I'm a student at CBIT, Hyderabad who builds AI systems that show their working. My projects combine machine learning, graph analytics, OCR and LLMs, with one rule: the data is the source of truth, not the model. Every score, ranking and generated sentence traces back to real evidence, and when the evidence isn't strong enough, the system abstains instead of guessing.
I take ideas all the way from research to a shipped product: data pipelines, interpretable models, sandboxed execution, FastAPI / Django backends, Next.js / React / React Native frontends, Docker deployments and 1,000+ automated tests across my flagship work.
focus: [Explainable AI, Graph ML, LLM systems & agents, Applied ML, Backend Engineering]
building: Forensics tools, health-screening apps, compliance AI, fair technical hiring
principles: [evidence over generation, abstain rather than guess, test everything]
learning: [System Design, Distributed ML pipelines, Computational Neuroscience]|
🇮🇳 Smart India Hackathon 2026: two problem statements 🎃 CBIT Hacktoberfest '26 (Team JL) 🩺 SFIC 2026: shipped an Android APK |
🧠 Open-source contributor to Brian2 📄 Wrote a research paper on neuroevolution 🧪 Testing is part of the product |
🏛️ Areté: Portfolio-Driven Coding Assessment Platform
A coding test built from the candidate's own portfolio, graded only by tests. Built at CBIT Hacktoberfest '26 with Team JL.
Areté turns a résumé, GitHub, LeetCode and Codeforces profile into a validated 60-minute assessment in the candidate's own domain and stack: one real-codebase task and one DSA task at a Novice, Adept or Master tier.
- 🏗️ Tasks prove themselves first: a validation gate checks that the reference passes 3× in a row, the starter fails and the statement doesn't leak the fix. Bugs come from a LibCST mutation engine
- 🔒 Sealed sandboxes: a fresh Docker container per run with no network and capped CPU, memory, PIDs and time. Hidden tests run in their own sandbox
- ⚖️ Deterministic scoring: fail-to-pass grading (SWE-bench style). No LLM ever decides a score, and a Re-grade button shows the same score every time
- 🪽 Guarded AI helpers: a hint assistant that refuses to give answers, a Labyrinth code-comprehension round, a consented AI voice interviewer, and a Trail Map replay of how the candidate moved through the code
FastAPI Next.js TypeScript Docker sandboxes PostgreSQL Redis + Celery Monaco LibCST
👂 Shravan Setu: Offline Hearing Screening That Follows Through
Others built a hearing test. We built the system that makes sure the child actually gets treated. Built for SFIC 2026.
A teacher screens a class on an ordinary Android phone with wired earphones and no internet. Each child who needs care becomes a referral case that the school tracks through triage, batched ear-camp requests, missed visits and reschedules until the child is treated.
- 🎧 Smart screening: PASS / RETEST / REFER at 0.5–4 kHz. Silent catch trials detect random tapping, and the mic rejects tones masked by noise
- 📴 Offline-first and Telugu-first: full Telugu UI and recorded voice instructions, plus idempotent sync to a laptop server
- 🔐 Tamper-evident: SHA-256 hash-chained audit log of consents, results and status changes
- ⚡ Cut audio start delay from 486 ms to 60 ms, and checked the screening rules against 5,456 real NHANES audiograms
| 486 | ₹0 | ₹14.9 | 📲 |
|---|---|---|---|
| tests passing | software cost | equipment per child per year (proposed) | APK release |
Expo / React Native TypeScript Django SQLite Offline sync Telugu + English
🔗 btc-intel: Offline Bitcoin Forensics for Investigators
Follow the coin back to the hand that sent it. Built for Smart India Hackathon 2026 (PS 26146, NTRO).
An air-gapped investigator console that clusters wallets into real-world owners, flags laundering patterns, traces funds to their cash-out points, and estimates which IP first broadcast a transaction from P2P relay timing.
- 🧠 Multi-engine detection: rules, anomaly detection and a Graph Neural Network, fused by a stacker with exact SHAP explanations
- 🛡️ Live red-team mode: inject a laundering pattern and see whether the stack catches it, with no retraining
- 📜 Court-ready evidence: a hash-chained chain-of-custody ledger (ISO/IEC 27037 style) and one-page PDF case reports
| 5 / 5 | 0.934 | 0 / 949 | 20 / 30 |
|---|---|---|---|
| illicit operations detected | alert precision | false alerts on a clean dataset | live red-team attacks caught |
Benchmarks on simulated generator data, as documented in the repo.
Python PyTorch (GNN) NetworkX SHAP React TypeScript Fully offline
⚖️ MetrIQ: Evidence-Backed Assistant for Indian Standards & BIS Services
Built for Smart India Hackathon 2026 (PS 26107).
Ask in English, Hindi or Telugu, or just photograph a product label. MetrIQ finds the governing Indian Standard, explains why it matched, and links the official BIS source behind every answer.
