Sai Ram Reddy Solleti

Twilio receives the call and Deepgram Nova‑2 transcribes it. An n8n workflow drives an LLM as a tool‑calling agent against a FastAPI booking backend on PostgreSQL, and ElevenLabs speaks the reply back through TwiML. Conversation history writes to PostgreSQL as JSONB every turn, so the workflow holds no in‑memory session state.

Tested across 90 staged calls covering six intent types — availability, reservation, confirmation lookup, cancellation, policy question, mixed conversation — fifteen runs each.

Most “quantum” voting prototypes call Python’s secrets module and describe it as quantum. This one runs real circuits on Qiskit’s Aer simulator for every security‑critical value. Ballot receipts draw 96 bits across six 16‑qubit Hadamard circuits — each qubit placed in equal superposition and collapsed by measurement, with no seed to predict.

BB84 key distribution runs on real single‑qubit circuits using X and H gates, with Alice’s bits, Alice’s bases and Bob’s bases all quantum‑generated. Voter identity is decoupled from ballots by SHA‑256 commitment, every event written to a hash‑chained audit log.

ResNet‑50 pretrained on ImageNet, fine‑tuned on the German Traffic Sign Recognition Benchmark. Phase one trains only the classification head against a frozen backbone — 1,071,147 trainable parameters, 4.36% of the network — taking validation accuracy from 79.2% to 92.6%.

Phase two unfreezes everything at a lower learning rate with cosine annealing and early stopping. Trained on 26,640 images across 43 classes, with EfficientNet‑B0 benchmarked as a lightweight alternative.

Hindi, Russian, Arabic, Greek, Chinese, Japanese, Korean, Hebrew, Telugu and Bengali. Phonetic transliteration runs client‑side, so raw non‑English text never reaches a server.

No benchmark existed for this, so I built one: 100 questions across 10 languages, 1,000 prompts total. Later extended with faster‑whisper for push‑to‑talk voice input.

128‑dimensional embeddings, 4 attention heads, 3 layers, trained on 1,748 chronologically ordered Bitcoin sequences distilled from an 87,000‑record dataset using 40‑step sliding windows.

Residuals concentrate in high‑volatility windows, which is where a model this size should be expected to struggle.

A custom spaCy named‑entity model with GloVe embeddings, trained on 525 annotated queries spanning 20 intents and 10 entity types, mapping requests onto a library of 100 code snippets behind a REST API and client.