I’m Ramkumar — a CSE (AI & ML) undergrad at Chennai Institute of Technology (CGPA 8.99) and an applied-AI + full-stack developer who turns models into things people can actually use. My work sits at the seam between machine learning and product — real-time computer vision, retrieval-augmented assistants, agentic workflows, and the FastAPI + React plumbing that ships them.

• 🧠 Working across AI and systems engineering — real-time computer vision and speech/audio ML (SpatialAI, AI Interview Twin), generative & agentic AI with RAG, embeddings, and LangChain/LangGraph workflows, and system-level database internals in Rust — Raft consensus, MVCC/LSM storage, distributed transactions (arcux).
• 🛠️ AI engineer × system-level engineer — I work at both altitudes: building and serving models, and hand-rolling the infrastructure underneath them (consensus, storage engines, gRPC services) — with the software engineering to ship both as real products.
• 🏆 Hackathon builder — international winner at AI for Sustainability (UAE) and ₹50K runner-up at TI FORGE; I like shipping working prototypes under pressure.
• 🌱 Currently going deeper on edge/on-device LLM inference, retrieval quality, latency-aware ML pipelines, and system-level DB internals — consensus, storage engines, and distributed transactions.
• 🤝 Open to collaborating on applied AI, computer vision, and developer tools.

A from-scratch, range-sharded, Raft-replicated database in Rust where each table declares its own consistency regime at creation — CP (Percolator 2PC + Raft + timestamp oracle → Snapshot Isolation) or AP (leaderless W=1 + HLC + Last-Writer-Wins → always available) — one storage engine, one cluster, two write paths.

• Architecture: keyspace range-sharded into regions (CockroachDB/TiKV-style); each CP region runs its own independent Raft group with majority commit, while AP regions take leaderless HLC-stamped writes with best-effort fan-out — all landing in the same WAL + MVCC-over-LSM storage engine.
• Engineering details: consensus, storage engine, and transaction protocol are hand-rolled (no raft-rs, no RocksDB) — leader election, log replication, snapshotting, and membership changes proven deterministically; crash recovery, property tests, a Placement Driver (timestamp oracle + region registry), and a region-aware async client SDK with transparent leader-following.
• Stack: Rust · tonic gRPC · Raft · Percolator 2PC · MVCC / LSM · HLC + LWW · dual MIT/Apache-2.0

An authorization and idempotency layer for agent-initiated payments (Razorpay AI Buildathon, Track 01). The LLM is untrusted input: it can only propose a purchase — a human confirms it, and a deterministic policy engine checks it against a merchant-defined mandate before anything reaches the payment provider.

• Architecture: LLM intent parser (Groq) returns a ParsedIntent with no price or approval field → human confirmation → price re-validated against the catalog → policy engine (per-transaction cap, item allowlist, expiry, sliding-window velocity limit) → idempotent executor → Razorpay order → double-entry ledger + reconciler, with every step written to an append-only, hash-chained audit log.
• Engineering details: at-most-once execution via a Postgres primary-key claim (ON CONFLICT + FOR UPDATE + advisory locks); a payment timeout is frozen as PENDING_VERIFICATION instead of retried; a stale claim is checked against the provider, with a fencing token, before it is reclaimed; ledger postings are refused at COMMIT unless they balance; audit-log UPDATE / DELETE / TRUNCATE are blocked by database triggers.
• Verification: a hostile-agent suite against the live API (14/14 attacks defended, 0 unintended charges) and a chaos harness that kills database connections and SIGKILLs processes mid-payment (100/100 rounds correct) — which exposed a real double-charge bug, since fixed. 685 tests. Order creation uses Razorpay's live test-mode API; capture is simulated and labelled as such.
• Stack: Python · FastAPI · PostgreSQL · Razorpay API · Groq LLM · PyNaCl · React

Turns a single RGB camera into spoken, metric spatial guidance — "backpack 0.9 m directly ahead blocking your path, step right" — running fully on-device, no cloud.

• Architecture: parallel perception branches — open-vocabulary detection (YOLO-World) + dense monocular depth (Depth Anything V2) → pinhole back-projection to 3D → tracked, dwell-filtered scene graph → on-device SLM (Phi-3-mini INT4, llama.cpp) → neural TTS (Kokoro-82M).
• Engineering details: threaded capture/inference so latencies overlap, IOU tracking + Hungarian matching + EMA smoothing for stability, a semantic change-gate to suppress repeats, camera-intrinsics calibration, a pytest suite, and CSV latency profiling.
• Stack: Python · PyTorch · YOLO-World · Depth Anything V2 · llama.cpp · sentence-transformers · OpenCV · Kokoro TTS

A full-stack system that runs a mock interview end-to-end: generate questions → record spoken answers → transcribe → score → coach.

• Pipeline: LLM question generation → in-browser audio capture (MediaRecorder) → Whisper transcription → Communication + Technical Correctness scoring → placement-ready feedback.
• Stack: FastAPI · faster-whisper · PyTorch · Transformers · librosa · spaCy · React · TypeScript · MIT-licensed

A security dashboard that flags phishing across two surfaces — suspicious URLs and email content — backed by ML models.

• Features: URL phishing detection, email phishing detection, and an analytics dashboard (screenshots in the repo).
• Stack: React · TypeScript · Tailwind · shadcn/ui · Python ML backend

Upload notes or textbooks (including scanned PDFs) and ask questions — a retrieval-augmented study assistant whose answers stay grounded in your own documents.

• Architecture: PDF/image ingestion with OCR (Tesseract + pdf2image) → sentence-transformer embeddings → FAISS vector search → LangChain + a Hugging Face model for grounded responses, served via FastAPI to a web frontend.
• Stack: FastAPI · LangChain · FAISS · sentence-transformers · Transformers · PyTorch · Tesseract OCR

Identifies fish/marine species from underwater audio — an applied-ML project implemented from a published research paper that's included in the repo.

• Approach: audio feature extraction → species classification model, with a runnable demo app and a model test harness.
• Stack: Python · audio ML · NumPy · grounded in peer-reviewed research

🔬 Also worth a look: Road Safety Intervention GPT — RAG over a curated road-safety knowledge base · RAG Chatbot for LCA Tool — retrieval assistant for life-cycle assessment · air_mouse — Android (Kotlin) gesture mouse.

_{Always building. Always learning. 🚀}