- 🔍 Retrieval-first design: deterministic code picks the answer, and the LLM only explains what retrieval found
- 📷 Local OCR label inspection: every extracted value links back to the exact pixel box it came from
- 🧪 Battle-tested: adversarial and prompt-injection queries produce zero hallucinated standard numbers
| 505 | 98.1% | 0 | 395 |
|---|---|---|---|
| verified standards indexed | correct standard in top-5 | invented standard numbers | tests passing |
FastAPI React 19 PostgreSQL ONNX OCR Local LLM (Qwen3) Multilingual
💬 Aminobots Chat Triage Agent: Bounded-Autonomy LLM Agent
An agent that reads code-mixed Telugu–English WhatsApp partner group chats and lists every real request: who asked and whether that's confirmed, a grade (act / seek approval / hand off / refuse), the numbered policy rule behind it, and the messages it relied on.
- 🧭 The LLM describes and code decides: identity, grade and rule come from a written policy (R1–R10) run in code, where the most cautious match wins
- 📈 Iterated from v0.1 to v1.0 with a proper eval harness: grade accuracy 0.89 → 0.94, approval load 0.44 → 0.18
- 🛡️ Zero over-autonomy on high-risk requests, 100% evidence validity, and prompt-injection cases included
- 💸 Runs on free models only ($0), with a local website that walks through every decision
| 0.89 | 0.94 | 1.00 | 0 |
|---|---|---|---|
| detection F1 | grade accuracy | sender attribution | high-risk over-autonomy |
Python OpenRouter (free LLMs) rule-engine Pydantic uv Eval harness
🔬 SideEffect2Cure AI: AI-Driven Drug Repurposing · Live demo ↗
For any disease, which already-approved drugs are promising repurposing candidates, and why?
- 🧬 A transparent 10-stage pipeline linking disease genes → biological pathways → drug targets → side-effect profiles
- 📉 Narrows a ~1,430-drug universe to biologically connected candidates, scored by an interpretable ML model
- 🗂️ Built on real biomedical data: SIDER · ChEMBL · Open Targets · Reactome
- 🤖 LLM-written explanations grounded in pipeline outputs, with a deterministic fallback
FastAPI scikit-learn NetworkX React + TypeScript Vercel
🐦 FlappyBirdAI: Neuroevolution Research Study
A genetic algorithm learns to play Flappy Bird, written up as a full research paper: "Neuroevolution of a Flappy Bird Controller: An Empirical Study of Genetic Operators, Evaluation Noise and Generalisation."
- ⚡ A headless simulator compiled with numba for high-speed training
- 📊 606 logged training runs, statistical analysis and fully reproducible figures
- 🧪 Tests how evolved controllers generalise to unseen levels
| Project | Contribution |
|---|---|
| 🧠 Brian2, a spiking neural network simulator | ✅ #1882 Replace deprecated pyparsing parseAll with parse_all in parse_statement |
| ✅ #1884 Remove obsolete Python linting settings from the devcontainer | |
| ✅ #1888 Add a VS Code task to run the test suite | |
| ✅ #1889 Skip submodule fetches when pulling upstream tags in the docs build workflow |
| Project | What it does | Stack |
|---|---|---|
| 🏭 Production Line Performance Checker | Predictive maintenance from sensor data with SHAP explainability, deployed on Render | FastAPI · Docker · SHAP |
| ⚙️ Minimal MLOps Batch Pipeline | Config-driven, validated and logged batch pipeline for financial time series | Python · Docker · YAML |
| ✨ Halo | Cinematic marketing site for an AI call copilot, with generative canvas backgrounds and pointer-reactive motion | Next.js 16 · Tailwind 4 · Motion |
| 👁️ Smart Vision Assistant | Real-time multi-object detection agent | YOLOv8 · OpenCV |
| 🧩 CppKit | VS Code extension with fast C++ and competitive-programming snippets | VS Code API |
| 🏏 IPL Prediction System | Match outcome prediction from historical IPL data | scikit-learn · Pandas |
| 📈 Sales Forecasting | Time-series forecasting with a Power BI insights dashboard | Python · Power BI |
| 📉 Customer Churn Prediction | Churn risk modelling with business-ready retention insights | scikit-learn · XGBoost |
| ⏱️ Focus Pomodoro | Productivity timer web app | React · TypeScript · Vite |
📚 ML foundations & practice notebooks (click to expand)
| Area | Repos |
|---|---|
| 🧠 Deep learning | 3-or-7 Neural Net Classifier · Cat & Dog Breed Classifier · MNIST 5-Classifier |
| 🌲 Classical ML | Breast Cancer (XGBoost) · Decision Trees on make_moons · MNIST with SVM · California Housing with SVR |
| 📊 Analytics | Instructor Effectiveness Modeling · Champions League 25/26 Analysis · Pearson's r on California Housing |
| 🧮 DSA | LeetCode Submissions · NeetCode Submissions (C++) |
Languages
AI / ML
Data
Backend
Frontend & Mobile
DevOps & Tools
| 🔎 Evidence over generation | LLMs explain results; they never invent them. |
| ⚖️ Tests decide, models assist | Scores, grades and verdicts come from deterministic code, never from a model's opinion. |
| 🤚 Abstain rather than guess | If the data can't support an answer, the system says so. |
| 🧪 Tests are part of the product | Hundreds of automated tests per project, adversarial and prompt-injection cases included. |
| 🔒 Offline-first when it matters | Sensitive and low-connectivity workloads run fully offline, with no cloud dependency. |
| 📖 Document like it's a paper | Every claim in a README is backed by a script or a documented condition. |