Building LLM Applications

Problem: General-purpose chatbots invent nutrition facts and figures, which is unsafe for health guidance and impossible to audit or trust.

Solution: A Coordinator routes each request to a Knowledge, Meal Planner, Meal Analyzer or Health Advisor agent — deterministic rules first at zero token cost, Granite JSON classification only for ambiguous messages. A RAG pipeline (ChromaDB, page-bounded chunking, top-k 5, 0.35 similarity threshold) grounds answers in the user's own PDFs with filename + page citations built in Python, so the model cannot invent a source; anything below the threshold is explicitly labelled general knowledge. A deterministic engine computes BMR/TDEE, macros and a 0–100 health score.

Impact: Ships as a full-stack Flask app across 36 endpoints with streaming (SSE) chat, offline + live modes, and 156 automated tests at 88% coverage running in CI with no credentials required — every reply exposes which agent answered and why.

Problem: Screening thousands of resumes by hand is slow and inconsistent, and the challenge banned GPUs, networking and heavyweight models (≤5 min wall-clock, ≤16 GB RAM).

Solution: Reverse-engineered the job description into machine-readable signals, then built a five-component interpretable scorer with decoy-profile detection and a fact-grounded reasoning engine that justifies each result and is mechanically verified against the source record. An embedding-based semantic layer was evaluated and deliberately rejected: the corpus collapsed to 44 unique text templates, so exact hash classification beat approximate similarity on accuracy, latency and reproducibility.

Impact: Ranked 100,000 candidates in under a minute on CPU alone — streaming a 465 MB dataset in ≤2 GB of memory, roughly 5× inside the time budget — with byte-identical output across repeated runs. Built for the Redrob × Hack2Skill INDIA.RUNS Data & AI Challenge.

Problem: Subscription businesses lose revenue when at-risk customers churn silently, and accuracy alone hides how many churners a model actually misses.

Solution: Cleaned and encoded 7,032 Telco records, ran EDA in Matplotlib/Seaborn, trained a scikit-learn logistic-regression classifier with class balancing to prioritise recall, and evaluated with precision, recall, F1 and confusion-matrix analysis.

Impact: Raised churn-class recall from 0.52 to 0.79 by accepting a precision trade-off (0.50, F1 0.61) — a deliberate choice, since a missed churner costs more than a false alarm. An audit of the pipeline also caught an encoding-order bug that had inflated the feature matrix from 30 to 7,061 columns; after the fix the model's true drivers (contract type, tenure, fiber service, electronic-check payment) came through cleanly in a stakeholder-ready Power BI dashboard.

Problem: Building a portfolio means balancing return against risk — hard to do well without simulating thousands of possible allocations.

Solution: Measured annualised returns, volatility and correlations for AAPL/MSFT/GOOGL/AMZN, then ran 5,000 Monte Carlo portfolios to find the allocation that maximises the Sharpe Ratio.

Impact: Improved the Sharpe Ratio from 0.87 (equal-weight) to ~0.92, reallocating toward Apple and Microsoft and away from higher-volatility Amazon.

Problem: Long-term climate questions require filtering and aggregating a large emissions dataset before any real trend becomes visible.

Solution: Queried a consolidated Our World in Data CO₂ dataset in Google BigQuery — filtering by country and year, aggregating emissions by decade, and ranking top emitters with a scalar subquery — then used Pandas and Matplotlib for time-series, per-capita and GDP-versus-emissions comparisons across India, China and the United States.

Impact: Showed China's steep emissions rise, the US plateau with apparent partial GDP decoupling, and India's steady per-capita growth — as clear, evidence-based visual narratives.