AI/ML engineering student who builds working systems, not just notebooks.
Third-year B.Tech, Computer Science & Engineering (AI) at Amrita Vishwa Vidyapeetham, Coimbatore (graduating 2028). I like problems where a model has to survive contact with real data: missing sensors, imbalanced labels, noisy feeds, users who don't read instructions. Most of my projects ship with tests, a one-command launcher, and a section that says what doesn't work yet.
Currently building VoxAgent — a multilingual AI phone receptionist for Indian clinics, salons and coaching institutes.
| Project | What it is | Stack | |
|---|---|---|---|
| 1 | InterviewAce AI | Full-stack interview prep: rule-based ATS score, voice mock interviews scored on a 5-dimension rubric, gap-driven study plan. Works fully offline; LLMs are optional. | FastAPI · React/TS · SQLAlchemy · 92 tests |
| 2 | disaster-risk-bayesian-network | 18-node Bayesian network for flood/landslide/cyclone/heatwave risk on live public weather feeds, with explanations and MEU decisions. Beats ML baselines once ~25% of sensors drop out. | pgmpy · FastAPI · Leaflet · 85 tests |
| 3 | cipl-eye-disease-diagnosis | Interpretable multi-label fundus diagnosis (8 diseases) with per-disease localisation maps and a text report. From-scratch prototype head vs. pretrained ResNet-18 + CBAM (0.85 validation mAUC). | PyTorch · Kaggle GPU |
| 4 | uav-apriltag-landing-sim | MuJoCo quadrotor that lands on addressed, moving pads using its own camera and real AprilTag decoding. 40/40 Monte Carlo landings, ~3 cm mean error. | MuJoCo · OpenCV · control |
| 5 | greenscale-cloud | Carbon-aware multi-cloud scheduler with an ECDSA-signed, Merkle-tree ledger of every placement decision. −72% CO₂ vs. nearest-region baseline in simulation. | Flask · scikit-learn · cryptography · 72 tests |
| 6 | polylag | Event-driven system that measures whether prediction markets lag the news, with paper trading and a three-layer risk manager. | asyncio · httpx · websockets · 68 tests |
Most of these were team projects; each README says who did what.
Languages — Python · TypeScript · SQL · MATLAB ML / DL — PyTorch · scikit-learn · TensorFlow/Keras · pgmpy (Bayesian networks) · OpenCV AI systems — LLM APIs (Gemini, Claude, OpenAI) behind provider-agnostic interfaces · speech-to-text / TTS pipelines · telephony (Twilio) Backend / web — FastAPI · Flask · Next.js · React · Tailwind · SQLAlchemy · Prisma · PostgreSQL / SQLite Tooling — Git · GitHub Actions · Docker · pytest · Kaggle · MuJoCo
- VoxAgent (private) — AI receptionist that answers calls in English, Telugu and Hindi and books appointments. Terminal agent → voice → real-time interruptible calls → Twilio phone numbers, with a business portal and an eval suite.
- Re-running the eye-disease experiments with a patient-level train/val/test protocol (the harness is done; the numbers aren't yet).
Agent tooling (MCP servers, LangGraph) · prototype-based interpretability · evaluation methodology for small, imbalanced medical datasets.
- Reinforcement Learning for 4×4 Tic-Tac-Toe: Q-Learning with Tactical Enhancements and Epsilon Decay — IEEE ESCI 2026 (co-author). DOI 10.1109/ESCI68015.2026.11493322
- Hybrid CNN–ViT image forgery detection — co-author, under review at IEEE Access.
Open to AI/ML and backend internships and to collaborating on applied AI projects. Reach me at saireddyannem2007@gmail.com.