Project Case Studies

# The Solution

Built a full RAG pipeline with hybrid search (70% semantic + 30% keyword) that transforms natural language queries into structured exercise plans. The system detects constraints like injuries and restrictions, ensuring safety-focused recommendations.

# Architecture

User Query
  → Constraint Detection (injuries, goals)
  → Hybrid Search (70% semantic + 30% keyword)
  → Context Preparation (top 15 exercises)
  → LLM Generation (structured prompts)
  → Structured Output (exercise plan)

# Impact

• ✓First production RAG system at Tenzr
• ✓Transformed 1,578 exercises into intelligent, queryable system
• ✓Created 40+ KB of technical documentation across 6 focused docs
• ✓Established foundation for future AI-powered healthcare features

# What I Learned

• →Hybrid search significantly outperforms pure semantic for domain-specific retrieval
• →Context window size matters - too much context causes LLM parse errors
• →LlamaIndex provides better RAG primitives than LangChain for this use case

# The Solution

Implemented PgBouncer connection pooling with optimized configuration. Created custom benchmarking tooling to measure before/after performance with statistical rigor.

# Impact

• ✓Connection time: 329ms → 2ms (99.4% reduction)
• ✓Concurrent handling: +458% capacity (5.6×)
• ✓Throughput: +6.3% queries per second
• ✓Complex queries: 13% faster execution

# What I Learned

• →Performance improvements without metrics are just opinions
• →Benchmarking tooling is as important as the optimization itself
• →Documentation enables future engineers to build on your work

# The Solution

Identified inefficient database queries using N+1 patterns. Implemented eager loading, query optimization, and strategic caching to reduce round trips.

# Impact

• ✓Page load time: 20s → 2s (90% reduction)
• ✓Significantly improved user adoption
• ✓Enabled usage by workers with limited technical background

# What I Learned

• →On-site user research reveals problems that logs never show
• →Performance is a feature, especially for users with limited connectivity
• →N+1 queries are the most common performance killer

# The Solution

Built a full-stack application with role-based access control, bilingual interface (English/Hindi), and simplified UI patterns informed by direct observation of users.

# Impact

• ✓Enabled shift logging across all technical literacy levels
• ✓Bilingual support for Hindi-speaking users
• ✓Role-based access control for workers, supervisors, and managers

# What I Learned

• →User research at actual work sites is invaluable
• →Accessibility means different things in different contexts
• →Simple UX often requires more complex engineering

# The Solution

Built a 3-tier microservices system with RAG-powered semantic search. Users describe courses in natural language like "easy science course with no prerequisites" and receive semantically relevant matches with personalized explanations for why each course fits.

# Architecture

User Query
  → Query Interpretation (Gemini LLM)
  → Embedding Generation (OpenAI)
  → Vector Search (ChromaDB)
  → Post-filtering (campus, level, WQB)
  → Ranking (semantic + elective score)
  → Match Reason Generation (Gemini LLM)
  → Personalized Results

# Impact

• ✓Semantic search over 1,200+ courses using ChromaDB vector database
• ✓5-step RAG pipeline: interpret → embed → search → rank → explain
• ✓Hybrid ranking: 80% semantic relevance + 20% elective quality score
• ✓Sub-second response times with rate limiting and query history

# What I Learned

• →RAG pipeline design with hybrid ranking outperforms pure semantic search
• →Post-filtering after vector search handles complex filter combinations efficiently
• →Microservices architecture enables independent scaling of AI workloads