conversation context
  • 5Reason about chunking strategy: sliding window vs semantic vs document-structure-based
  • 6Detect and address the 'lost in the middle' phenomenon in long contexts
  • 7Compare Naive RAG vs RAG 2.0 quantitatively using Project 3's eval harness
  • 8Decide when to use CAG (Cache-Augmented Generation) vs RAG
  • 9Implement basic GraphRAG: entity graph + community detection + traversal-based retrieval
  • 10Articulate when fine-tuning would beat RAG — backed by experiment data showing it doesn't (for your use case)

Tools you'll learn

Hands-on with these tools across the paid and open-source stacks. The Paid / OSS / Both label shows which stack each tool belongs to.

LL
LlamaIndex
Both stacks
PI
Pinecone (hybrid)
Paid stack
WE
Weaviate (hybrid)
Paid stack
QD
Qdrant
OSS stack
CO
Cohere Rerank API
Paid stack
BGE-reranker
BGE-reranker
OSS stack
Sentence Transformers
Sentence Transformers
OSS stack
Neo4j (GraphRAG)
Neo4j (GraphRAG)
Both stacks

Week-by-week checkpoints

Each checkpoint is a Pull Request to your public GitHub repo. Rohan reviews on the PR; merge = checkpoint complete.

CP1: Hybrid retrieval (BM25 + dense)
Days 1–3
What ships

BM25 added alongside dense retrieval. Reciprocal rank fusion combines results. Eval harness shows Δ vs Project 1's baseline.

Acceptance criteria

Hybrid beats dense-only on ≥2 of 4 metrics by >5%; eval scorecard published.

New skills
  • BM25 / sparse retrieval
  • Reciprocal rank fusion
  • Eval-driven retrieval decisions
CP2: Cross-encoder reranking
Days 4–5
What ships

Two-stage retrieval: top-50 from hybrid → cross-encoder rerank → top-5 to LLM. Eval shows further improvement. Latency cost measured.

Acceptance criteria

Reranked top-5 beats fused top-5 on Faithfulness / Context Precision; latency increase quantified.

New skills
  • Cross-encoder rerankers
  • Two-stage retrieval
  • Quality-vs-latency tradeoff
CP3: Query rewriting + HyDE
Days 6–7
What ships

Query rewriter for multi-turn context. HyDE transformation applied. Eval shows improvement on ambiguous queries.

Acceptance criteria

Multi-turn test set passes; HyDE wins on technical-jargon queries.

New skills
  • HyDE (Hypothetical Document Embeddings)
  • Query rewriting
  • Multi-turn context handling
CP4: GraphRAG variant + comparative write-up
Days 8–10
What ships

Optional GraphRAG over a structured corpus (codebase or research papers). Comprehensive write-up comparing every RAG variant — when each wins, when each loses. Blog post draft.

Acceptance criteria

Write-up published as PR or blog; comparison table with eval numbers; clear 'use X when Y' guidance.

New skills
  • GraphRAG basics (entity graph + community detection)
  • Comparative analysis writing
  • Tech-blog craft

Pick your stack

Same skills, different credentials. Start free with the OSS stack or use the paid stack to learn directly on production tools.

ToolPurposeLicenseHardwareMarket equivalent
Qdrant or Weaviate self-hostedHybrid retrieval (dense + BM25)Apache-2.0DockerPinecone hybrid
BGE-reranker via Sentence TransformersCross-encoder rerankMITCPU OK, GPU fasterCohere Rerank
Ollama (Qwen 2.5 14B)LLMApache-2.016GB+ RAMGPT-4o / Claude
Neo4j Community EditionGraph DBGPLDockerNeo4j Aura
LlamaIndexPipeline compositionMITFreeSame

Estimated cost: $0/month — fully self-hosted.

Interview questions you'll be ready for

Real questions from 13 sources, mapped to the skills this project builds.

What is hybrid search, and why is it better than pure vector search?
Source: GitHub — amitshekhariitbhu/ai-engineering-interview-questions
Prepares you for: CP1 IS this — and you can defend the choice with metrics.
Why is the cross-encoder typically used as the re-ranker rather than the bi-encoder?
Source: GitHub — KalyanKS-NLP/RAG-Interview-Questions-and-Answers-Hub
Prepares you for: CP2 forces the quality / latency tradeoff to be quantified.
Explain how the HyDE query transformation technique works.
Source: GitHub — KalyanKS-NLP/RAG-Interview-Questions-and-Answers-Hub
Prepares you for: CP3 is the implementation.
What is GraphRAG and how does it improve upon standard RAG?
Source: Analytics Vidhya — 30 Agentic AI Interview Questions (2026)
Prepares you for: CP4 builds a GraphRAG variant and compares to vanilla — you've measured the answer.
Is RAG still relevant in the era of long context LLMs?
Source: GitHub — KalyanKS-NLP/RAG-Interview-Questions-and-Answers-Hub
Prepares you for: CP4 includes a RAG-vs-CAG decision write-up — your blog post IS the answer.

Ready to start building?

Book a free 30-minute project planning session. We'll tailor this project to your stack, team, and goals.