Sandeep Kumar Mahto · Software Engineer

• 0.911faithfulness
• 0.969context precision @4
• 0unsafe / 50 tickets
• 6.4msescalation p50

A six-layer multi-agent support engine: four MCP servers (tickets, knowledge base, CRM, action executor) over one shared SQLite/FTS5 store, a System-1 classifier that returns typed intent/sentiment/urgency JSON in milliseconds instead of burning an LLM on routing, hybrid RAG (dense MiniLM + BM25 -> reciprocal-rank fusion -> cross-encoder rerank, with query decomposition), then a LangGraph supervisor running research -> resolution -> QA with a cyclic retry when grounding fails. Auto-reply fires only on QA pass AND confidence >= 0.8, and the >$50 refund rule lives inside the MCP server so no client can bypass it. The whole pipeline runs offline with zero API keys: faithfulness 0.911, context precision @4 0.969, 50 tickets swept with 0 unsafe actions.

what i learned: guardrails belong in the server. anything a client can skip, a client will skip.

• LangGraph
• MCP
• Hybrid RAG
• RAGAS

• 34passing tests
• ~16×gpu throughput
• ≤20msbatch window
• 4microservices

A distributed LLM inference serving platform: a gateway doing hashed-key auth, tiered rate limits, quota billing and circuit-breaker degradation; a Redis Stream with pull-based load balancing feeding GPU workers that dynamic-batch inside <=20ms windows for ~16x throughput; two-tier caching (shared exact + local semantic cosine >= 0.90) that keeps answering while the fleet is down; and row-locked reservations settling idempotently, so billing is exactly-once under at-least-once delivery. One codebase runs as a single process for dev or four containers in Docker Compose, because the bus is an interface with two implementations. 34 tests, Prometheus metrics, live ops dashboard.

what i learned: the bus being an interface is why one process became four containers without a rewrite.

• Redis Streams
• Dynamic Batching
• Circuit Breaker
• Prometheus

• 82.4%orders completing
• p=3.1e-6z-test
• +3.5%relative lift
• £298Kper year

Find out why shoppers abandon carts, then prove the fix actually works. Dual conversion funnels, time-aware refund matching and 25 monthly cohorts over two years: 82.4% of started orders complete, 19% of completed orders are refunded, and 27.7% of customers never come back. Then a two-proportion z-test (p = 3.1e-6) with power analysis, guardrail metrics and SRM checks certifies a +3.5% relative lift worth roughly £298K a year. Five scripts reproduce every number.

what i learned: an a/b test you can't reproduce is just a screenshot.

• A/B Testing
• Cohort Analysis
• Statistics
• pandas

• 12labeled sections
• 3models compared
• 5-foldcross-validation
• 0api keys needed

One file, twelve labeled sections, the entire classical ML workflow: synthetic customer data, EDA, a Pipeline + ColumnTransformer for preprocessing, logistic regression vs random forest vs gradient boosting, cross-validation, GridSearchCV tuning, precision/recall/ROC-AUC evaluation, feature importance, PCA and K-Means, then joblib persistence and inference on unseen data. Clone and run: no downloads, no API keys.

what i learned: my first grid-searched pipeline. overfitting taught me to stop trusting training accuracy.

• Scikit-Learn
• GridSearchCV
• PCA
• joblib

• 4.5GBper model, 4-bit
• 7Bparameters
• 8GBmacbook air m2
• 0cloud calls

An assistant that never phones home: Qwen2 7B and Mistral 7B running fully offline on an 8GB MacBook Air M2 through Apple's MLX framework. 4-bit quantization shrinks each model from ~15GB to ~4.5GB so nothing swaps, with multi-model routing and session memory. Zero cloud calls after the download.

what i learned: quantization is the difference between "runs on my laptop" and "runs on someone else's server".

• MLX
• Local LLM
• 4-bit Quantization
• Apple Silicon