name:      Charan
focus:     DSA, System Design
building:  AI agents, interactive dashboards, and analytics tools
learning:  DSA, System Design, backend development
approach:  Fewer projects, finished properly

• 📊 I build data-driven web applications — mostly turning unstructured public datasets into interfaces people can explore and make decisions with.
• 🗺️ Recent work is geospatial: mapping drought risk across Telangana and healthcare capacity across India.
• 🔬 Completed the Infosys Springboard Virtual Internship 6.0, delivering a healthcare operations analytics platform as the capstone.
• 🧠 Currently going deeper on DSA and system design.
• 🎯 Focused on shipping fewer, more complete projects rather than many partial ones.

Languages
Python · Java · JavaScript · SQL · C

Data & Visualisation
Pandas · NumPy · Plotly · Matplotlib · Seaborn · GeoPandas

Web & Frameworks
Streamlit · Flask · TensorFlow · OpenCV · HTML5 · CSS3

Tools
Git · GitHub · VS Code · Jupyter

AI finance controller for multi-source settlement reconciliation — built for Razorpay merchants

Python · Llama 3.2 via Ollama · JavaScript · pytest

• Problem — Razorpay settles orders in bundled lump sums days later, net of fees and GST. A merchant's bank statement shows one credit with no indication of which orders it covers — someone reconciles that by hand, daily.
• Approach — Tested the AI on the actual matching first: asking an LLM which orders summed to a bank line scored 0/9, even with retries. Deterministic Python scored 3/3, instantly. So code does every calculation; the LLM only writes the explanation for lines it can't resolve.
• Evaluation — A sealed answer key the agent is never allowed to open. 93.1% settlement match rate on held-out months, 0 false positives, every duplicate charge and stray credit caught. A 248-run sensitivity analysis over the assumed parameters gives the honest range: 76.3%–98.1%.
• Engineering — 127 tests, including one that runs identical inputs through both the Python and browser reconciliation engines and fails on any disagreement. Zero third-party Python dependencies. CI on GitHub Actions.

Geospatial water resource monitoring for Telangana

Python · Streamlit · GeoPandas · Plotly

• Problem — Rainfall alone is a misleading drought signal. A district with poor rainfall but large reservoirs may be secure, while one with good rainfall and no storage stays vulnerable.
• Approach — Normalised monsoon rainfall and dam storage capacity to a 0–100 scale and combined them into a single drought index, so districts can be ranked against each other.
• Features — Interactive Plotly choropleth over district boundaries, click-through drill-down per district, covering all 33 Telangana districts across 2024–2025.
• Finding — Districts around Hyderabad with negligible reservoir storage score worst consistently, leaving them wholly dependent on same-year rainfall.

Interactive analytics platform for healthcare infrastructure planning across India

Python · Streamlit · Pandas · Plotly

• Problem — Healthcare planning data arrives unstructured and fragmented, making it hard to assess hospital capacity, workforce availability and emergency preparedness.
• Approach — Consolidated infrastructure metrics, workforce indicators and readiness assessments into a single interactive Streamlit application.
• Features — Bed capacity and hospital-type analysis, emergency readiness scoring, region-level geographic coverage comparison, interactive filtering, dataset export.
• Context — Capstone project for the Infosys Springboard Virtual Internship 6.0, mentored by Mousami Shrivastava.

Estimates age and gender from a photo or a live camera capture

Python · Flask · OpenCV · TensorFlow · JavaScript

• Pipeline — YuNet locates faces and returns five landmarks, each face is rotated onto a level eye axis, then a fine-tuned MobileNetV2 predicts age as a number and gender.
• Detector choice — YuNet over RetinaFace and MTCNN: within a few points on accuracy, roughly 100× faster on CPU, 340 KB, and no extra dependency. Accuracy that cannot run in real time is not accuracy you have.
• Train/serve consistency — the training set is re-cropped with the same detector the app uses at inference, so the model does not learn on one distribution and predict on another.
• Full-stack — Flask JSON API with a plain HTML/CSS/JS frontend; camera capture runs in the browser via getUserMedia so it still works when hosted.

Front-end storefront built from scratch without a framework

HTML5 · CSS3 · JavaScript

Authentication screens complete. Catalogue, product pages and cart are next.

Infosys Springboard — Virtual Internship 6.0
_{Applied software development · Mentor: Mousami Shrivastava}

Built the Healthcare Operations Analytics Dashboard as the capstone deliverable, covering data cleaning, metric design and interactive visualisation.

MILESTONE 1  ████████████████████  Python & web fundamentals          ✅
MILESTONE 2  ████████████████████  Data analysis & visualisation      ✅
MILESTONE 3  ████████████████████  Professional internship            ✅
MILESTONE 4  ███████████████░░░░░  Full-stack applications            🚧
MILESTONE 5  ████████░░░░░░░░░░░░  DSA & System Design                🚧
MILESTONE 6  ████░░░░░░░░░░░░░░░░  Production-level products          ⏳